Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 113 appraisals
PloS one
Multi-view graph-regularized deep metric subspace clustering network
Multi-view subspace clustering has progressed significantly by using deep neural networks to handle nonlinear data representations. A recent advancement, the Multi-view Self-Expressive Subspace Clustering (MSESC) network, achieves markedly higher computational efficiency by substituting the traditional self-expression layer with a deep metric learning approach. Nevertheless, MSESC still suffers from two notable limitations: it fails to adequately capture the high-order geometric structures inherent in multi-view data, and it lacks effective guidance from the underlying clustering distribution. To overcome these shortcomings, we propose a novel framework termed Multi-View Graph Regularized Deep Metric Subspace Clustering (MVGR-DMSC). The proposed method introduces two key components into MSESC to enhance the discriminability of representations. First, a dual-order graph regularization module is devised to maintain both first-order and second-order manifold structures, thereby allowing the model to capture more complex local geometric relationships. Second, an adaptive view-weighted deep clustering module is incorporated, which employs the Kullback-Leibler divergence to guide representation learning while dynamically adjusting the contributions of different views. Through evaluations on five benchmark datasets, we show that MVGR-DMSC consistently yields better results than several state-of-the-art approaches, including the direct baseline MSESC, in both accuracy and robustness.
26 July 2026
Read appraisal →PloS one
Fundus image analysis of retinitis pigmentosa using artificial intelligence
Retinitis pigmentosa (RP) is a group of inherited retinal diseases that are caused by genetic defects that lead to progressive photoreceptor loss and eventual blindness. Early diagnosis would be helpful for effective management of the disease; however, many patients stay unaware of early symptoms. Meanwhile, fundus images are widely obtained during routine medical checkups but are underused for detecting RP. This study explores the effectiveness of finetuning deep learning models, pre-trained for general visual tasks, to identify RP from color fundus images. The dataset comprised 321 color fundus images from 201 Japanese subjects at Keio University Hospital, including 200 images from 107 patients with retinitis pigmentosa and 121 images from 94 non-retinitis pigmentosa subjects. Multiple images were available for some subjects. Using transfer learning, pretrained convolutional neural network models -VGG16, Resnet50, and InceptionV3- were finetuned to detect RP. As a result, Inception V3 achieved the best accuracy of 96.97%, which matches the average diagnostic accuracy of ophthalmologists. Gradient-weighted Class Activation Mapping (Grad-CAM) suggested that the model attended to clinically relevant fundus regions, including the peripheral retina and posterior pole, which may reflect features such as peripheral degenerative changes and retinal vascular attenuation. These findings support the potential interpretability of the finetuned model and suggest that deep learning may assist ophthalmologists in RP screening as a supportive tool.
25 July 2026
Read appraisal →IEEE transactions on pattern analysis and machine intelligence
50 Years of Automated Face Recognition
Over the past five decades, automated face recognition (FR) has progressed from handcrafted geometric and statistical approaches to advanced deep learning architectures that now approach, and in many cases exceed, human performance. This paper traces the historical and technological evolution of FR, encompassing early algorithmic paradigms through to contemporary neural systems trained on extensive real and synthetically generated datasets. We examine pivotal innovations that have driven this progression, including advances in dataset construction, loss function formulation, network architecture design, and feature fusion strategies. Furthermore, we analyze the relationship between data scale, diversity, and model generalization, highlighting how dataset expansion correlates with benchmark performance gains. Recent systems have achieved near-perfect large-scale identification accuracy, with the leading algorithm in the latest NIST FRTE 1:N benchmark reporting a False Negative Identification Rate (FNIR) of 0.15 percent at False Positive Identification Rate (FPIR) of 0.001 on a gallery of over 10 million identities. Larger galleries increase false positive rates and deployments at greater scales will see higher error rates. We delineate key open problems and emerging directions, including scalable training, multi-modal fusion, synthetic data, and interpretable recognition frameworks.
1 July 2026
Read appraisal →Journal of integrative medicine
Deep learning in traditional Chinese medicine
As an alternative medicine approach, traditional Chinese medicine (TCM) is mainly built on an ancient understanding of the world. In order to translate the promises of TCM into modern therapies, rigorous scientific techniques are applied, which generates big data from experiments, clinical practices and medical literature. Recently, deep learning (DL) has attracted the attention of TCM researchers due to its application in data mining. In this review, we summarized recent examples of DL application in the study of TCM, including medical image processing, medicinal substances investigation, data fusion and natural language processing. We find that DL methods have considerable potential to extract valuable information from TCM data, but high-quality TCM data are still needed for further development of DL techniques. Please cite this article as: Liu LP, Yang C, Cen SX, Suo TC. Deep learning in traditional Chinese medicine. J Integr Med. 2026; 24(4):471-480.
1 July 2026
Read appraisal →PloS one
A lightweight hybrid deep learning approach for fashion mnist classification with explainable attention visualization
The classification of fashion images is an essential task in the e-commerce sector, where accurate categorization improves user experience and refines product discovery. Convolutional Neural Networks (CNNs) and Transformers have demonstrated strong performance in image classification tasks due to their ability to learn complex visual features. However, deep variants of these architectures, such as VGG-19, ResNet-50, Vision Transformer (ViT), and Swin Transformer, contain tens of millions of parameters, requiring high memory and powerful GPUs for training, which makes them less suitable for low-resource and edge device environments. To address these limitations, this research proposes a lightweight hybrid architecture, TinyCNN with Linear Self-Attention (LSA), optimized for resource-constrained settings. The proposed model contains fewer than half a million parameters and is trainable on a CPU, achieving a classification accuracy of 91.47% on the Fashion-MNIST dataset. In addition, multiple Explainable Artificial Intelligence (XAI) techniques are implemented, including Self-Attention visualization, Multi-Head Attention, Attention Flow, Attention Rollout, Fixed query position attention maps, Integrated Gradients, LIME, and SHAP, to provide visual interpretability of the model's predictions and enhance transparency in its decision-making process.
25 June 2026
Read appraisal →International journal of medical informatics
Human in the loop artificial intelligence in healthcare: applications, outcomes, and implementation challenges
BACKGROUND: The integration of artificial intelligence in healthcare has transformed clinical practice and research methodologies. However, concerns regarding algorithmic accountability, interpretability, and safety have necessitated human oversight in AI systems. Human in the loop artificial intelligence represents a collaborative paradigm where human expertise and machine intelligence converge to enhance decision making while maintaining ethical standards and clinical safety. AIM: This review synthesizes current evidence on human in the loop AI in healthcare delivery and research, examining implementation frameworks, clinical outcomes, comparative advantages over fully automated and clinician-only approaches, and challenges. METHOD: A comprehensive narrative review was conducted using PubMed, Scopus, Web of Science, and IEEE Xplore databases covering studies from 2018 to 2025. Data were thematically synthesized to identify patterns, frameworks, and outcomes. This narrative approach enables comprehensive conceptual synthesis across diverse HITL-AI applications and contexts. RESULTS: Human in the loop AI demonstrates significant applications across diagnostic imaging, clinical decision support, patient monitoring, drug discovery, and research data analysis. Evidence indicates improved diagnostic accuracy, reduced medical errors, enhanced patient safety, and increased clinician trust compared to both automated AI and traditional approaches. Implementation requires EHR interoperability, clear liability frameworks, adaptive training protocols, and quantum-safe cryptographic security. Challenges include workflow integration, regulatory gaps for adaptive systems, and sustainability concerns. CONCLUSION: This review advances the field by synthesizing cross-domain implementation patterns, mapping collaboration models to risk-stratified contexts, identifying regulatory gaps for adaptive systems, and proposing future directions including post-quantum cryptographic integration, AI-driven adaptive architectures, and multi-center scalability frameworks for optimizing human-machine collaboration in healthcare.
15 June 2026
Read appraisal →Computer assisted surgery (Abingdon, England)
DMRNet: a dynamic multi-scale residual network for Shamrock view and lumbar plexus segmentation
Lumbar plexus block (LPB) is a regional anesthesia technique widely used for hip and knee surgeries. However, despite the assistance of ultrasound guidance, the complex anatomical structure of the lumbar plexus poses significant challenges for anesthesiologists during the procedure. To accurately identify the lumbar plexus located in the posterior third of the psoas major in the Shamrock view, a deep learning-based segmentation model named DMRNet was proposed. This model is designed to precisely delineate muscles, nerves, and bony structures in Shamrock view ultrasound images. DMRNet integrates several innovative modules, including Adaptive Multi-Scale Dilated (AMD) Module that enhances the model's ability to capture multi-scale features; Dense Attention Residual (DAR) Module that adaptively selects salient feature regions; Attention-Enhanced Hybrid (AEH) Module that emphasizes critical features while suppressing irrelevant ones; and two attention mechanisms, Boundary-Aware Spatial Attention Mechanism (BASA) and the Enhanced Residual Multi-Head Attention Mechanism (ER-MHA), that improve the model's capacity to recognize complex contextual patterns. Experimental results demonstrated that DMRNet achieved a mean Intersection over Union of 0.863 and a mean Dice coefficient of 0.926 across all target structures, outperforming other state-of-the-art models. These findings suggest that DMRNet may serve as a useful assistive tool for sonoanatomical recognition in Shamrock view ultrasound images and may provide educational support for ultrasound-guided LPB training. A dynamic multi-scale residual network is proposed for lumbar lexus (LP) segmentation.Two attention mechanisms were improved, and three modules were proposed to improve downsampling, skip connection, and upsampling.We built our own Shamrock view ultrasound database containing the lumbar plexus nerves, and the Anesthesiologist physician of the anesthesiology department annotated it using Labelme.External validation was performed using the Ultrasonic Brachial Plexus Dataset (UBPD).
14 June 2026
Read appraisal →JMIR formative research
Automated Machine Learning Frameworks for Radiomics: Comparative Evaluation Study
BACKGROUND: Automated machine learning (AutoML) frameworks can lower technical barriers for predictive and prognostic model development in radiomics by enabling researchers without programming expertise to build models. However, their effectiveness in addressing radiomics-specific challenges remains unclear. OBJECTIVE: This study aimed to evaluate the performance, efficiency, and accessibility of general-purpose and radiomics-specific AutoML frameworks on diverse radiomics classification tasks, thereby guiding researchers and highlighting development needs for radiomics. METHODS: A total of 10 public and private radiomics datasets with varied imaging modalities (computed tomography and magnetic resonance imaging), sizes, anatomies, and end points were used. Six general-purpose and 5 radiomics-specific frameworks were tested with predefined parameters using standardized cross-validation. Evaluation metrics included area under the receiver operating characteristic curve, runtime, and qualitative aspects related to software status, accessibility, and interpretability. RESULTS: Simplatab, a radiomics-specific tool with a no-code interface, achieved the best overall balance between performance and computational efficiency, recording the highest average test area under the receiver operating characteristic curve (mean 78.46%, SD 12.22%) with a moderate runtime (1.1 h). However, its performance was not statistically superior to the most intensive general-purpose solutions. Most radiomics-specific frameworks were excluded from the performance analysis due to obsolescence, extensive programming requirements, or computational inefficiency. Conversely, general-purpose frameworks demonstrated higher accessibility and ease of implementation. CONCLUSIONS: While no single framework demonstrated absolute predictive superiority, Simplatab provides an effective balance of performance, efficiency, and accessibility for radiomics classification problems. However, continued efforts are needed to further mature AutoML solutions in the radiomics domain.
14 June 2026
Read appraisal →Biomedical physics & engineering express
Enhancing brain tumor classification with a simplified CNN through hyperparameter optimization
Brain tumors remain a major public health challenge because of their high mortality rate and the need for timely and accurate diagnosis. Magnetic resonance imaging (MRI) is widely used for brain tumor assessment, yet conventional diagnostic workflows still depend largely on the subjective interpretation of radiologists, which may lead to inter-observer variability. In this context, deep learning approaches-particularly convolutional neural networks (CNNs)-have gained increasing attention for automated medical image analysis, as they are capable of learning discriminative features directly from imaging data and improving classification accuracy. Nevertheless, the performance of CNN-based models is strongly influenced by the choice of hyperparameters, whose optimization remains a non-trivial task due to complex and non-linear interactions within the network.This study investigates the impact of different hyperparameter optimization strategies on the performance of a simplified CNN model for brain tumor classification from MRI images. Three widely used optimization methods, Genetic Algorithms (GA), Bayesian Optimization (BO), and HyperBand-are considered and systematically compared. To emphasize efficiency and practical applicability, the CNN architecture is deliberately kept lightweight, aiming to reduce computational complexity while preserving sufficient modeling capacity. Experiments are carried out on a publicly available brain MRI dataset that includes glioma, meningioma, pituitary tumor, and healthy subjects.To account for performance variability and ensure statistical reliability, each optimized model is evaluated over ten independent runs using different random initializations. The reported results therefore include both the mean performance values and the corresponding standard deviations. The experimental analysis shows that the GA-based approach achieves a mean accuracy of 96.41%±0.36%, while BO improves this performance to 97.37%±0.30%. Among the evaluated strategies, HyperBand yields the highest average accuracy of 97.73%±0.48%, indicating its ability to consistently identify effective hyperparameter configurations for the proposed CNN architecture.
13 June 2026
Read appraisal →Nature communications
Adaptive Riemannian optimization for multi-scale diffeomorphic matching
Image matching is a fundamental task in quantitative biomedical and biological image analyses, enabling researchers to compare, integrate, and interpret imaging data across subjects, time points, modalities, and experimental conditions. Existing state-of-the-art registration methods are slow due to inefficient implementations and poor convergence rates because of the ill-conditioned nature of the optimization problem. Deep learning methods offer fast inference but require extensive training time, substantial inference memory, and fail to generalize across long-tailed distributions or diverse image modalities, necessitating costly retraining. We address these challenges by proposing FireANTs, a training-free, GPU-accelerated, multi-scale adaptive Riemannian optimization algorithm for fast and accurate dense diffeomorphic image matching. FireANTs more than doubles the speed of the community standard ANTs registration tool on a CPU, and is two orders of magnitude faster on a GPU. On the GPU, FireANTs performs competitively with deep learning methods on inference runtime while consuming up to 10 × less memory. FireANTs demonstrates robustness on a wide variety of matching problems across modalities, species, and organs, without any domain-specific training or tuning. Our framework allows hyperparameter grid search studies with less resources and time compared to traditional and deep learning registration algorithms alike.
12 June 2026
Read appraisal →Physics in medicine and biology
Unsupervised multimodal deformable medical image registration based on feature perceptual contrast learning
Objective.Multimodal medical image registration has extensive applications in clinical diagnosis and is fundamental for a series of medical analysis tasks. However, the presence of modality differences makes the registration process challenging. Existing methods often employ modality-independent feature descriptors that are sensitive to noise, or attempt to bridge differences within networks, which typically results in translation inaccuracies and misaligned anatomical information.Approach.In this paper, we propose a novel unsupervised approach utilizing a Feature Perceptual Contrast Learning Network (FP-net) to learn descriptors that bridge modality differences while accurately capturing common details. We unify the feature representation of anatomical information under homogeneous and heterogeneous intensity distributions through local sampling-based feature perceptual contrast learning and image reconstruction learning. The trained FP-net is subsequently employed to drive an unsupervised registration framework without requiring ground-truth deformation fields.Main results.We extensively evaluated our method on two public benchmarks: the BraTS 2021 dataset for brain T2-T1 and T1-T1ce registration, and the Learn2Reg 2021 dataset for challenging abdominal CT-MR registration. By passing multimodal image pairs with shape differences through the fixed FP-net, we generate optimization gradients that successfully update the registration network. Quantitative evaluations demonstrate our method's superiority over state-of-the-art baselines. Specifically, our model achieved a dice similarity coefficient (DSC) of 76.3% and 77.7% in tumor-bearing T2-T1 and T1-T1ce tasks, respectively. Furthermore, in the complex abdominal CT-MR task, it reached a DSC of 50.1%, significantly improving structural alignment.Significance.Our method effectively shifts the burden of bridging the modality gap away from the registration network, enabling standard U-Net architectures to achieve state-of-the-art deformable registration. This provides a robust, accurate, and easily deployable unsupervised solution for complex clinical multimodal image analysis.
12 June 2026
Read appraisal →Physics in medicine and biology
Scan-wise generalized PET denoising with contrastive adversarial learning
Objective.Deep learning has significantly advanced low-count positron emission tomography (PET) denoising. However, models trained on specific distributions often yield biased outputs when applied to scans with different activity distributions caused by anatomical and physiological variations (distribution shifts). Existing methods fail to generalize well across these scan-wise variations. Our goal is to formulate PET denoising as a scan-wise domain generalization (DG) problem to mitigate these variations and achieve robust, unbiased denoising for unseen scans.Approach.We propose a contrastive adversarial DG (CADG) framework to learn scan-invariant features efficiently. We leverage the property that multiple low-count noise realizations can be generated from a single raw list-mode PET scan to form a scan-wise domain distribution. Unlike conventional adversarial training with cross-entropy (CE) loss, we propose a contrastive adversarial framework that minimizes the mutual information between feature and scan-wise domains. Considering that each subject can have multiple scans in longitudinal studies, we further propose an ordered and memory-queued contrastive adversarial framework. This method efficiently includes realizations from the same scan as positive pairs, different scans of the same subject as pseudo-positive pairs, and different subjects as negative pairs in a memory batch. We explicitly exploit their ordered relationship as prior knowledge using a novel noisy-robust multi-positive ordinal contrastive loss.Main results.We systematically validated the effectiveness of our methods using 1920 noise realizations derived from 80 subjects with 192 longitudinal scans of MK-6240 tau PET data. The proposed CADG approach demonstrated superior denoising performance compared to CE-based adversarial methods and standard baselines. The ordered contrastive loss successfully improved the peak signal-to-noise ratio and structural similarity index consistently while reducing bias and standard deviation in Alzheimer-related regions and the whole brain.Significance.To our knowledge, this is the first attempt to alleviate the performance degradation in cross-scan denoising from the perspective of DG. Our study is also a pioneering work in utilizing longitudinal scans as pseudo-positives within an ordered contrastive learning scheme to exploit fine-grained relationships for robust clinical PET imaging applications.
9 June 2026
Read appraisal →Scientific data
A Dataset of Widefield Fundus Images From Patients With Birdshot Chorioretinitis and Matched Control
Birdshot chorioretinitis is a rare uveitis where research is limited by the lack of large, publicly available imaging datasets. To address this gap, we present Birdshot-Wide, a standardized retinal imaging dataset comprising 5,042 widefield fundus photographs from 742 eyes with birdshot chorioretinitis (BSCR) and 1,310 images from 742 age-and sex- matched healthy eyes controls. The BSCR cohort is longitudinal, with a median follow-up of 4.31 years. Each image in the BSCR cohort was manually reviewed for quality and classified by dominant lesion subtype. To demonstrate its utility, the dataset was benchmarked on a classification task, where deep learning models trained to differentiate BSCR from control images achieved an area under the receiver operating characteristic curve of up to 0.96. This dataset, with its detailed metadata, provides a valuable resource for biomarker discovery, the development of automated diagnostic tools, and disease progression modeling.
8 June 2026
Read appraisal →Radiation oncology (London, England)
HyperSightTM-CBCT based monitoring of sarcopenia during definitive radiotherapy of prostate cancer: a longitudinal feasibility study
BACKGROUND: Sarcopenia and treatment-related muscle loss are associated with increased toxicity and reduced resilience in cancer patients. HyperSight™ cone-beam CT (CBCT) provides improved soft-tissue contrast compared with conventional CBCT and may enable reliable longitudinal skeletal-muscle monitoring during image-guided radiotherapy (IGRT). This study evaluated the feasibility of HyperSight™ CBCT-based muscle assessment in prostate cancer patients and explored the clinical relevance of intratreatment muscle decline. METHODS: Eighty-eight patients with non-metastatic prostate cancer were treated on an Ethos® system with HyperSight™ CBCT. Patients received ultrahypofractionated (UHF) or normo-/moderately hypofractionated (NF/MHF) radiotherapy; 28% received androgen-deprivation therapy. Cross-sectional muscle area (CSA) and radiodensity (HU) were assessed on planning CT (PCT) and serial CBCTs (first, second, last fraction [FF, SF, LF]); HU values were z-standardized to subcutaneous fat (zHU). Agreement between PCT and CBCT was evaluated using concordance correlation coefficients (CCC), regression, and Bland-Altman analysis. Longitudinal changes were analyzed using mixed-effects models, and associations with acute genitourinary/gastrointestinal (GU/GI) toxicity (RTOG ≥ 2) using regression models. RESULTS: HyperSight™ CBCT showed excellent agreement with PCT for CSA (CCC = 0.981). The NF/MHF cohort showed a significant longitudinal CSA decline (FF-LF: -153.0 mm², p = 0.021), whereas no significant within-cohort change was observed in UHF. Muscle density (zHU) decreased significantly in both cohorts (NF/MHF: p < 0.001; UHF: p = 0.0017). Weight loss and age were independently associated with zHU change, while no independent predictors of CSA decline were identified. In the NF/MHF cohort, CSA loss > 5% was associated with acute GU toxicity in univariate analysis (OR 4.90, p = 0.033; Fisher's exact p = 0.014), but not after adjustment for age, weight loss, ECOG performance status, and ADT; no associations were found for GI toxicity or in the UHF cohort. CONCLUSIONS: HyperSight™ CBCT enables quantitative skeletal-muscle assessment during prostate radiotherapy and detects intratreatment reductions in muscle density, which may represent an early imaging marker of muscle decline. Muscle loss showed an exploratory association with acute GU toxicity in NF/MHF patients in univariate analysis, but not after adjustment. These findings support the feasibility of CBCT-based muscle monitoring and warrant further evaluation in larger cohorts. CLINICAL TRIAL NUMBER: Not applicable.
1 June 2026
Read appraisal →Biomedical physics & engineering express
Biwt-UNet: lung nodule segmentation via wavelet transform and multi-scale feature fusion
Lung nodule segmentation is a crucial step in computer-aided diagnosis of lung diseases and is of great significance for early detection and diagnosis of lung cancer. Traditional segmentation methods face many challenges, such as blurred nodule boundaries that are difficult to distinguish effectively, and the diversity of nodule sizes and shapes, which together increases the difficulty of segmenting lung nodules. Existing approaches often fail to address these issues adequately, highlighting the need for more effective solutions. To meet these challenges, this paper proposes a hybrid lung nodule segmentation approach-Biwt-UNet that integrates improved Haar wavelet transform and multi-scale feature information from the multi-scale fusion module and bi-encoder fusion module proposed in this study. The approach enables accurate segmentation of lung nodules of varying sizes, diverse shapes and, in particular, indistinct boundaries. Experimental results on the public LIDC-IDRI dataset demonstrate that Biwt-UNet achieves excellent segmentation performance. It significantly surpasses other state-of-the-art methods with an average dice similarity coefficient of 90.18%, an average intersection over union of 82.67%, and an average normalized surface dice of 96.39%, fully validating the effectiveness and accuracy of the proposed model in CT-based lung nodule analysis. Moreover, ablation studies further confirm the individual contributions of each architectural component and show that the best performance is achieved within an acceptable parameter budget. This study provides a new technical perspective for automatic lung nodule segmentation and is expected to assist physicians in more efficient diagnosis.
29 May 2026
Read appraisal →European radiology experimental
Enhancing resolution and image quality in musculoskeletal MRI using deep learning reconstruction
OBJECTIVE: Deep learning-based noise reduction enhances image quality, overcoming the tradeoff among acquisition time, spatial resolution, and signal-to-noise ratio (SNR). We implemented deep learning reconstruction (DLR) into a 1.5-T musculoskeletal (MSK) magnetic resonance imaging (MRI) protocol to improve image quality without compromising SNR. MATERIALS AND METHODS: We retrospectively analyzed 39 MRI examinations performed on a 1.5-T scanner using standard-resolution (SR) sequences and sequences with higher resolution reconstructed with DLR (HR-DLR). Exams of the knees, shoulders, ankles, and hips were evaluated. The included sequences were: three-dimensional T2-weighted fast advanced spin-echo; T1-weighted and proton density-weighted fast spin-echo. One expert reader and two junior readers in agreement evaluated the visibility of various structures using a 5-point Likert scale in a blind manner. A fourth reader estimated the SNR and contrast-to-noise ratio (CNR) in bone and muscle. A mixed model was used to compare HR-DLR versus SR measures. The agreement between radiologists was assessed with the Kendall τ coefficient. RESULTS: The HR-DLR sequences globally had a smaller pixel size and shorter acquisition time. A good inter-reader agreement was obtained for SR sequences (0.613 ≤ τ ≤ 0.788) and even higher levels of agreement for HR-DLR sequences (0.682 ≤ τ ≤ 0.961). All the structures had higher or similar Likert scores in HR-DLR sequences (p < 0.001), regardless of joint and sequence contrast. Apparent SNR and CNR of HR-DLR and SR were similar. CONCLUSION: Incorporating DLR into 1.5-T MSK MRI protocols enhances resolution and maintains SNR and CNR, improving MSK structure visualization. RELEVANCE STATEMENT: This study demonstrated the effectiveness of deep learning reconstruction in improving the efficiency of 1.5-T musculoskeletal MRI exams. Despite shorter acquisition times, the visibility of key MSK structures was consistently rated as superior or similar in higher-resolution images with DLR. KEY POINTS: MRI is one of the primary diagnostic tools for evaluating MSK injuries and disorders. Deep learning reconstruction (DLR) implemented in a 1.5-T MSK protocol improved resolution, still preserving SNR and CNR. Higher scores were assigned by different raters to the DLR images, showing image quality improvement.
29 May 2026
Read appraisal →PloS one
Development and evaluation of AI model with deep learning for segmentation of extraocular muscles in thyroid eye disease
PURPOSE: To develop and evaluate an AI model for the segmentation of extraocular muscles (EOMs) using Magnetic Resonance Imaging (MRI). DESIGN: Single-center study with retrospective study. METHODS: The study included 52 patients with thyroid eye disease (TED) who underwent MRI examination of the orbital region at the Department of Ophthalmology, Osaka Metropolitan University, between October 2020 and June 2023. Manual labelling of all EOMs was performed on all slices. An AI model was created and compared with the manually labelled data using 5-fold cross validation on data sets of 12, 32, and 52 cases. The average signal intensity ratio (SIR) within the EOMs was measured. The primary outcome was the comparison of DICE similarity coefficients (DSC) as the agreement rate between AI's result and manual labelling amongst the three data sets (52 cases vs 32 cases vs 12 cases). The secondary outcome was the correlation of SIR between the AI and manual labelling. RESULTS: A significant difference in DSC was observed between the 12-, 32-, and 52-case data sets for the inferior rectus muscle and medial rectus muscle. A significant correlation was observed between the AI and manual labelling for SIR in all EOMs. CONCLUSIONS: An AI model was successfully developed for the automatic segmentation of EOMs. There was little measurement error between AI and manual labelling, and the AI model using 52 cases had improved measurement accuracy compared to the AI models using 32 and 12 cases.
29 May 2026
Read appraisal →European radiology experimental
Who labels best? Radiologists, rules, or large language models for CT reports on pulmonary embolism.
OBJECTIVE: To compare open-weight and proprietary large language models (LLMs), a rule-based extractor (RBE) and radiologists for labelling pulmonary embolism CT reports, and to test whether a hybrid RBE-LLM workflow improves labelling performance. MATERIALS AND METHODS: This single-centre retrospective study included structured CT reports from October 2021 to March 2025. Three labelling pipelines were evaluated: an RBE; a model-agnostic LLM extractor (18 open-weight, four GPT-4 variants); and a hybrid pipeline routing only RBE failures to an LLM. Ground truth was defined at the report-text level by deterministic schema matching for initially RBE-valid fields and blinded adjudication of RBE-invalid fields by two attending radiologists. Eight radiologists provided a human baseline. Outcomes included F1 scores, accuracy, LLM-based salvage of RBE failures, and labelling time. RESULTS: In total, 2,923 reports from 2,923 patients (mean age 66 ± 17 years; 1,465 women) were included. Falcon3-10b and GPT-4.1-mini achieved similar item-level performance (F1 0.98 [95% CI, 0.97-0.98] for both; p = 0.70) and both exceeded the RBE (F1 0.81 [95% CI, 0.80-0.82]; p < 0.001). Salvage of RBE failures was comparable between open-weight and proprietary models (88.1% vs 91.9%; p = 0.12). The hybrid RBE-LLM workflow achieved 99.8% accuracy and F1 0.99 (0.98-0.99), exceeding both the RBE and pooled radiologists (F1 0.92 [95% CI, 0.90-0.93]; all p < 0.001). CONCLUSION: Schema-constrained open-weight and proprietary LLMs exceeded rule-based extraction and, at the upper end of performance, matched a pooled radiologist label-transfer baseline. A rules-first, targeted LLM workflow enabled near-perfect extraction from finalised structured pulmonary embolism CT reports. RELEVANCE STATEMENT: A rules-first LLM workflow can automate high-fidelity extraction of structured CT findings from finalised radiology reports, enabling scalable, auditable, and more consistent cohort curation for clinical research, registries, and quality improvement. KEY POINTS: A hybrid rules-first workflow combining a rule-based extractor (RBE) with targeted large language model (LLM) salvage achieved the highest overall performance for labelling of pulmonary embolism CT reports (F1, 0.99; accuracy, 99.8%). The top standalone open-weight and proprietary LLMs (Falcon3-10b and GPT-4.1-mini) both exceeded the RBE and, at the upper end of performance, matched a pooled radiologist label-transfer baseline. The hybrid workflow reduced cohort-curation time from 32.2 h for radiologists to 1.0 h while reducing LLM calls by 85.6%, because the LLM was only triggered for rule-failed fields.
28 May 2026
Read appraisal →Scientific reports
Uncertainty estimation and probabilistic skull shape reconstruction using bayesian neural networks
3D shape reconstruction is an active area of research and a fundamental problem in computer vision, with growing applications in the medical domain, where it enables the recovery of missing or fine anatomical structures. Numerous approaches have been proposed for medical shape reconstruction, emphasizing accurate and anatomically plausible reconstructions. However, 3D shape reconstruction remains an inherently ill-posed problem, meaning (1) both neural network-based methods and conventional shape modeling approaches naturally introduce uncertainty in their predictions, and (2) multiple anatomically plausible reconstructions exist for a given partial or low-resolution input. While the uncertainty aspects have been widely explored in general computer vision, they remain relatively under-explored in the context of medical shape reconstruction. In this paper, we developed a 3D Bayesian U-Net and investigated its use for uncertainty estimation and probabilistic reconstructions across three key tasks: cranial reconstruction, facial reconstruction, and skull shape super-resolution. Our findings show that the Bayesian model is able to produce a range of anatomically plausible skull reconstructions while capturing natural skull variations arising from the learned weight uncertainty. Notably, these variations are primarily expressed through differences in bone thickness, which aligns with anatomical expectations, particularly relevant in real-world applications like cranial implant design. Additionally, we propose a principled framework to study the relationship between weight uncertainty and reconstruction uncertainty by analyzing the learned posterior distribution of the weights, demonstrating that our Bayesian U-Net achieves comparable reconstruction performance to a deterministic U-Net baseline while providing reliable uncertainty estimates. Our study also reveals a clear cross-task uncertainty pattern, where tasks with stronger structural constraints, like super-resolution, yield lower predictive uncertainty, while less constrained tasks, like facial reconstruction, result in higher uncertainty. Refer to the project page for more visual results https://git.zib.de/jli/uncertainty-aware-skull-reconstruction/ .
28 May 2026
Read appraisal →Journal of radiation research
Deep learning-based dose prediction in proton beam therapy for hepatocellular carcinoma: comparison of network architectures and loss functions
In proton beam therapy (PBT) for hepatocellular carcinoma (HCC), deep learning (DL)-based dose prediction offers clinical value by providing immediate reference dose distributions as a guideline tool for treatment planners and enabling virtual PBT dose assessment at institutions lacking PBT facilities to support clinical decision-making. This study proposes a dose gradient-aware DL training approach and a beam arrangement-free framework that predicts dose distributions from computed tomography (CT) images and target/organs-at-risk structures without requiring beam arrangement inputs. Using data from 172 HCC patients, we systematically compared 20 DL models combining five architectures with four loss functions, including a novel dose gradient-aware loss capturing PBT-specific dose distribution characteristics arising from the Bragg peak. Prediction accuracy was evaluated using dose metrics for clinical target volume, planning target volume, and normal liver, 10% dose region volume assessment, visual examination, Dice similarity coefficient (DSC) for 10% dose region, and dose-volume histogram (DVH) analysis. DL models employing dose gradient-aware loss demonstrated reduced prediction errors in target volume dose metrics, achieved the lowest errors in 10% dose region volume when combined with U-Net (MAE: 35.4 cm3, RMSE: 51.3 cm3), attained the highest DSC (0.88 ± 0.05), showed the closest DVH agreement with clinical plans for both target structures and normal liver. A comprehensive evaluation consistently confirmed the superiority of dose gradient-aware loss across all metrics, demonstrating the importance of physics-informed, dose gradient-aware learning for accurate PBT dose distribution prediction. Particularly, DL models combining simple architectures with dose gradient-aware loss predicted doses that closely approximated clinical doses in PBT.
27 May 2026
Read appraisal →Dental materials journal
Domain shift analysis of deep learning models for tooth detection in pediatric panoramic radiographs
The aim of this study was to externally validate object detection models for comprehensive tooth detection on pediatric panoramic radiographs, quantify the impact of domain shift across institutions and imaging protocols, and compare YOLOv8 and YOLOv10. Two datasets were used: an internal set of 200 images of early mixed dentition without bite blocks, and an external, open-source set of 192 images acquired with bite blocks. Performance was assessed using mean average precision (mAP), per‑class AP, precision‑recall curves, and confusion matrix. Overall performance was high but domain dependent. mAP was 0.958/0.941 (YOLOv8/v10) in Experiment 1; 0.910/0.906 in Experiment 2; and 0.906/0.901 in Experiment 3. Error analysis revealed domain‑specific failure modes: bite blocks were occasionally mistaken for the primary central incisor. Deep learning enables accurate and comprehensive detection of teeth on pediatric panoramic radiographs. However, domain shift alters error patterns and impairs the detection of rare yet clinically important classes.
25 May 2026
Read appraisal →World journal of urology
From conventional imaging software to artificial intelligence: tools for stone volume assessment in urolithiasis. Α review by the EAU and YAU sections of endourology.
INTRODUCTION: Accurate assessment of stone burden is fundamental in urolithiasis, as it directly influences treatment selection and prognostic evaluation. Although maximum stone diameter on non-contrast computed tomography remains the most widely used parameter, it does not adequately reflect the three-dimensional complexity of urinary calculi. This review aimed to summarize the evolution of stone burden assessment from conventional imaging-based methods to software-assisted volumetry and artificial intelligence (AI)-driven tools, with emphasis on their accuracy, reproducibility, and clinical utility. METHODS: A narrative review of the literature was performed using PubMed/MEDLINE, Scopus, and Google Scholar for English-language studies published up to March 2026. Original studies, validation studies, technical reports, reviews, and guideline-related papers addressing conventional CT-based measurement, software-assisted volumetry, AI-based stone segmentation, and the clinical significance of stone volume were included. Due to heterogeneity in study design and reported outcomes, the evidence was synthesized narratively. RESULTS: Maximum stone diameter remains simple and widely available, but it incompletely represents true stone burden, particularly in larger or irregular stones. Formula-based ellipsoid calculations are practical yet show limited accuracy in complex geometries. Semi-automated CT-based segmentation software provides more reliable volumetric assessment, with excellent agreement with reference standards and reduced interobserver variability. AI-based approaches have further improved efficiency by enabling rapid and highly accurate automated stone detection and volume calculation. Across the reviewed literature, stone volume was consistently shown to be more clinically informative than linear dimensions for predicting spontaneous passage, stone-free rates after shockwave lithotripsy and ureteroscopy, and future symptomatic events during surveillance. CONCLUSIONS: Stone volume offers a more accurate and clinically meaningful estimate of stone burden than maximum stone diameter alone. The transition from formula-based methods to software-assisted and AI-driven volumetry represents an important advance in urolithiasis imaging. Wider adoption will depend on standardized imaging protocols, improved software accessibility, and validation of volume-based thresholds for routine clinical practice.
25 May 2026
Read appraisal →Journal of medical Internet research
Alignment Between Cardiologists and AI-Driven Diagnostic Systems: Mixed Methods Study
BACKGROUND: The clinical value of artificial intelligence (AI)-based diagnostic systems depends not only on their accuracy but also on how well their outputs integrate with clinicians' judgments in practice. Critical knowledge gaps remain regarding diagnostic concordance between AI and clinicians in stress echocardiography interpretation, patient characteristics predicting discordance, and how cardiologists respond when AI recommendations conflict with their clinical diagnoses. OBJECTIVE: This study examined the diagnostic alignment between an AI-driven stress echocardiography system (EchoGo Pro [EGP]) and cardiologists' diagnoses of coronary artery disease (CAD), identified predictors of concordance and AI scan rejection, and explored cardiologists' decision-making strategies when disagreements arise. METHODS: We conducted mixed methods research. The quantitative study analyzed concordance between EGP and cardiologists using data from 854 participants with suspected CAD in the multicenter PROTEUS randomized controlled trial. Logistic regression identified predictors of agreement, disagreement, and scan rejection, adjusting for age, sex, smoking status, BMI, and cardiovascular risk factors (hypertension, hypercholesterolemia, diabetes, family history of CAD, and prior CAD events). To gain deeper insight into discordance, we conducted a qualitative study analyzing survey responses from 61 UK consultant cardiologists recruited via Qualtrics, exploring their perceptions of AI tools, the risks of following discordant AI recommendations, and their typical responses to AI-clinician disagreement. RESULTS: EGP and cardiologists agreed in 60% (512/854) of the cases, but agreement was significantly lower among patients with hypertension (OR 0.58, 95% CI 0.38-0.89; P=.01), diabetes (OR 0.56, 95% CI 0.35-0.90; P=.02), and pre-existing CAD (OR 0.48, 95% CI 0.30-0.77; P=.002). EGP rejected 26.1% (222/854) of the scans due to insufficient image quality, with rejection significantly more common in male patients (β=0.35; P=.03) and those with a family history of CAD. If a positive CAD diagnosis was assigned when either cardiologists or EGP identified CAD, the proportion of positive cases increased from 17.9% (153/854) to 22.1% (189/854), potentially identifying additional at-risk patients. Survey respondents (50/60, 85% male; 26/57, 46% aged 40-49 years; 39/61, 64% White) required 65% to 69% confidence in their initial diagnosis to justify disregarding contradictory AI recommendations. The survey findings revealed cardiologists treated AI recommendations as advisory rather than definitive. When facing discordance, they retained confidence in their judgment and sought corroboration through additional testing, data review, or second opinions rather than deferring to AI. Paradoxically, cardiologists with higher confidence in AI tools required greater confidence in their own diagnosis to disregard AI recommendations (β=7.73; P=.02). Cardiologists attributed discordance primarily to AI's inability to incorporate patient history, comorbidities, and broader clinical context. CONCLUSIONS: EGP shows promise as an adjunctive tool but struggles with multimorbid patients and exhibits high, uneven rejection rates. Cardiologists use AI to prompt scrutiny, not replace judgment. Future systems need to integrate wider patient data with imaging and minimize bias through representative training to avoid exacerbating inequities.
21 May 2026
Read appraisal →PloS one
A novel hybrid framework integrating GA-driven 3D ResUNetGAN for MRI brain tumor segmentation
Accurate brain tumor segmentation by multi-modal MRI is crucial for diagnosis, treatment planning, and prognostic assessment. This work proposes a novel hybrid architecture within a GA optimized by 3D residual U-Net (ResUNet) with generative adversarial network (GAN) for the accurate segmentation of tumor subregions, namely the tumor core (TC), whole tumor (WT), and enhancing tumor (ET), using a dataset collected from BraTS2023 (437 training and 188 validation). By using GAs, we enable the computer to automatically find the best model structure and training settings, which helps keep the model simple enough to avoid overfitting while being powerful enough to generalise well to new data. The GAN part of our method trains the model to draw more realistic and consistent boundaries around tumor regions in MRI scans, thereby improving the overall quality of the segmentation. In our experiments, this GA-ResUNetGAN approach performed better than several leading methods, such as 3D UNet, DynUNet, UNETR, SwinUNETR, and the standard ResUNet, especially in accurately identifying TC, WT, and ET parts-achieving Dice scores of 0.94, 0.96, and 0.91 on the validation set. In addition, K-Fold cross-validation was performed to ensure consistent performance across different data splits. While the method involves more processing work than some baseline techniques, it provides improved segmentation quality and boundary delineation on the evaluated dataset. These results demonstrate the advantages of combining evolutionary optimization with adversarial deep learning for challenging medical imaging tasks.
19 May 2026
Read appraisal →Radiation oncology (London, England)
Comparative study between PET-CT simulation and CT for head and neck cancer treatment and their potential effect on the thyroid gland: a prospective comparative study
BACKGROUND: Radiotherapy (RT) plays a key role in multidisciplinary treatment of Head and neck cancers (HNCs). Over the past two decades, advances have improved RT's tumor control and toxicity profiles. Precise delineation of gross tumor volume and organs at risk using CT and PET-CT is crucial for safe and effective RT. Additionally, RT increases the risk of thyroid dysfunction, especially hypothyroidism. AIM: The present study compared PET-CT simulation and CT-alone planning in head and neck cancer (HNC) treatment, emphasizing their effects on thyroid gland function. METHODS: The patients were divided into two groups according to the simulation method (PET-CT or CT), two fresh blood samples (5 mL/ sample) were collected from each patient, before and after radiotherapy. TSH, FT3, and FT4 (Thyroid biomarkers) were assessed by electrochemiluminescent immunoassay (ECLIA). RESULTS: PET-CT simulation demonstrated a significant improvement in GTVp delineation accuracy by 71.8% compared with CT. After radiotherapy, TSH levels show minor decline changes, whereas FT3 and FT4 levels declined significantly, with greater declines observed in stages II and III. The most common tumor site was the larynx (60% in PET-CT, 75% in CT), followed by the tongue, nasopharynx, and parotid gland. CONCLUSION: PET-CT simulation offered superior GTVp delineation accuracy over CT, improving HNCs radiotherapy planning. Immediately after RT, radiotherapy caused significant declines in serum FT3 and FT4 levels, with only a minor, non-significant TSH declining, particularly in stage II/III disease. Notably FT4 is little more affected in PET-CT simulation patient group, highlighting the need for thyroid function monitoring. These results advocate PET-CT integration in planning and routine endocrine monitoring for treatment complications.
19 May 2026
Read appraisal →Physics in medicine and biology
Nonperiodic dynamic CT reconstruction using backward-warping implicit neural representation with diffeomorphism regularization
Objective.Motion artifacts remain a major obstacle in dynamic computed tomography (CT) reconstruction, particularly for nonperiodic rapid motion such as cardiac imaging in patients with fast or irregular heart rates, where ECG-gated approaches are unreliable. The extreme limited-angle problem arising from insufficient angular coverage within a single cardiac phase makes accurate reconstruction of fine cardiac structures especially challenging. This work aims to develop a self-supervised framework that recovers high-resolution dynamic CT images from nonperiodic motion scenarios without external training datasets.Approach.We propose BIRD, an implicit neural representation (INR) framework for nonperiodic dynamic CT reconstruction, with three components: (1) a dual-feature representation decomposing the dynamic scene into topology-preserving and free-form features, enabling modeling of both deformable motion and residual variations such as contrast kinetics; (2) a backward-warping deformation model enabling direct ray-based training, overcoming the resolution limitations of prior forward-warping INR approaches at clinically relevant sub-millimeter resolution; and (3) a diffeomorphism-based regularization enforcing anatomically plausible deformation vector fields through bidirectional inverse consistency, without restricting representational capacity.Main results.BIRD was validated on digital and physical cardiac phantoms and retrospective patient data. In the XCAT phantom study under severe limited-angle conditions (0.5 s rot-1, 120 bpm), BIRD improved peak signal-to-noise ratio over the next-best method by 1.43 dB on the right coronary artery and 1.21 dB on the right ventricle. Curved planar reformation confirmed improved vessel continuity and boundary sharpness. On real projection data from a physical cardiac phantom and a retrospective patient scan, BIRD reduced motion artifacts and improved depiction of cardiac and vascular structures compared with conventional reconstructions.Significance.The proposed framework performs self-supervised reconstruction of nonperiodic dynamic CT images solely from projection data. It offers potential clinical applications including non-ECG-gated cardiac imaging for patients with arrhythmia, cinematic image sequences, and retrospective motion artifact correction in conventional CT scans.
19 May 2026
Read appraisal →Scientific reports
RDE-DR: robust deep ensemble CNNs for automated diabetic retinopathy detection from fundus images
Diabetic retinopathy (DR) is a leading cause of preventable blindness, motivating the development of reliable automated screening systems. This work proposes a Robust Deep Ensemble for Diabetic Retinopathy detection (RDE-DR) by analyzing ensemble fusion strategies. Four pre-trained convolutional neural networks (ResNet50, VGG16, VGG19, and DenseNet121) are trained using CLAHE-enhanced APTOS 2019 fundus images and integrated through seven heterogeneous fusion mechanisms, including voting-based, rank-based, and fuzzy-integral-inspired strategies. A consistent evaluation protocol is adopted, incorporating threshold optimization and probabilistic calibration analysis to validate robustness, decision margins, and accuracy-precision trade-offs. Experimental results show that multiple fusion techniques achieve comparable high performance and stable behavior on the APTOS 2019 benchmark, with the best configuration reaching 98.64% accuracy, 98.40% precision, 98.92% recall, 98.66% F1-score, and 99.78% Area-Under-Curve (AUC). Beyond peak accuracy, the study provides insights into ensemble reliability, calibration characteristics, and practical design choices for medical image classification systems. These results show that integrating transfer learning with CLAHE preprocessing and ensemble fusion yields stable experimental performance on the APTOS 2019 benchmark, suggesting potential for future medical decision support.
18 May 2026
Read appraisal →Drug discoveries & therapeutics
Artificial intelligence (AI)-assisted ultrasound in clinical trials: Endpoint automation, decentralized monitoring, and regulatory readiness
Ultrasound is among the most widely used imaging modalities in clinical trials, and yet its dependence on operator skill and equipment settings has historically limited the reproducibility of ultrasound-based endpoints in multi-center studies. Artificial intelligence (AI) now addresses this limitation across two complementary dimensions: automated measurement algorithms that quantify cardiac function, organ volume, and vascular parameters with reproducibility approaching or, in select settings, exceeding that of trained human readers and real-time acquisition guidance systems that enable clinicians with no formal sonography training to perform diagnostic-level examinations, making remote and decentralized assessment increasingly feasible. This narrative review synthesizes current evidence and regulatory developments across three interconnected domains. First, use of automated ultrasound to ascertain endpoints has advanced from single-institution validation to prospective and randomized evidence, with deep learning measurement of the left ventricular ejection fraction demonstrating formal equivalence to expert readers across multiple echocardiographic parameters and AI-first workflows shortening the time to diagnosis in a blinded non-inferiority trial. Second, AI-guided ultrasound acquisition by nurses and other non-expert operators has achieved a high rate of diagnostic acceptability in cardiac and pulmonary ultrasound, laying the groundwork for use of ultrasound-based endpoints in decentralized clinical trial designs, as reflected in Food and Drug Administration (FDA) guidance on decentralized trial elements. Third, the regulatory frameworks governing AI-enabled medical devices - including the U.S. FDA's Predetermined Change Control Plan guidance, the EU Artificial Intelligence Act, and internationally harmonized good machine learning practice relevant to Japan and other jurisdictions - increasingly emphasize overlapping principles such as specification of prospective performance, post-marketing oversight, and transparent reporting. Addressing remaining challenges in domain generalization across vendors, subgroup fairness, and algorithm change management during ongoing trials will be essential for AI-assisted ultrasound to fulfill its potential as a robust, scalable endpoint in clinical research worldwide.
18 May 2026
Read appraisal →Computer methods and programs in biomedicine
Geometry-aware PointNet for rapid prediction of cerebral aneurysm hemodynamics
BACKGROUND AND OBJECTIVE: Cerebral aneurysms affect 2-5% of the global population and pose a significant health risk upon rupture. While computational fluid dynamics provides detailed hemodynamic information for risk assessment, its high computational demands limit routine clinical use. This study aims to develop a deep learning model to rapidly predict three-dimensional velocity fields and wall shear stress at peak systole using aneurysm geometry. METHODS: We synthesized a dataset of 984 idealized middle cerebral artery bifurcation aneurysms. For each case, computational fluid dynamics simulations were conducted with pulsatile boundary conditions to generate ground-truth data, and peak-systolic snapshots were extracted. We developed a single-input point-cloud network augmented with a distance-to-wall feature to predict both velocity fields and wall shear stress. Model performance was evaluated using mean absolute error (MAE), normalized MAE (NMAE), and relative L2 error (rL2). RESULTS: On the test set, the model achieved velocity-field accuracy of NMAE 4.05% and rL2 19.2% over the full geometry domain, and WSS accuracy of NMAE 2.59% and rL2 23.9%. The mean inference time was approximately 1.6 seconds for velocity and 0.3 seconds for wall shear stress per case. Out-of-distribution evaluation on non-idealized geometries showed substantial zero-shot degradation (NMAE 19.1%, rL2 62.4%), while leave-one-out fine-tuning improved performance (NMAE 10.8%, rL2 37.0%). CONCLUSIONS: The proposed geometry-aware point-cloud surrogate provides fast peak-systolic hemodynamic prediction on idealized aneurysm geometries. However, out-of-distribution evaluation indicates that broader patient-specific training, physiological boundary conditions, and reliability assessment are required before routine clinical application.
17 May 2026
Read appraisal →Computers in biology and medicine
LADNET: An MRI-based deep learning approach for Alzheimer's disease detection.
Alzheimer's disease (AD) is considered a leading form of dementia, found in the majority of individuals, and it accounts for the highest prevalence of all the types of dementia. It is a gradually progressing condition which can begin with mild memory impairment and may lead to the inability to engage in meaningful conversations in the future. In addition to that, it is also capable of reacting appropriately to the environment around it. According to official data, there has been an increase in reported cases of AD-related mortality in recent years, along with rising rates of the disease itself. Therefore, it is beneficial for patients to be diagnosed with AD as early as possible in order to maximize their chance of survival. In medical field, deep learning (DL) is employed in many applications, especially the identification of Alzheimer's disease, has shown considerable success. Based on the fact that these methods are capable of acquiring and extracting features from extensive datasets autonomously, they are highly suitable for analyzing intricate medical images with the aid of these methods. Design the Lightweight Alzheimer's Disease Net (LADNET) model for the current task. Accurate detection of AD using magnetic resonance imaging (MRI)-based scans is presented in this work. The model classifies the three dementia types, namely Mild Dementia (MD), Moderate Dementia (MDD), and Very Mild Dementia (VMD), while also accurately categorizing the non-demented (ND) category based on the severity of the condition. We measure the performance of the model we recommend against a publicly available Kaggle dataset that contains over six thousand images from four different types of images. LADNET model achieves 99.4% accuracy and 99% AUC, outperforming existing methods. With a lightweight design of approximately 1.2 million parameters and fast inference (4.2 ms per image), LADNET demonstrates strong potential for practical clinical deployment, where it could help extract biomarker information from conventional MRIs, potentially reducing patient burden and diagnostic costs.
17 May 2026
Read appraisal →Computers in biology and medicine
Mapping intratumoral heterogeneity through PET-derived washout and deep learning after proton therapy
The distribution of produced isotopes during proton therapy can be imaged with Positron Emission Tomography (PET) to verify dose delivery. However, biological washout, driven by tissue-dependent processes such as perfusion and cellular metabolism, reduces PET signal-to-noise ratio (SNR) and limits quantitative analysis. In this work, we propose an uncertainty-aware deep learning framework to improve the estimation of washout parameters in post-proton therapy PET, not only enabling accurate correction for washout effects, but also mapping intratumoral heterogeneity as a surrogate marker of tumor status and treatment response. We trained the models on Monte Carlo-simulated data from eight head-and-neck cancer patients, and tested them on four additional head-and-neck and one liver patient. Each patient was represented by 75 digital twins with distinct tumoral washout dynamics and imaged 15 min after treatment, when slow washout components dominate. We also introduced washed-out maps, quantifying the contribution of medium and fast washout components to the loss in activity between the end of treatment and the start of PET imaging. Trained models significantly improved resolution and accuracy, reducing average absolute errors by 60% and 28% for washout rate and washed-out maps, respectively. For intratumoral regions as small as 5.1 mL, errors predominantly fell below thresholds for differentiating vascular status, and the models generalized across anatomical areas and acquisition delays. This study shows the potential of deep learning in post-proton therapy PET to non-invasively map washout kinetics and reveal intratumoral heterogeneity, supporting dose verification, tumor characterization, and treatment personalization. The framework is available at https://github.com/pcabrales/ppw.git.
17 May 2026
Read appraisal →Computers in biology and medicine
Imaging and machine learning reveal pigment ring spatial characteristics in vitiligo progression
BACKGROUND: Vitiligo is a chronic autoimmune disorder causing skin depigmentation due to melanocyte loss, with pigment ring dynamics being key to understanding and managing the disease. METHODS: Data from 105 vitiligo patients were collected using reflectance confocal microscopy (RCM) imaging to analyze pigment ring patterns across progressive and stable stages. A multi-task self-supervised learning framework was used to identify and classify pigment ring types (complete, depressed, incomplete, and absent). Graph-based structural analysis was applied to characterize the spatial and network dynamics of pigment rings, incorporating metrics such as clustering coefficients, centrality measures, and Ripley's statistics. RESULTS: Progressive-stage patients exhibited denser clustering and spatial propagation of absent and depressed pigment rings, which played dominant roles in the network. Stable-stage patients showed higher clustering coefficients for complete rings, indicating restored structural stability. The learning framework demonstrated high accuracy in identifying pigment ring types, with an AUC of 0.99 and PR score of 0.95. Ripley's K-function analysis highlighted significant clustering of absent and depressed rings in the progression stage, correlating with active disease. CONCLUSION: Integrating imaging, machine learning, and graph-based analysis offers key insights into vitiligo progression, enabling better staging and personalized treatments, with future research needed for broader validation and model refinement.
16 May 2026
Read appraisal →Physics in medicine and biology
A novel deep-learning approach for monitoring gastrointestinal air variation during radiotherapy in young patients using radiographs
Objective.To develop deep-learning (DL) predictive models of internal gastrointestinal (GI) gas from radiographs, a surrogate for dose degradation in abdominal proton beam therapy (PBT).Approach.DL architectures (convolutional neural networks and U-Nets) were trained to predict GI air local path length and/or global volume from radiographs. Frontal/lateral radiographs with paired volume and path length labels were simulated from CTs and GI gas segmentations using the Tomographic Iterative GPU-based Reconstruction (TIGRE) toolbox. CTs from The Cancer Imaging Archive Paediatric-CT-SEG dataset (N= 349, 0-16 year, North America) were used for model development and internal evaluation; variation in GI filling, organ motion and body size were simulated for augmentation. External evaluation utilised data from two international radiotherapy institutions: one consisting mostly of children and young people (CYP) (N= 21, 2-19 year, Europe); the other predominantly of teenagers and young adults (TYAs) (N= 14, 10-24 year, Asia). Finally, real-life use case was demonstrated on two subjects treated with PBT.Main results.The optimal configuration was the dual U-Net, accurately predicting both volume (mean absolute error (MAE) = 24.0 ± 32.9 ml) and path length (MAE = 0.80 ± 0.59 mm) on the held-out Paediatric-CT-SEG dataset, while also generalising well to external radiotherapy datasets (volume MAE = {18.9 ± 22.4, 34.6 ± 35.2} ml, and path length MAE = {0.52 ± 0.38, 0.67 ± 0.55} mm on the CYP and TYA datasets, respectively). We found a lower Dice similarity coefficient for gas segmentations in TYA vs Paediatric-CT-SEG and CYP (0.56 ± 0.24 vs 0.87 ± 0.09 and 0.84 ± 0.16, respectively), reflecting cohort-specific gas morphology differences. When applied longitudinally, the network detected both monotonic and realistic intrasubject GI gas variation. Proof-of-concept evaluation on clinical radiographs showed promising performance (volume MAE = 29.5 ± 25.9 ml, and path length MAE = 1.10 ± 0.75 mm).Significance.DL-powered radiographs quantified internal GI gas changes with promising performance and generalisability to diverse ages and demographics. This was the first step towards a novel personalised image-guided radiotherapy traffic-light workflow for abdominal PBT with the goal of reducing the need for CBCT to required fractions.
14 May 2026
Read appraisal →Scientific reports
Deep learning-based automatic field of view planning for prostate MRI in oblique coronal and oblique axial planes
Accurate field-of-view (FoV) prescription in oblique coronal and axial planes is essential for high-quality prostate MRI but remains operator-dependent and variable. We developed and evaluated a ResNet-based deep learning framework for automated FoV planning. In this retrospective multicenter study, FoV prescriptions were annotated on PI-CAI dataset. Three readers assessed intra- and inter-rater variability to establish reference consistency. Three neural network variants were trained on 1,474 examinations from PI-CAI dataset (2012-2021), and the optimal model was selected by internal validation. Generalizability and clinical utility were tested on three external cohorts totaling 530 examinations (2021-2024) using a non-inferiority design. The selected model achieved non-inferior performance for slice positioning, with differences ranging from 0.16 ± 0.99 to 0.37 ± 0.48. Across sites, FoV overlaps ranged from 82.4 ± 4.1% to 88.7 ± 6.0%, and the angle differences between predicted and reference planes were 4.66 ± 4.89° (Site I), 3.46 ± 2.80° (Site II), and 2.99 ± 2.90° (Site III). Clinical utility was high at all sites, with acceptability rates of 97.9%, 97.7%,98.8%, 98.1% and 98.1% for Site I (Raters 1-5), 95.7%, 97.8%, 100%, 95.7% and 97.8% for Site II (Raters 1-5), and 100% for all raters at Site III. These findings demonstrate the feasibility of automated FoV positioning for prostate MRI and indicate excellent clinical utility.
14 May 2026
Read appraisal →Scientific reports
Federated learning with swarm intelligence for efficient and secure medical image analysis
Collaborative learning in healthcare faces challenges, including strict regulations and fragmented data. This research introduces a federated learning framework that employs swarm intelligence to augment communication and enhance the analysis of medical images. The method optimizes hyperparameters, selects features, and assigns aggregation weights to federated clients simultaneously by combining Particle Swarm Optimization (PSO) and the Firefly Algorithm (FA) with deep Convolutional Neural Networks (CNNs). The framework was tested on three medical datasets: COVID-19 chest X-rays (5,856 images), monkeypox skin images (569 images), and breast cancer mammograms (320 images). These datasets were shared among four fake healthcare institutions. It strives to strike a balance between privacy, communication costs, and classification accuracy. The results showed that the test was 96.71% accurate in detecting COVID-19, 96.06% accurate in classifying monkeypox, and 97.0% accurate in diagnosing breast cancer. The framework was able to handle noise and attacks from individuals who sought to disrupt it, which reduced communication rounds by 25-30%. A privacy-utility analysis revealed that there were acceptable trade-offs, with accuracy remaining above 94%. This study employs robust privacy measures and statistical validation. It also shows how to use medical AI in smaller healthcare settings without putting patients' privacy at risk.
14 May 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Deep-learning endomicroscope with large field-of-view and depth-of-field for real-time in vivo imaging of epithelial cancer hallmarks
In vivo microscopy (IVM) has shown great promise to improve early detection of epithelial precancer, but it suffers from fundamental trade-offs that limit the resolution, field-of-view (FOV) and depth-of-field (DOF). Here, we present PrecisionView, a compact, deep learning-enabled endomicroscope that breaks these constraints and achieves 20 mm2 FOV and 500 µm DOF with 4 µm resolution, representing approximately 5× increase in FOV and 8× larger DOF compared to conventional IVM with similar resolution. PrecisionView integrates a deep learning-optimized phase mask and real-time reconstruction, enabling rapid in vivo assessment of two key hallmarks of cancer: epithelial cell nuclear morphology and subsurface microvasculature through fluorescence and reflectance imaging. By imaging the oral cavity of healthy volunteers and cervical specimens with precancerous lesions, PrecisionView generates large-scale (1 to 3 cm2) coregistered maps of cellular and vascular structures, revealing distinct microscopic patterns associated with anatomic structures and precancerous lesions. Our results suggest the potential of this computational endomicroscope to address the unmet need for early cancer detection at the point of care.
13 May 2026
Read appraisal →Neuroinformatics
Instance-Aware Pseudo-Labeling and Class-Focused Contrastive Learning for Weakly Supervised Domain Adaptive Segmentation of Electron Microscopy
Annotation-efficient segmentation of the numerous mitochondria instances from various electron microscopy (EM) images is highly valuable for biological and neuroscience research. Although unsupervised domain adaptation (UDA) methods can help mitigate domain shifts and reduce the high costs of annotating each domain, they typically have relatively low performance in practical applications. Thus, we investigate weakly supervised domain adaptation (WDA) that utilizes additional sparse point labels on the target domain, which require minimal annotation effort and minimal expert knowledge. To take full use of the incomplete and imprecise point annotations, we introduce a multitask learning framework that jointly conducts segmentation and center detection with a novel cross-teaching mechanism and class-focused cross-domain contrastive learning. While leveraging unlabeled image regions is essential, we introduce segmentation self-training with a novel instance-aware pseudo-label (IPL) selection strategy. Unlike existing methods that typically rely on pixel-wise pseudo-label filtering, the IPL semantically selects reliable and diverse pseudo-labels with the help of the detection task. Comprehensive validations and comparisons on challenging datasets demonstrate that our method outperforms existing UDA and WDA methods, significantly narrowing the performance gap with the supervised upper bound. Furthermore, under the UDA setting, our method also achieves substantial improvements over other UDA techniques.
13 May 2026
Read appraisal →Acta oncologica (Stockholm, Sweden)
Improving confidence in MRI-based auto-segmentation via uncertainty assessment.
BACKGROUND AND PURPOSE: Accurate delineation of organs of interest (OOIs, also commonly referred to as organs at risk, OARs) is crucial for safe radiotherapy. While deep learning-based segmentation using convolutional neural networks has achieved high geometric accuracy, clinical translation is hindered by overconfident, uncalibrated predictions in anatomically ambiguous regions. Uncertainty quantification and model calibration are prerequisites for safe clinical workflows. This study compared the standard nnU-Netv2 against its residual-encoding variant (ResEncM), hypothesizing that ResEncM would demonstrate superior reliability and calibration while maintaining comparable geometric accuracy. Patient/material and methods: T1-weighted contrast-enhanced MRI scans from 70 brain cancer patients were used (55 training/validation, 15 testing). Ground-truth contours for brainstem, hippocampi, chiasm, optic nerves, optic tracts, and pituitary were delineated per Danish Neuro Oncology Group guidelines. Both architectures were trained using five-fold cross-validation with identical preprocessing. Epistemic uncertainty was quantified using mutual information, and Expected Calibration Error (ECE) was computed within a 10-mm isotropic margin around reference contours. RESULTS: Both models achieved high geometric accuracy (brainstem dice similarity coefficient [DSC] > 0.93, hippocampi DSC > 0.81). No significant geometric differences were found for large structures. ResEncM showed significantly lower DSC for the pituitary (p = 0.003) and chiasm (p = 0.018). However, ResEncM demonstrated significantly lower epistemic uncertainty and ensemble variance across all structures (p < 0.05), and significantly reduced ECE for the optic chiasm, optic tracts, and pituitary. INTERPRETATION: Integrating a deep residual encoder into the standard U-Net framework significantly improves reliability and calibration of automated brain OOI contours while maintaining strong geometric performance. The ResEncM architecture provides a more trustworthy tool for clinical radiotherapy by reliably flagging high-uncertainty voxels, supporting confidence-aware clinical workflows.
13 May 2026
Read appraisal →Physics in medicine and biology
A delta-based multi-omics model to predict adaptive radiotherapy benefit in patients with non-small cell lung cancer during carbon-ion radiotherapy
Objective.Patients with non-small cell lung cancer (NSCLC) treated with carbon-ion radiotherapy (CIRT) are potential candidates for adaptive radiotherapy (ART). However, ART is time- and resource-intensive. This study aimed to develop a decision-support tool to estimate ART demand before treatment.Approach.This retrospective study included 69 patients with NSCLC treated with CIRT. Baseline clinical characteristics, dose-volume histogram (DVH) parameters, and radiomic and dosiomic features from the plan-computed tomography (CT) (1-2 weeks before treatment) and confirm-CT (1-2 d before treatment) were collected, and relative change rates in features (ΔDVH, Δradiomics, and Δdosiomics) were calculated. Feature selection involved univariable testing, correlation analysis, least absolute shrinkage and selection operator regression, and stepwise regression with 1000 bootstrap resamples. A multivariable logistic regression model was constructed and validated using 200 iterations of five-fold cross-validation.Main results.ART was triggered in 32 (46.4%) patients. A five-feature delta-omic model (one ΔDVH, two Δradiomic, and two Δdosiomic features) achieved a validated median area under the curve of 0.897 (interquartile range, 0.833-0.950) with good calibration. At a high-sensitivity threshold (0.11), the model identified all ART candidates while potentially reducing the frequency of weekly-CT evaluations in 21.4% of patients; at the maximum-Youden threshold (0.41), the median accuracy reached 78.6%, with up to 50.0% classified as low ART demand.Significance.A delta-omic model predicting ART demand in patients with NSCLC receiving CIRT was developed. The model may help optimize resource allocation and streamline ART workflows, although prospective validation in larger cohorts is warranted before widespread clinical adoption.
13 May 2026
Read appraisal →Biomedical physics & engineering express
Vitiligo state assessment based on progressive transfer learning and multimodal domain adaptation
Vitiligo is a common skin depigmentation disorder; assessing its state is crucial for the treatment outcome. Collecting multimodal data for vitiligo assessment is complex and costly in clinical practices, and the limited data size restricts the performance of deep learning models. Transfer learning can alleviate the shortage of training data in medical image recognition, however, its applications in vitiligo state assessment are constrained by feature differences between natural and medical images and insufficient generalization to different modalities. To address these challenges, this paper introduces a vitiligo state assessment method based on progressive transfer learning and multimodal domain adaptation. The scheme uses a large set of unlabeled medical images as a bridge to reduce the discrepancies between natural and medical images through multi-step fine-tuning. An adaptive parameter unfreezing strategy is then applied for the accurately labeled target data to improve the adaptability and accuracy of the model. In addition, an integrated multimodal domain adaptation approach based on CycleGAN and hue, saturation, value color space transformation is proposed to reduce the impact of modality differences on transfer learning. Experimental results demonstrate that compared to the conventional transfer learning method, the proposed method improves the accuracy by 2.4% and 3.6% on the clinical and Wood's lamp modality datasets, respectively. This accurate vitiligo state assessment method can be applied to a wide range of multimodal dermatological diseases where labeled data is limited.
13 May 2026
Read appraisal →European radiology experimental
Reconstruction algorithms and arm positioning effects on abdominal CT image quality and radiation dose: a phantom study
OBJECTIVE: To evaluate the effects of arm positioning and reconstruction algorithms on radiation dose and image quality of abdominal CT. MATERIALS AND METHODS: Abdominal CT scans were performed using a customized body phantom at 100 kVp with automatic tube current modulation based on noise indices (NI) of 9, 11, and 13 across seven arm positions: arms up (AU), arms down (AD), arms down with one or two-layered air cushions (ADAC, ADAC2), arms on the belly (AB), and arms on the belly with one or two-layered air cushions (ABAC, ABAC2). Images were reconstructed using filtered back projection (FBP), iterative reconstruction, and deep learning-based reconstruction (DLIR). Radiation dose was recorded. Quantitative image quality metrics included image noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), sharpness, and structural similarity. Qualitative assessment of sharpness, noise, artifacts, and overall image quality was performed using fixed-dose AU-FBP images as the reference standard. RESULTS: Compared with the AU position, radiation dose increased by up to 83% in non-AU configurations, with similar trends across all NI settings. In contrast, air cushion use caused only minor, configuration-dependent dose changes. Qualitatively, non-AU FBP images showed marked degradation, whereas DLIR enabled selected non-AU configurations to approach near-reference diagnostic quality. Air cushion-related benefits were modest overall but more apparent in AB than AD configurations. CONCLUSION: Arm positioning is the primary determinant of radiation dose and image quality in abdominal CT. When arm elevation is not feasible, DLIR effectively mitigates image quality degradation, while air cushions provide limited, configuration-dependent benefits. RELEVANCE STATEMENT: When arms cannot be raised during abdominal CT, avoiding arms-on-belly and using air cushions may help maintain image quality without substantially increasing radiation dose, while deep learning image reconstruction further enhances diagnostic performance under restricted positioning conditions. KEY POINTS: Arm positioning is a major determinant of radiation dose and image quality in abdominal CT. Arms-on-belly positioning produces poorer image quality despite increased radiation dose. DLIR improves image quality across arm positions but does not eliminate positioning-related degradation.
12 May 2026
Read appraisal →Biomedical physics & engineering express
Enhanced ResU-Net for brain tumor segmentation using EfficientNetB0, channel attention, and ASPP
Accurate and efficient segmentation of brain tumors is critical for diagnosis, treatment planning, and monitoring in clinical practice. In this study, we present an enhanced ResU-Net architecture for automatic brain tumor segmentation, integrating an EfficientNetB0 encoder, a channel attention mechanism, and an atrous spatial pyramid pooling (ASPP) module. The EfficientNetB0 encoder leverages pre-trained features to improve feature extraction efficiency, while the channel attention mechanism enhances the model's focus on tumor-relevant features. ASPP enables multi-scale contextual learning, which is crucial for handling tumors of varying sizes and shapes. The proposed model was evaluated on two benchmark datasets: The Cancer Genome Atlas Low Grade Glioma and brain tumor segmentation (BraTS-2020). Experimental results demonstrate that our method consistently outperforms the baseline ResU-Net and its EfficientNet variant, achieving dice similarity coefficient of 0.903 and 0.851, and HD95 scores of 9.43 and 3.54 for whole tumor and tumor core (TC) regions on the BraTS 2020 dataset, respectively. Compared to state-of-the-art methods, our approach shows competitive performance, particularly in whole tumor and TC segmentation. These results indicate that combining a powerful encoder with attention mechanisms and ASPP can significantly enhance BraTS performance. The proposed approach holds promise for further optimization and application in other medical image segmentation tasks.
11 May 2026
Read appraisal →Biomedical physics & engineering express
SAM2HIPT: a hybrid deep learning framework integrating SAM2 and HIPT with joint loss optimization for immunohistochemical cell nucleus segmentation
Nucleus segmentation in immunohistochemistry images plays a critical role in cancer diagnosis and treatment assessment. However, existing methods remain limited in segmentation accuracy and boundary delineation due to staining heterogeneity, densely packed cell distributions, and complex background interference. To address these challenges, this paper proposes a two-stage nucleus segmentation framework, termed SAM2HIPT. In the first stage, the pre-trained Segment Anything Model 2 (SAM2) is employed to generate initial segmentation predictions for input images, wherein the image encoder is kept frozen to preserve the pre-trained visual representation capacity while the mask decoder is fine-tuned to adapt to the characteristics of the pathological image domain; local texture, morphological, and boundary information are extracted through visual feature encoding to produce initial nucleus segmentation masks and spatial prior representations. In the second stage, the Hierarchical Image Pyramid Transformer is introduced to refine the initial segmentation results, performing multi-scale, multi-level feature representation and fusion of morphological, textural, and spatial structural information through a hierarchical vision Transformer architecture, thereby enhancing nuclear structural representation and boundary consistency. To enable collaborative optimization across both stages, a joint loss function is designed to impose unified constraints on segmentation accuracy and feature representation. Evaluated on two public histopathological benchmark datasets, BCData and DeepLIIF, the proposed method achieves Dice coefficients of 0.92 and 0.91, respectively, and HD95 boundary error values of 1.05 pixels and 1.10 pixels, demonstrating superior segmentation performance and robustness over multiple state-of-the-art baseline methods.
11 May 2026
Read appraisal →AJNR. American journal of neuroradiology
Predicting Intracranial Aneurysm Rupture Risk and Intervention Outcomes Using Denoising-Enhanced CTA
BACKGROUND AND PURPOSE: Our aim was to develop a comprehensive multimodal framework for assessing the rupture risk of intracranial aneurysms and predicting intervention outcomes. In addition, it seeks to a novel denoising algorithm to enhance the quality of CTA images, thereby improving morphologic profiling. MATERIALS AND METHODS: This retrospective multicenter study included 352 patients with intracranial aneurysms who underwent CTA. A multimodal framework was developed, integrating 3 complementary feature sets: clinical variables, radiomic texture features, and deep learning-derived aneurysm morphologic data. A novel denoising algorithm was applied to improve image quality, thereby enhancing prediction performance. Model validation was performed through cross-validation, using multiple end points, including the Hunt-Hess, World Federation of Neurosurgical Societies (WFNS), and mRS grading systems. RESULTS: The multimodal framework demonstrated robust performance, achieving an area under the curve (AUC) of 0.896 [0.819-0.973] for aneurysm rupture prediction, outperforming conventional single-technique models (radiomics-based model: 0.752 [0.693-0.809]; deep learning-based model: 0.823 [0.789-0.827]). Incorporating the denoising technique further enhanced performance, with the AUC for rupture prediction increasing to 0.908 [0.836-0.981]. In clinical grading tasks, the model showed strong efficacy, achieving an AUC of 0.907 [0.845-0.968] for Hunt-Hess grading, 0.883 [0.662-0.988] for World Federation of Neurosurgical Societies grading, and 0.926 [0.879-0.973] for mRS. CONCLUSIONS: Our system demonstrates promising performance in predicting aneurysm rupture and clinical grading assessments, indicating its potential for comprehensive aneurysm evaluation. Moreover, the proposed denoising method effectively mitigates noise interference, enhances morphologic edge features, and improves the accuracy of existing models.
11 May 2026
Read appraisal →EuroIntervention : journal of EuroPCR in collaboration with the Working Group on Interventional Cardiology of the European Society of Cardiology
Derivation and external validation of a deep learning model to predict changes in coronary plaque burden
BACKGROUND: Predicting the progression/regression of coronary plaque burden is challenging. AIMS: We aimed to develop a deep learning model to forecast changes in percent atheroma volume (ΔPAV) using intravascular ultrasound (IVUS). METHODS: We analysed data from IBIS-4 and PACMAN-AMI. Core lab measurements of plaque burden were available from IVUS pullbacks. Each model consists of a bidirectional Long Short-Term Memory (biLSTM) layer followed by two fully connected layers with one neuron each, resulting in both a classification for input progression/regression and an estimation of the ΔPAV. RESULTS: For the derivation and validation, a total of 1,960 regions of interest (ROIs) from the IBIS-4 dataset were used. The mean±standard deviation of the model accuracy was 0.85±0.02, the Matthews correlation coefficient was 0.70±0.04, and the F1 score was 0.85±0.02 for both progression and regression classes. In the testing (external validation) process with the PACMAN-AMI dataset, 5,283 ROIs were utilised. The mean ΔPAV was -0.31±5.63, for which 2,665 featured regression with a mean ΔPAV of -4.57±3.73, and 2,618 presented progression with a mean ΔPAV of 4.02±3.55, representing 49.6% of plaque progression prevalence. The predictive performance across the 100 trained models in the testing dataset showed an accuracy of 0.84, a Matthews correlation coefficient of 0.68, and an F1 score for the progression and regression classes of 0.84. CONCLUSIONS: This is the first deep learning model capable of detecting changes in plaque progression by analysing the rate of plaque burden change between adjacent frames.
10 May 2026
Read appraisal →Biomedical physics & engineering express
Enhancing CNN regressors with contour encoding and self-supervision for improved 3D/2D x-ray to CT registration in spinal surgery navigation
With advances in deep learning, regression-based methods have shown promising results in 3D/2D medical image registration. However, strict intraoperative radiation dose constraints produce low-dose x-ray images with severe blur and reduced contrast, significantly degrading registration accuracy and limiting precise image-guided spinal interventions. We propose the Contour Encoding Regressor (CER), a novel end-to-end CNN framework that extracts highly discriminative features directly from binary contour masks of intraoperative x-rays without any restrictions on contour length, shape, or morphology. These contour features are efficiently encoded by a dedicated module and seamlessly fused into the regression pipeline to improve robustness against image degradation. To further enhance pose estimation, CER employs a dual-branch architecture that explicitly decouples rotational and translational parameters, thereby reducing mutual interference and improving overall accuracy. In addition, a self-supervised fine-tuning strategy with a tailored multi-component loss function is introduced to adapt the model to blurred low-dose conditions and minimize residual errors. On low-dose x-ray images, CER achieves a mean target registration error of 1.39 mm-a clinically acceptable threshold-while outperforming state-of-the-art methods in accuracy and enabling real-time performance (0.03-0.06 s per frame on clinically accessible GPUs). These improvements meet the stringent precision and speed requirements of intraoperative navigation, offering strong potential to enhance surgical safety and outcomes in minimally invasive spinal procedures.
10 May 2026
Read appraisal →Diagnostic and interventional radiology (Ankara, Turkey)
Foundation models for radiology: fundamentals, applications, opportunities, challenges, risks, and prospects.
Foundation models (FMs) represent a significant evolution in artificial intelligence (AI), impacting diverse fields. Within radiology, this evolution offers greater adaptability, multimodal integration, and improved generalizability compared with traditional narrow AI. Utilizing large-scale pre-training and efficient fine-tuning, FMs can support diverse applications, including image interpretation, report generation, integrative diagnostics combining imaging with clinical/laboratory data, and synthetic data creation, holding significant promise for advancements in precision medicine. However, clinical translation of FMs faces several substantial challenges. Key concerns include the inherent opacity of model decision-making processes, environmental and social sustainability issues, risks to data privacy, complex ethical considerations, such as bias and fairness, and navigating the uncertainty of regulatory frameworks. Moreover, rigorous validation is essential to address inherent stochasticity and the risk of hallucination. This international collaborative effort provides a comprehensive overview of the fundamentals, applications, opportunities, challenges, and prospects of FMs, aiming to guide their responsible and effective adoption in radiology and healthcare.
10 May 2026
Read appraisal →Medical physics
Spectral deep learning-based patient and bowtie scatter correction for clinical photon-counting CT
BACKGROUND: The presence of scatter in computed tomography degrades image quality, and can be caused by the patient and by other components in the beam path, such as the bowtie filter. While conventional energy-integrating detectors do not provide spectral distinction, photon-counting (PC) detectors are energy-selective and provide spectral information about the incoming X-ray photons. Since each energy threshold is affected differently by scatter, this spectral information implicitly encodes the scatter content of a projection. PURPOSE: The purpose of this work is to investigate how the spectral information can be exploited to improve deep learning (DL)-based scatter correction. Furthermore, the performance of joint and separate patient and bowtie scatter correction will be investigated, addressing that bowtie scatter has not been considered in current DL-based approaches. METHODS: We present a DL-based approach that can estimate bowtie and patient scatter jointly and compare it against a separate correction. We also introduce neural network-based methods that incorporate the spectral information inherent in PCCT for scatter correction. We present networks that estimate scatter for up to four energy thresholds simultaneously. Training and validation was performed with Monte Carlo data as well as with real data measured by a clinical PCCT system. RESULTS: When comparing joint and separate patient and bowtie scatter estimation, both methods reduce the mean absolute error (MAE) from 8 HU to 1 HU. All proposed DSE methods effectively reduce scatter artifacts and perform better than the convolution-based reference approach. Incorporating the spectral information further improves the performance, with the DSE variant with four energy thresholds achieving the best overall results for all thresholds. For all energy thresholds tested, the spectral DSE methods reduced scatter errors originating from the patient and the bowtie in PCCT from up to 8 HU to below 1 HU. In addition to the global MAE, we report a critical MAE (MAE10) restricted to voxels with uncorrected errors > $>$ 10 HU, as such deviations are visually perceptible in soft tissue and exceed the noise level of modern CT systems. In all test cases, the proposed spectral methods reduced the MAE10 from 23.8 HU in the uncorrected images to 1.6 HU after spectral correction. The affected voxels comprised on average 25 % of the image volume, indicating a significant reduction in artifact intensity in the most affected areas. In virtual monoenergetic images (VMI), the application of spectral neural networks resulted in a significant reduction in MAE from ∼ $\sim$ 16 HU to ∼ $\sim$ 2 HU at 45 keV, from ∼ $\sim$ 8 HU to ∼ $\sim$ 1 HU at 70 keV, and from 5 HU to under 1 HU at 100 keV. CONCLUSIONS: This paper presents a combined method for correcting patient and bowtie scatter that delivers results equivalent to separate corrections and thus eliminates the need for multiple networks. Further, we demonstrate that deep scatter estimation can effectively exploit the spectral information available to improve scatter correction, especially for spectral applications, like VMIs. Spectral DSE networks slightly outperformed non-spectral variants, with multiple energy thresholds lead to more accurate estimations. This enables the use of one network for scatter correction and eliminates the need for multiple ones, thereby saving computational cost and complexity.
10 May 2026
Read appraisal →Medical image analysis
Context-enriched contrastive auto-encoder with topology learning for medical hyperspectral image classification to diagnose tumors
Deep learning has emerged as a highly effective approach for the automatic classification of medical hyperspectral images (MedHSIs), facilitating the accurate diagnosis of diverse tumors. Most of current methods suffer from the challenges in the separability of latent representation and the preservation of topological structure. To remedy such deficiencies, we propose a novel context-enriched contrastive auto-encoder with topology learning (CCAET) based MedHSI classification for tumor diagnosis. Concretely, we initially employ an auto-encoder (AE) network as the foundational framework to achieve latent representation and data reconstruction. Subsequently, to enhance the separability of the latent representation, we integrate a context-enriched contrastive loss into the AE framework, where both individual pixels and their contextual information are used to redefine contrastive learning, so that small distances between intra-class features and large distances between inter-class features are achieved. Further, we construct a topology-preserved loss by incorporating a graph based approach into the foundational framework, which aims to maintain the topological structure during latent representation learning. Furthermore, we devise a gradient descent-based optimization scheme to optimize the overall loss of the proposed CCAET method, so as to obtain the final latent representation of MedHSI. Finally, we implement soft-max-based label prediction to obtain the classification results. Different from most existing methods, the CCAET approach considers both the separability of latent representations and the preservation of topological structure to enhance MedHSI classification performance. Extensive experimental evidence has been presented to substantiate that the proposed CCAET method outperforms several state-of-the-art techniques in terms of MedHSI classification based tumor diagnosis.
9 May 2026
Read appraisal →Medical image analysis
MICCAI STS 2024 challenge: Semi-supervised instance-level tooth segmentation in panoramic X-ray and CBCT images
Orthopantomogram (OPGs) and Cone-Beam Computed Tomography (CBCT) are vital for dentistry, but creating large datasets for automated tooth segmentation is hindered by the labor-intensive process of manual instance-level annotation. This research aimed to benchmark and advance semi-supervised learning (SSL) as a solution for this data scarcity problem. We organized the 2nd Semi-supervised Teeth Segmentation (STS 2024) Challenge at MICCAI 2024. We provided a large-scale dataset comprising over 90,000 2D images and 3D axial slices, which includes 2380 OPG images and 330 CBCT scans, all featuring detailed instance-level FDI annotations on part of the data. The challenge attracted 114 (OPG) and 106 (CBCT) registered teams. To ensure algorithmic excellence and full transparency, we rigorously evaluated the valid, open-source submissions from the top 10 (OPG) and top 5 (CBCT) teams, respectively. All successful submissions were deep learning-based SSL methods. The winning semi-supervised models demonstrated impressive performance gains over a fully-supervised nnU-Net baseline trained only on the labeled data. For the 2D OPG track, the top method improved the Instance Affinity (IA) score by over 44 percentage points. For the 3D CBCT track, the winning approach boosted the Instance Dice score by 61 percentage points. This challenge demonstrates the potential benefit benefit of SSL for complex, instance-level medical image segmentation tasks where labeled data is scarce. The most effective approaches consistently leveraged hybrid semi-supervised frameworks that combined knowledge from foundational models like SAM with multi-stage, coarse-to-fine refinement pipelines. Both the challenge dataset and the participants' submitted code have been made publicly available on GitHub (https://github.com/ricoleehduu/STS-Challenge-2024), ensuring transparency and reproducibility.
9 May 2026
Read appraisal →Reproductive biomedicine online
AI-driven paradigm shift in follicle ultrasound monitoring: from automated segmentation to clinical decision support
This commentary delineates the developmental pathway of artificial intelligence (AI) in ultrasound follicular monitoring, highlighting a paradigm shift from automated segmentation to clinical decision support. The deep learning-based CR-Unet and C-Rend models have enabled accurate follicle segmentation and measurement from two-dimensional to three-dimensional imaging, substantially improving boundary segmentation accuracy and measurement consistency. Building on this foundation, the study further establishes two-dimensional follicle area and three-dimensional follicle volume as novel biomarkers, providing quantitative criteria for predicting oocyte maturity and optimizing the timing of human chorionic gonadotrophin triggering. Through seamless integration of algorithms into the Acclarix LXK9 ultrasonography equipment, an intelligent monitoring platform with real-time analytical capabilities has been developed, demonstrating significantly superior measurement accuracy and consistency compared with manual operations. These advancements represent a transformative leap from image segmentation to AI-driven clinical decision making, offering robust technical support for standardized and precise management in assisted reproduction.
9 May 2026
Read appraisal →Scientific reports
From intelligent models to clinical tools: the evolving landscape of AI in medical imaging
Integration of Artificial Intelligence (AI), particularly deep learning, into medical imaging represents a profound shift in diagnostic medicine, moving from purely descriptive analysis to advanced predictive and prescriptive analytics. This Collection explores the rapid advancement of AI-driven tools in their specific fields such as oncology, cardiology, ophthalmology and so on, highlighting their potential to improve diagnostic accuracy, workflow efficiency, and personalized treatment planning. However, significant challenges remain, including the heterogeneity of medical image data, the "black box" nature of some intelligent models, and the critical hurdles of clinical integration and validation. The research presented here addresses these frontiers, showcasing innovations in algorithm development, explainable AI, and translational application. This Editorial synthesizes the contributions and outlines the essential collaborative pathway-uniting computer scientists, clinicians, and regulatory bodies-required to translate algorithmic promise into robust, trustworthy, and equitable clinical tools that genuinely improve patient care.
9 May 2026
Read appraisal →Medical image analysis
Diversity-driven MG-MAE: Multi-granularity representation learning for non-salient object segmentation
Masked Autoencoders (MAEs) have grown increasingly prominent as a powerful self-supervised learning paradigm. They are capable of effectively leveraging inherent image prior information and are gaining traction in the field of medical image analysis. However, their application to feature representations of the non-salient objects, such as microvasculature, accessory organs, and early-stage tumors-is fundamentally limited by dimensional collapse problem, which diminishes feature diversity critical for non-salient structure discrimination. To address this, we propose a Multi-Granularity Masked Autoencoder (MG-MAE) framework for feature diversity learning: (1) We extend the conventional MAE into a multi-granularity framework, a global branch reconstructs global pixels, with a local branch recovering Histogram of Oriented Gradients (HOG) features, enabling hierarchical representation of both coarse-grained and fine-grained patterns; (2) Critically, in the local branch, a diversity-enhanced loss function incorporating Nuclear Norm Maximization (NNM) constraint to explicitly mitigate feature space collapse through orthogonal embedding regularization; and (3) A Dynamic Weight Adjustment (DWA) strategy that dynamically prioritizes hard-to-reconstruct regions via entropy-driven gradient modulation. Comprehensive evaluations across five clinical benchmarks-CCTA139, BTCV, LiTS, ACDC, and MSD Pancreas Tumour datasets-demonstrate that MG-MAE achieves statistically significant improvements in Dice Similarity Coefficient (DSC) scores for non-salient object segmentation, outperforming state-of-the-art methods. The code is available at https://github.com/zhangbbin/mgmae.
9 May 2026
Read appraisal →Canadian Association of Radiologists journal = Journal l'Association canadienne des radiologistes
Training With Local Data Remains Important for Deep Learning MRI Prostate Cancer Detection
OBJECTIVES: Domain shift has been shown to have a major detrimental effect on AI model performance however prior studies on domain shift for MRI prostate cancer segmentation have been limited to small, or heterogenous cohorts. Our objective was to assess whether prostate cancer segmentation models trained on local MRI data continue to outperform those trained on external data with cohorts exceeding 1000. METHODS: We simulated a multi-institutional consortium using the public PICAI dataset (PICAI-TRAIN: 1241 exams, PICAI-TEST: 259) and a local dataset (LOCAL-TRAIN: 1400 exams, LOCAL-TEST: 308). IRB approval was obtained and consent waived. We compared nnUNet-v2 models trained on the combined data (CENTRAL-TRAIN) and separately on PICAI-TRAIN and LOCAL-TRAIN. Accuracy was evaluated using the open source PICAI Score on LOCAL-TEST. Significance was tested using bootstrapping. RESULTS: Just 22% (309/1400) of LOCAL-TRAIN exams would be sufficient to match the performance of a model trained on PICAI-TRAIN. The CENTRAL-TRAIN performance was similar to LOCAL-TRAIN performance, with PICAI Scores [95% CI] of 65 [58-71] and 66 [60-72], respectively. Both of these models exceeded the model trained on PICAI-TRAIN alone which had a score of 58 [51-64] (P < .002). Reducing training set size did not alter these relative trends. CONCLUSION: Domain shift limits MRI prostate cancer segmentation performance even when training with over 1000 exams from 3 external institutions. Use of local data is paramount at these scales.
9 May 2026
Read appraisal →Nuclear medicine communications
Refining deep learning segmentation in gallium-68-prostate-specific membrane antigen-11 positron emission tomography: evaluation of small lesion filtering and intersection-over-union thresholds
PURPOSE: Evaluate how small lesion filtering and different intersection-over-union (IoU) thresholds influence deep learning segmentation of Ga-68-PSMA-11 PET images in prostate cancer. METHODS: A 3D U-Net was trained on 115 patient scans with manual contours as ground truth. Performance was assessed at voxel, lesion and patient levels. Lesions less than 8 voxels (195 mm 3 ) or less than 27 voxels (658 mm 3 ) were optionally discarded, and lesion-level metrics were computed across 10-50% IoU thresholds. RESULTS: Excluding lesions less than 27 voxels increased voxel-level Dice from 0.7975 to 0.8173 and improved precision. Lesion-level metrics were stable across IoU thresholds of 20-40% after 27-voxel filtering, while sensitivity declined at higher overlap thresholds. Patient-level sensitivity and positive predictive value reached 96.6 and 94.5%, respectively. CONCLUSION: Implementing small lesion filtering criteria improves the reliability of automated segmentation outputs, particularly for quantitative evaluation. For lesion-level metrics, defining true positives within the range of 20-40% IoU thresholds is optimal. Further validation through multicenter studies and larger datasets is essential to ensure the generalizability of these findings.
9 May 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
MGML: A plug-and-play meta-guided multi-modal learning framework for incomplete multimodal brain tumor segmentation
Leveraging multimodal information from Magnetic Resonance Imaging (MRI) plays a vital role in lesion segmentation, especially for brain tumors. However, in clinical practice, multimodal MRI data are often incomplete, making it challenging to fully utilize the available information. Therefore, maximizing the utilization of this incomplete multimodal information presents a crucial research challenge. We present a novel meta-guided multi-modal learning (MGML) framework that comprises two components: meta-parameterized adaptive modality fusion and consistency regularization module. The meta-parameterized adaptive modality fusion (Meta-AMF) enables the model to effectively integrate information from multiple modalities under varying input conditions. By generating adaptive soft-label supervision signals based on the available modalities, Meta-AMF explicitly promotes more coherent multimodal fusion. In addition, the consistency regularization module enhances segmentation performance and implicitly reinforces the robustness and generalization of the overall framework. Notably, our approach does not alter the original model architecture and can be conveniently integrated into the training pipeline for end-to-end model optimization. We conducted extensive experiments on the public BraTS2020 and BraTS2023 datasets. Compared to multiple state-of-the-art methods from previous years, our method achieved superior performance. On BraTS2020, for the average Dice scores across fifteen missing modality combinations, building upon the baseline, our method obtained scores of 87.55, 79.36, and 62.67 for the whole tumor (WT), the tumor core (TC), and the enhancing tumor (ET), respectively. We have made our source code publicly available at https://github.com/worldlikerr/MGML.
9 May 2026
Read appraisal →Annals of surgical oncology
CT-based Prediction of Visceral Pleural Invasion in Lung Adenocarcinoma ≤ 3 cm: Enhancing Deep Learning Specificity by Waiving Chest Wall Information
BACKGROUND: Accurate preoperative prediction of visceral pleural invasion (VPI) in lung adenocarcinoma is essential for guiding surgical decision-making. However, existing prediction models often sacrifice specificity when optimized for high sensitivity, increasing the risk of overtreatment. This study aimed to develop a computed tomography (CT)-based deep learning (DL) model that improves specificity by excluding chest wall information. METHODS: We retrospectively enrolled 835 patients with pathologically confirmed lung adenocarcinoma who underwent complete resection at two medical centers. A VPI-DL model was developed using a four-layer convolutional neural network incorporating a novel attention mechanism trained on chest-wall-masked CT inputs. The model was trained on 692 cases and externally validated on an independent cohort of 143 patients. Performance was benchmarked against the consolidation-to-tumor ratio (CTR), a published DL model, and assessments by three thoracic surgeons. Evaluation metrics included area under the curve (AUC), sensitivity, specificity, and accuracy under a high-sensitivity operating threshold. RESULTS: In external validation, VPI-DL achieved an AUC of 0.91, sensitivity of 91%, specificity of 82%, and accuracy of 83%, outperforming the CTR (specificity 63%), the previously published model (specificity 54%), and thoracic surgeons (specificity 57%) at comparable sensitivity levels. The attention-guided architecture effectively reduced spurious correlations and improved interpretability. CONCLUSION: The proposed VPI-DL model improves the specificity of VPI prediction while preserving high sensitivity. By intentionally excluding chest wall information, the model offers a robust and interpretable tool to aid preoperative planning and minimize the risk of both under- and overtreatment in early-stage lung adenocarcinoma.
6 May 2026
Read appraisal →Microscopy research and technique
Deep Learning Integration in Optical Microscopy: Advancements and Applications
Optical microscopy is a cornerstone imaging technique in biomedical research, enabling visualization of subcellular structures beyond the resolution limit of the human eye. However, conventional optical microscopy faces challenges such as optical aberrations, diffraction-limited resolution, low signal-to-noise ratio (SNR), and poor contrast. The exponential growth of bioimaging data further underscores the need for advanced computational tools. Deep learning (DL) is a subset of machine learning that has emerged as a transformative approach to address these limitations, offering enhanced precision, reduced manual intervention, and diminished reliance on domain-specific expertise for image reconstruction, enhancement, and analysis. This review explores the integration of DL into optical microscopy, focusing on key applications including image classification, segmentation, and computational reconstruction. We examine prominent DL architectures such as convolutional neural networks (CNNs), U-Nets, residual networks (ResNets), and generative adversarial networks (GANs)-and their role in advancing diverse microscopy modalities. These frameworks enhance image quality, improve quantitative analysis, and democratize access to high-performance microscopy. Additionally, we discuss persisting challenges, including the demand for large, annotated datasets, dynamic sample variability, model interpretability, and potential data biases. Collectively, DL is poised to revolutionize optical microscopy, shaping its future developments in biomedical imaging.
6 May 2026
Read appraisal →Medical image analysis
Robust fine-grained echocardiographic view classification with supervised contrastive learning
Accurate classification of echocardiographic views is fundamental for automated cardiac analysis. However, clinical practice relies on a large, heterogeneous set of fine-grained acquisitions that introduce substantial inter-observer variability. Existing studies have primarily focused on limited view sets, often collapsing specialised views into broad categories, which limits their clinical relevance. We address this limitation by introducing TTE47, the first publicly available benchmark comprising 47 clinically meaningful views annotated independently by three experts. This dataset enables the rigorous quantification of inter-observer agreement and establishes a foundation for reproducible, clinically relevant evaluation. To tackle the dual challenges of subtle inter-class distinctions and structured label variability, we propose a novel supervised contrastive learning framework incorporating a tailored loss function. Our method outperforms cross-entropy and standard supervised contrastive baselines, achieving leading performance among evaluated methods on TTE47 and surpassing prior work on TMED-2 without dataset-specific pretraining, using a model pretrained on TTE47. Beyond accuracy, we introduce clustering-based metrics, Detection Rate and Label Recovery Precision, that measure semantic coherence and the model's ability to resist annotation variability. Results show that the learned feature space aligns more strongly with underlying anatomical structure than with any single annotator's style, enabling resilience to label shifts and maintaining robustness comparable to human-level disagreement. By integrating multi-expert evaluation, robust representation learning, and interpretable feature-space analysis, this work establishes a scalable and clinically relevant framework for fine-grained echo view classification. Our findings highlight the potential of contrastive pretraining to standardise interpretation, mitigate subjectivity, and enhance the reliability of AI-assisted echocardiography in diverse clinical settings.
6 May 2026
Read appraisal →Artificial intelligence in medicine
TADynFed: Dynamic modality-adaptive federated learning with tissue-aware disentanglement for cross-disease analysis
Federated learning (FL) enables collaborative medical image analysis across decentralized institutions while preserving data privacy. However, real-world deployment faces critical challenges: modality heterogeneity, where clients possess incomplete or varying medical image modalities, and cross-disease generalization, requiring models to adapt to unseen pathologies with anatomical consistency. Existing methods like FedAvg, FedProx, IOP-FL, and PntTranForFL assume uniform modality availability and static client participation, leading to poor performance under realistic clinical constraints. We propose TADynFed, a novel framework for the Heterogeneous Federated Learning (HFL) paradigm, which addresses both data and modality heterogeneity through, a tissue-aware disentanglement strategy that decouples modality-tailored and modality-shared features, a dynamic prototype memory bank for missing modality compensation, an adaptive aggregation mechanism that accounts for client reliability and tenure. We evaluate TADynFed using multi-disease MRI datasets including BraTS21 along with cross-domain imaging data from CheXpert (chest X-ray) and Hep-2 (microscopy), simulating a 13-client FL environment. The proposed framework achieves an average mDice score of 66.03%, significantly outperforming baseline methods such as PointTransformerFL at 58.60% and FedAvg at 53.06%. It also attains the lowest boundary alignment error with an ASD of 1.85 mm and HD95 of 8.70 mm, surpassing existing approaches by notable margins. In terms of calibration stability, TADynFed records an ECE of 0.09, indicating superior confidence reliability. Furthermore, it demonstrates high communication efficiency with only 76 MB of data exchanged per round, compared to 95-105 MB in other frameworks. These results validate TADynFed's ability to maintain high segmentation accuracy, boundary precision, and calibration stability while minimizing bandwidth usage. Outperforming existing frameworks by significant margins, TADynFed demonstrates robust boundary alignment, superior calibration, and efficient communication. It also exhibits strong cross-disease transferability without retraining. By integrating structured representation decomposition, prototype-guided fusion, and client-adaptive learning, TADynFed establishes a new benchmark in realistic, heterogeneous, and mix-modal federated medical imaging systems. Code supporting this study TADynFed.
5 May 2026
Read appraisal →Journal of nuclear medicine : official publication, Society of Nuclear Medicine
Validation of a Deep-Learning Coregistration Framework for Long-Axial-Field-of-View PET/CT Using Low-Radiation-Exposure Protocols Across Various Tracers
The advent of long-axial-field-of-view (LAFOV) PET/CT systems has significantly improved whole-body imaging by providing higher sensitivity and extended torso coverage. However, PET/CT potential misalignment remains a challenge, potentially introducing artifacts and quantification errors. Moreover, PET protocols with reduced scanning duration and dose, as well as the use of ultra-low-dose CT (ULD-CT), are increasingly relevant in clinical practice and research. This study aimed to evaluate the robustness and generalizability of a deep-learning coregistration framework for PET/CT alignment using low-dose PET and ULD-CT across various tracers with a LAFOV system. Methods: In total, 63 scans with 4 different tracers (89Zr-trastuzumab, 15O-H2O, 18F-MC225, and 18F-FDG) were included to assess PET and CT alignment improvements. Further evaluation was performed by comparing CT scans coregistered to the original PET images (rCT) with those coregistered to low-count (50%, 25%, and 12.5% of original counts) PET images (LC-rCT). Dice similarity coefficient, Jaccard similarity coefficient, Hausdorff distance, and average surface distance were used as evaluation metrics. Furthermore, ULD-CT coregistered to PET (rULD-CT) was compared with low-dose CT coregistered to the same PET (rLD-CT) using the same metrics. PET accuracy was evaluated by calculating SUVs. Results: The robustness of the deep-learning coregistration framework was demonstrated in reduced PET counts and ULD-CT scenarios across tracers. All metrics indicated robust performance when comparing LC-rCT with rCT and rULD-CT with rLD-CT. Consistent SUVs across varying PET counts and ULD-CT conditions further validated the quantitative accuracy of the approach. Conclusion: This work highlights that neither the use of low-dose PET protocols nor ULD-CT compromise the performance of this coregistration framework across 4 tracers. For LAFOV PET, these findings support the feasibility of such a framework in imaging protocols with reduced radiation exposure.
5 May 2026
Read appraisal →Journal of neuroradiology = Journal de neuroradiologie
Automated estimation of perivascular space and brain morphometry from deep learning-reconstructed three-dimensional T1-weighted MRI: comparison with the conventional technique
PURPOSE: To evaluate the utility of deep learning-based reconstruction (DLR) three-dimensional T1-weighted imaging (T1-WI) in improving fine structural delineation through the automated estimation of brain morphometry and perivascular space (PVS). MATERIALS AND METHODS: In this retrospective cohort study, participants who underwent T1-WI between August 2021 and June 2022 were included. Participants were categorized into normal and disease groups based on the presence of visible cerebral lesions. Cortical thicknesses and cerebral volumes from DLR T1-WI were estimated using FreeSurfer and compared to those derived from conventional T1-WI. Additionally, participants were grouped by MRI-visible PVS severity according to the number of enlarged PVS in centrum semiovale and basal ganglia. PVS volumes from DLR and conventional T1-WI were compared across these groups. Statistical comparisons were performed using a paired independent t-test with Benjamin-Hochberg adjustment. RESULTS: 240 participants (mean age: 40.8 ± 17 years) were included. In the normal group (n = 144), cortical thicknesses and volumes were significantly greater in DLR T1-WI compared to conventional T1-WI (P < 0.001). This trend was also observed in the disease group (n = 96) and in external datasets (n = 63). Across all three PVS severity groups (mild, n = 111; moderate, n = 71; severe, n = 58), DLR T1-WI significantly reduced PVS volumes and variability compared to conventional T1-WI. DLR T1-WI demonstrated PVS volumes (mild: 738.4 mm³ vs. 1408.1 mm³; moderate: 919.0 mm³ vs. 1895.8 mm³; severe: 1328.5 mm³ vs. 2711.7 mm³; all P < 0.001) and variability (mild: 224.4 mm3 vs. 420.9 mm³; moderate: 229.7 mm3 vs. 462.9 mm³; severe: 382.4 mm3 vs. 653.6 mm³) that were consistently smaller. CONCLUSION: DLR T1-WI yielded cerebral cortical thickness and volume measurements that were more consistent with expected anatomical values than those from conventional T1-WI.PVS volumes measured from DLR T1-WI exhibited less variability and better concordance with radiologists' visual assessments.
5 May 2026
Read appraisal →Pediatric radiology
Deep learning based gestational age estimation from multi-view fetal brain magnetic resonance imaging
BACKGROUND: Gestational age (GA) is essential for assessing fetal development, but conventional methods such as last menstrual period and ultrasound are often inaccurate, particularly in late pregnancy. Recent advances in deep learning (DL) and MRI offer more reliable and consistent GA estimation by capturing detailed fetal brain development. OBJECTIVE: This study aimed to develop deep learning models for GA prediction using multi-view fetal brain MRI and to compare their performance with conventional biometric regression techniques. MATERIALS AND METHODS: A total of 1,321 fetal MRI scans were used to train and evaluate various DL models, while an additional 80 publicly available MRI scans served as an external test set. Two training strategies were explored: transfer learning versus training from scratch, and single-view versus multi-modality input. RESULTS: The pre-trained ResNet-101 model achieved a mean absolute error (MAE) of 4.47 days and a coefficient of determination (R2) of 0.96 on the internal test set. On the external test set, the model yielded an MAE of 6.57 days, outperforming the biometric regression method, which achieved an MAE of 9.42 days. Explainability analysis revealed that the model predominantly focused on the lateral ventricles, cerebellum, and surrounding brain regions for GA prediction. CONCLUSIONS: The integration of multi-view MRI and transfer learning significantly enhanced the predictive accuracy of DL models for GA estimation. The proposed approach outperformed conventional biometric regression and highlighted clinically relevant anatomical regions, demonstrating its potential for use in prenatal diagnostic applications.
5 May 2026
Read appraisal →Medical image analysis
PH2ST: Prompt-guided hypergraph learning for spatial transcriptomics prediction in whole slide images
Spatial Transcriptomics (ST) reveals the spatial distribution of gene expression in tissues, offering critical insights into biological processes and disease mechanisms. However, the high cost, limited coverage, and technical complexity of current ST technologies restrict their widespread use in clinical and research settings, making obtaining high-resolution transcriptomic profiles across large tissue areas challenging. Predicting ST from H&E-stained histology images has emerged as a promising alternative to address these limitations but remains challenging due to the heterogeneous relationship between histomorphology and gene expression, which is affected by substantial variability across patients and tissue sections. In response, we propose PH2ST, a prompt-guided hypergraph learning framework, which leverages limited ST signals to guide multi-scale histological representation learning for accurate and robust spatial gene expression prediction. Extensive evaluations on two public ST datasets and multiple prompt sampling strategies simulating real-world scenarios demonstrate that PH2ST not only outperforms existing state-of-the-art methods, but also shows strong potential for practical applications such as imputing missing spots, ST super-resolution, and local-to-global prediction, highlighting its value for scalable and cost-effective spatial gene expression mapping in biomedical contexts.
5 May 2026
Read appraisal →Abdominal radiology (New York)
A CT-based 2.5D deep learning model for preoperative T-staging in gastric cancer: a retrospective multicenter study
PURPOSE: To develop and validate a 2.5D multi-angle deep learning (MADL) model for preoperative T-staging in patients with gastric cancer (GC) and to explore the predictive potential for survival outcome. MATERIALS AND METHODS: In this retrospective study, GC datasets from four centers were used to develop a quaternary for preoperative T-staging model by integrating CT images with nine 2D slicers and three angles (0°, + 45°, -45°) from transverse, sagittal, and coronal views. Diagnostic performance (accuracy, sensitivity, specificity, F1-score) were compared between the 2.5D MADL and radiologists. The prognostic potential of 2.5D MADL-derived features was evaluated using Kaplan-Meier analysis and multivariate Cox regression. RESULTS: A total of 433 patients were divided into the training (n = 346, mean age, 64.8 years ± 10.5 [SD]; 95 female) and the internal validation (n = 87, mean age, 63.0 years ± 11.9; 23 female) sets from Centers 1-3, and 41 patients (mean age, 64.6 years ± 8.9; 10 female) formed the external testing set from Center 4. The 2.5D MADL model exhibited excellent diagnostic performance for T-staging in external testing set (AUCs of T1-T4: 0.962, 0.722, 0.913, 0.962). For T1 staging, it outperformed radiologists in accuracy (95% vs. 83%), sensitivity (75% vs. 12%), and F1 score (86% vs. 22%) in external testing set. Survival analysis showed significant overall survival differences between high- and low-risk groups stratified (P < 0.001), with high-risk patients having shorter survival (HR = 1.632, 95% CI:1.239-2.150). CONCLUSION: The 2.5D MADL model achieves accurate and reliable preoperative T staging of gastric cancer, with performance comparable to radiologists and superior accuracy for early-stage disease.
5 May 2026
Read appraisal →Journal of clinical ultrasound : JCU
Illuminating Research Dynamics: Medical Ultrasound and Deep Learning
BACKGROUND: Deep learning (DL) has enabled advances in ultrasound imaging, but challenges like limited datasets and device variability hinder progress. This study provides the first bibliometric overview of DL research in medical ultrasound. METHOD: We retrieved related publications (2004-Apr 2025) on medical ultrasound and DL from the Web of Science Core Collection. Bibliometric tools (Bibliometrix, VOSviewer, CiteSpace) were used to quantify publication growth, country/institution contributions, and thematic evolution. RESULTS: A total of 3386 publications were included, with the annual output increasing steadily at an average growth rate of 31.02%. While China led all countries in publication volume (1355), the United Kingdom achieved the highest citation rate per article (55.5 citations). Fudan University was the most prolific institution (110), and key authors such as Luca Saba and Jasjit S. Suri formed the principal collaboration hubs. In terms of venues, Diagnostics published the most papers, whereas IEEE Transactions on Medical Imaging exerted the greatest citation impact (5646 citations; IF 8.9). Keyword analysis revealed core themes in classification, segmentation, and CNN-based diagnosis. From 2020 onward, ultrasound-specific topics (e.g., diagnosis, images) dominate; advanced architectures (e.g., U-Net) and emerging clinical concepts such as noninvasive assessment. CONCLUSION: This mapping highlights a rapidly growing, internationally collaborative field. Emerging trends include advanced artificial intelligence (AI) architectures (U-Net, attention-based networks, vision Transformers) and a shift toward surgical and therapeutic applications. These insights underscore the need for multicenter validation, standardized data protocols, and clinically interpretable AI models to accelerate safe translation into practice.
5 May 2026
Read appraisal →Magnetic resonance in medicine
Dynamic-Guided Diffusion Probability Model for Cranial Nerves Segmentation
PURPOSE: Segmentation of cranial nerves (CNs) bundles using magnetic resonance imaging (MRI) provides a valuable quantitative approach for analyzing the morphology and orientation of individual CNs. Currently, the CN regions can be segmented directly using deep learning-based methods. However, existing methods overlook the unique characteristics of CNs, particularly their environmental features and representation in multimodal images that may lead to suboptimal segmentation outcomes. METHODS: We proposed a dynamic-guided diffusion probability model for CNs segmentation, which enhances segmentation performance by integrating the intrinsic characteristics of CNs. A dynamic-guided mechanism approach called the SE-A-NL module was proposed. The module is capable of addressing both the varying characterization abilities of multimodal images and the long-range connections of CNs within images. RESULTS: Quantitative and qualitative experiments demonstrate that the proposed method surpasses current state-of-the-art approaches, delivering accurate and effective segmentation of five pairs of cranial nerves. Notably, the method outperforms existing techniques in 16 out of the 20 evaluated metrics. CONCLUSION: The overall network model effectively integrates multimodal information and anatomical priors by combining multi-channel attention and non-local attention mechanisms, thereby improving CNs segmentation performance. Thorough comparative and ablation studies highlight the superior performance of the proposed method.
5 May 2026
Read appraisal →Graefe's archive for clinical and experimental ophthalmology = Albrecht von Graefes Archiv fur klinische und experimentelle Ophthalmologie
Automatic measurement and evaluation of anterior segment anatomical structures via UBM images using a deep learning-based approach
PURPOSE: To develop a deep-learning model capable of measuring essential anterior segment (AS) parameters derived from preoperative ultrasound biomicroscopy (UBM) images of candidates for implantable collamer lens (ICL) surgery. SETTING: Tianjin Medical University Eye Hospital, Tianjin, China. DESIGN: Cross-sectional retrospective study. METHODS: A dataset comprising 1,480 preoperative panoramic UBM images taken from 638 eyes of 320 subjects was collected and was divided into training and testing subsets at a proportion of 7:3. Using the YOLOv8-pose algorithm, the model identified ten anatomical key point coordinates and computed six relevant AS parameters. Both manual and anterior segment optical coherence tomography-based non-contact measurements served as reference standards for evaluating the model's accuracy. The relationship between postoperative vault and preoperative parameters measured by the model was analyzed using multiple linear regression. RESULTS: On the test dataset, the model achieved an intraclass correlation coefficient (ICC) exceeding 0.978, with a mean Euclidean distance of 67.65 ± 54.25 μm across all point locations. The ICC values for anterior chamber depth (ACD), pupil diameter, and sulcus-to-sulcus distance were above 0.980 (95% CI: 0.975 to 0.985), with average relative error below 1.7%. Additionally, postoperative vault at one month was significantly correlated with model-measured parameters, including crystalline lens rise, iris concavity, and ACD (P < 0.001). CONCLUSION: This study introduces a robust program capable of quantitatively measuring AS parameters with accuracy comparable to that of experienced ophthalmologists. The findings provide valuable guidance for ICL sizing and vault prediction.
5 May 2026
Read appraisal →Experimental eye research
Automated segmentation of pterygium lesions using multiscale deep learning networks
Pterygium is an eye condition that needs to be identified at an early stage so that its progression can be mitigated in order to avoid the possible threat of visual impairment. One of the important low-level modules in determining the severity level of the pterygium automatically is to measure the extent of fibrovascular tissue encroachment onto the corneal and pupil regions. Therefore, it is important to develop a semantic segmentation method to map and quantify the lesion tissues accurately, especially through the deep learning technique. As the pterygium condition progresses from the trace stage to the severe stage, the lesions will generally become bigger but maintain more or less the same shape, which looks like a wedge. Taking inspiration from this unique pattern of pterygium lesions, this work aims to explore multiscale deep learning networks to capture the variable scales of the lesion features effectively. All the proposed multiscale modules are embedded at the bottleneck layer of the base UNet architecture, whereby the feature map size is the smallest to reduce the potential increment in the computational burden. Apart from that, the total number of convolutional channels is fixed to 1024, whereby the multiscale module replaces the first convolutional set at the UNet's bottleneck layer. Two main multiscale modules have been explored, which are spatial pyramid pooling (SPP) and atrous spatial pyramid pooling (ASPP) modules. Furthermore, each of these modules has been expanded by analyzing the equal-flow (EF) and waterfall-flow (WF) patterns in constructing the parallel paths that represent the multiscale features. According to the simulation results, the best segmentation network is UNet, which has been embedded with three parallel paths of the EF-ASPP module that produces the lowest Hausdorff distance of 16.75 pixels. As a result, the severity level of pterygium can be better predicted through accurate extraction of the lesion map. This work can be further improved by analyzing different network architectures of the multiscale module, especially networks that focuses on the larger feature maps size, but surely with a trade-off in computational burden.
5 May 2026
Read appraisal →Abdominal radiology (New York)
Deep learning-based combined noise reduction and contrast enhancement for post-neoadjuvant pancreatic cancer CT: does improved image quality translate to better resectability assessment?
PURPOSE: To evaluate whether deep learning-based combined noise reduction and contrast enhancement reconstruction (DLR) improves image quality and resectability prediction accuracy compared to conventional iterative reconstruction (IR) in post-neoadjuvant pancreatic cancer CT assessment. METHODS: This retrospective study included 114 patients with pancreatic cancer following neoadjuvant therapy. Contrast-enhanced CT images were reconstructed using conventional IR and vendor-neutral ClariACE. Three abdominal radiologists independently assessed image quality (based on 8 parameters: tumor conspicuity, tumor margin, image noise, sharpness of the main pancreatic duct, arterial depiction, venous depiction, plasticity and overall image quality) and determined tumor resectability with confidence levels. Quantitative analysis included aortic and portal venous attenuation measurements and pancreas-to-tumor contrast-to-noise ratio (CNR). Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC) with DeLong's test. Sensitivity, specificity, accuracy, and reader confidence were compared using McNemar's test. RESULTS: DLR demonstrated significantly superior vessel enhancement (p < 0.001) and improved CNR (p < 0.001) versus conventional IR. Two readers consistently rated DLR images higher across all qualitative categories (p < 0.001) except plasticity, while the third reader favored DLR in five of eight parameters (p < 0.001 to 0.020). However, all readers noted increased artificial appearance in DLR images (p < 0.001). Despite image quality improvements, no significant differences were observed in resectability assessment accuracy (62.3%-65.8%), AUC values (0.485-0.520), or high-confidence diagnosis rates between reconstruction methods. CONCLUSION: Although deep learning-based combined noise reduction and contrast enhancement reconstruction significantly improved quantitative and subjective image quality metrics, it did not enhance diagnostic accuracy for predicting R0 resectability in post-neoadjuvant pancreatic cancer patients.
5 May 2026
Read appraisal →European radiology
A deep learning model to reduce agent dose for contrast-enhanced MRI of the cerebellopontine angle cistern
OBJECTIVES: To evaluate a deep learning (DL) model for reducing the agent dose of contrast-enhanced T1-weighted MRI (T1ce) of the cerebellopontine angle (CPA) cistern. MATERIALS AND METHODS: In this multicenter retrospective study, T1 and T1ce of vestibular schwannoma (VS) patients were used to simulate low-dose T1ce with varying reductions of contrast agent dose. DL models were trained to restore standard-dose T1ce from the low-dose simulation. The image quality and segmentation performance of the DL-restored T1ce were evaluated. A head and neck radiologist was asked to rate DL-restored images in multiple aspects, including image quality and diagnostic characterization. RESULTS: 203 MRI studies from 72 VS patients (mean age, 58.51 ± 14.73, 39 men) were evaluated. As the input dose increased, the structural similarity index measure of the restored T1ce increased from 0.639 ± 0.113 to 0.993 ± 0.009, and the peak signal-to-noise ratio increased from 21.6 ± 3.73 dB to 41.4 ± 4.84 dB. At a 10% input dose, using DL-restored T1ce for segmentation improved the Dice from 0.673 to 0.734, the 95% Hausdorff distance from 2.38 mm to 2.07 mm, and the average surface distance from 1.00 mm to 0.59 mm. Both DL-restored T1ce from 10% and 30% input doses showed excellent image quality (3, interquartile range (IQR) [Q3-Q1] = 3-3 and 3, IQR [Q3-Q1] = 4-3), with the latter being considered more informative (4, IQR [Q3-Q1] = 4-3). CONCLUSION: The DL model improved the image quality of low-dose MRI of the CPA cistern, which makes lesion detection and diagnostic characterization possible with 10-30% of the standard dose. KEY POINTS: Question Deep learning models that aid in the reduction of contrast agent dose are not extensively evaluated for MRI of the cerebellopontine angle cistern. Findings Deep learning models restored the low-dose MRI of the cerebellopontine angle cistern, yielding images sufficient for vestibular schwannoma diagnosis and management. Clinical relevance Deep learning models make it possible to reduce the use of gadolinium-based contrast agents for contrast-enhanced MRI of the cerebellopontine angle cistern.
4 May 2026
Read appraisal →Journal of medical imaging and radiation sciences
Explainability-informed benchmarking of two deep learning models for organ-at-risk segmentation in MR-guided adaptive radiotherapy
INTRODUCTION/BACKGROUND: Segmentation of gastrointestinal (GI) organs-at-risk (OARs) is a critical yet time-consuming step in MR-guided adaptive radiotherapy (MRgRT), with manual delineation prone to inter- and intra-observer variability. While deep learning approaches have shown promise, their clinical adoption requires not only accuracy but also interpretability and reliability. This study benchmarks two widely used convolutional architectures, U-Net and Residual U-Net (ResUNet), for abdominal OAR segmentation, with an emphasis on explainability-oriented quantitative analysis. METHODS: An anonymized abdominal MRI dataset was used to train and evaluate U-Net and ResUNet using a 5-fold stratified group cross-validation strategy. Segmentation performance was assessed using the Dice Similarity Coefficient (DSC), Intersection-over-Union (IoU), and the 95th percentile Hausdorff Distance (HD95). Explainability was investigated using Gradient-weighted Class Activation Mapping (Grad-CAM) computed from the final convolutional layer of each network. To enable objective analysis beyond qualitative visualization, Grad-CAM activation maps were quantified using numerical localization metrics relative to ground-truth organ masks, including in-organ energy ratio, boundary energy ratio, pointing accuracy, activation Dice coefficient, centroid distance and activation entropy. Grad-CAM metrics were aggregated across gastrointestinal organs and averaged over the five validation folds. RESULTS: Both architectures demonstrated comparable segmentation performance across organs, with no statistically significant differences across evaluated metrics. Grad-CAM analysis showed similar region-level attention patterns, with in-organ activation ratios of 71.4 ± 8.6% for U-Net and 66.2 ± 9.1% for ResUNet, boundary energy ratios of 24.1 ± 4.9% and 21.8 ± 5.2%, respectively, and pointing accuracies exceeding 70% for both models. Uncertainty analysis based on inter-fold variability and boundary error dispersion indicated comparable stability and bounded worst-case behavior. DISCUSSION/CONCLUSION: By integrating performance, uncertainty and explainability quantitative indicators, this study provides an informed benchmarking of two deep learning models for abdominal OAR segmentation. The results suggest that both U-Net and ResUNet exhibit stable and interpretable behavior under the evaluated configurations, supporting their potential use in MR-guided adaptive radiotherapy workflows where reliability and clinical trust are essential.
4 May 2026
Read appraisal →Journal of medical imaging and radiation sciences
Automatic field-of-view planning for magnetic resonance shoulder imaging using Deep Learning
INTRODUCTION: Accurate prescription of oblique coronal and oblique sagittal field of views (FOV) is essential for diagnostic shoulder MRI. Manual planning is radiographer-dependent, time-consuming, and subject to inter- and intra-operator variability, leading to inconsistent image quality and incomplete coverage. Although deep learning (DL) has advanced automated scan planning in non-oblique planes, oblique shoulder prescriptions remain underexplored; an automated DL approach could standardize FOV prescription, reduce operator dependence, and improve reproducibility and workflow without compromising diagnostic quality. METHODS: In this retrospective multicenter study, 575 shoulder MRI examinations (2019-2025) from four sites were included. Sites A (n=151) and B (n=220) were used for training; testing was performed on sites C (n=61), and D (n=143). A two-stage pipeline was implemented using five oriented bounding box (OBB) variants of YOLOv11 (n, s, m, l, x): Stage 1 performed slice selection; Stage 2 performed FOV prescription. Performance was evaluated against radiographers' prescriptions using mean absolute slice difference (MASD, slices), intersection over union (IoU), and mean absolute angle difference (MAAD, degrees). Clinical utility was assessed by three raters. RESULTS: The YOLOv11-OBB-l model achieved the lowest MASD for Stage 1 (1.016±0.153 slices). For Stage 2, YOLOv11-OBB-x performed best (coronal IoU, 0.847±0.003; sagittal IoU, 0.852±0.007; MAAD, 3.259±0.190°). During testing across each site, MASD ranged from 0.700±0.837 to 1.192±2.550 slices; MAAD from 2.811±2.348 to 4.396±7.158°; coronal IoU from 0.800±0.092 to 0.872±0.065; and sagittal IoU from 0.824±0.111 to 0.887±0.047. Mean clinical utility was 97.2%. Performance was noninferior to interrater variability across all sites and metrics. CONCLUSION: DL-based automated FOV prescription for shoulder MRI achieves performance comparable to radiographers, generalizes across institutions, and demonstrates high clinical utility.
4 May 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
MMC-CS: Multi-branch multi-stage contrastive learning for self-supervised compressed sensing
In recent years, optimization-inspired networks that integrate optimization theory into deep neural networks (DNNs) have achieved remarkable success in image compressed sensing (ICS). However, current DNN-based ICS methods face two key challenges: (1) the difficulty of acquiring ground-truth (GT) data from labeled measurements and (2) the underutilization of measurement data. To address these issues, a novel self-supervised deep learning framework is proposed to solve the inverse problem in ICS. Building upon traditional self-supervised learning, our proposed method leverages measurement values through a multi-branch, multi-stage progressive cross-contrast structure, enabling effective learning of the common prior of the original image in the absence of GT data. Based on this framework, a novel Multi-branch Multi-stage Cross-contrast CS (MMC-CS) end-to-end DNN is designed, which unfolds the Proximal Gradient Descent (PGD) algorithm. Our approach further implements multi-scale co-optimization in the image reconstruction process, integrating both the image paths (IPs) and convolutional feature paths (CFPs). Additionally, the reconstruction performance across multiple spatial levels is enhanced by incorporating wavelet convolution (WTConv). Extensive experiments demonstrate that our proposed method outperforms existing self-supervised approaches, achieving an average PSNR improvement of 0.3-1.6 dB. Moreover, it shows strong potential to compete with state-of-the-art supervised methods.
4 May 2026
Read appraisal →NMR in biomedicine
Deep Learning Reconstruction Enables Diagnostic-Quality 0.4T Knee and Spine MRI in One-Third of the Time
There has been a growing interest in low-field MRI due to its lower costs, enabling an increase in accessibility of MRI worldwide. Long scan times are currently needed to obtain high-quality images. The aim of our study was to accelerate low-field 0.4T musculoskeletal MRI using deep learning and evaluate maximum acceleration via assessment by two expert radiologists. A cascades of independently recurrent inference machines (CIRIM) model was trained using knee and spine 2D multislice scans, retrospectively undersampled in k-space. The heterogeneous dataset included varied contrasts, orientations, and matrix sizes. To determine the maximum possible acceleration factor, the undersampling pattern and loss function of the CIRIM were optimized using the knee dataset. Models were trained for acceleration factors 1.5-4. The different acceleration factors were evaluated quantitatively and clinically, with radiologist scores compared using Kendall's Tau and Cohen's Kappa. To assess the model's generalizability, two prospectively undersampled scans were evaluated, and the knee-trained models were tested on spine data. CIRIM accurately reconstructed all data, with slight metric degradation at higher accelerations. Combining L1, structural similarity index measure (SSIM), and perceptual loss improved sharpness, while undersampling pattern differences were minimal. Clinical scores from both radiologists declined with acceleration, reflected in Kendall's Tau values of 0.64, 0.64, 0.54, and 0.55 for image quality, sharpness, anatomical conspicuity, and artifacts. Inter-observer agreement was low (Cohen's kappa: 0.227-0.299). The model generalized well, reconstructing spine data with a knee-trained model at comparable quality and successfully handling prospectively undersampled scans. CIRIM consistently delivered high-quality reconstructions, demonstrating excellent robustness to variations in contrast, orientation, matrix size, and anatomy. Its ability to generalize across anatomies and handle prospectively undersampled scans highlights its practical utility. The differences in reading for higher accelerations between radiologists suggest that the optimal acceleration factor is context-dependent and should be tailored to the specific diagnostic task.
4 May 2026
Read appraisal →Taiwanese journal of obstetrics & gynecology
Segmentation of uterus and myoma from MRI image based on deep learning
OBJECTIVE: Myoma is a common gynecologic condition with abnormal muscular and fibrotic tissue growth in the uterus. Compared with sonography, magnetic resonance imaging (MRI) provides more detailed information with a better diagnostic accuracy. However, the interpretation of the imagery is shrewdness- and experience-dependent. The therapeutic strategy of myoma can be directly affected by the relative location and size in the uterus. Hence, this study aimed to generate an objective standardized modality in interpreting the imagery. MATERIALS AND METHODS: Deep learning (DL) was used to train two models using MRI imagery labeled by radiologists: 1) Model Uterus + Endometrium: segmentation of uterus and endometrium; 2) Model Myoma: segmentation of myoma. The myomas were classified based on their relative position in the uterus according to FIGO criteria. Each set of MRI image sequence was obtained from uterine sagittal sections with 5.5-mm interval of 30 slices per case. The dataset was divided into training, validation, and test sets. The number of images was first increased by data augmentation before implementing a DL semantic segmentation DeepLabV3+model. Cross entropy and Dice function were used for the loss and evaluation functions, respectively. RESULTS: The Dice values of uterus, endometrium and myoma segmentation reached 81.55%, 73.49% and 67.34%, respectively. Thus, the 3-D prediction of the uterine, endometrial and myomatous regions had a positive effect on determining the classification of myomas. CONCLUSION: DL can be used as a tool to improve the diagnosis and classification of uterine myoma. This model can also be applied to other gynecologic diseases.
4 May 2026
Read appraisal →Journal of occupational and environmental medicine
Combining MEA-Net and LAP-Net for Pneumoconiosis Staging Framework
OBJECTIVE: Pneumoconiosis is a common and highly hazardous occupational disease. The staging of pneumoconiosis is mainly carried out by experienced doctors on the basis of the shadows and textures on lung x-ray films. Despite well-defined criteria, the process remains influenced by individual clinical judgment. METHODS: To improve the subjective process of pneumoconiosis diagnosis, this study proposes a new deep learning framework for pneumoconiosis staging framework using MEA-Net and LAP-Net. RESULTS: The experimental results show that the accuracy, precision, recall, specificity, F1-score, and area under the curve in the four-stage classification reached 95.24%, 95.15%, 95.15%, 90.58%, 94.85%, and 98.87%, respectively. CONCLUSION: The proposed method can help doctors to identify the different stages of pneumoconiosis more accurately in the diagnosis of the disease.
4 May 2026
Read appraisal →Knee surgery, sports traumatology, arthroscopy : official journal of the ESSKA
Synthetic computed tomography from magnetic resonance imaging: An editorial on deep learning approaches for hip and knee image translation.
This editorial provides a concise overview of synthetic computed tomography (CT) generation from magnetic resonance imaging (MRI) for musculoskeletal care, with a focus on hip and knee applications. We outline why interest in this area is growing, including the potential to reduce radiation exposure, consolidate imaging into a single visit, and align measurements across modalities for preoperative planning and longitudinal follow-up. At a high level, we summarize how deep learning methods transform MRI into CT-like volumes and what that could mean for bone morphology, implant planning, and combined assessment of bone and soft tissue. We briefly introduce representative model classes, including conditional generative adversarial networks that pair realism objectives with voxelwise constraints and diffusion models that replace adversarial training with iterative denoising for stable optimization and strong anatomical fidelity. Our goal is to orient clinicians, engineers, and researchers to the key promises and open questions, and to motivate collaborative studies that test whether synthetic CT can deliver reliable, reproducible, and clinically useful information at the point of care.
4 May 2026
Read appraisal →Magnetic resonance in medical sciences : MRMS : an official journal of Japan Society of Magnetic Resonance in Medicine
Visualization of the Trochlear Nerve Using Deep Learning-enhanced 3D T2-weighted MR Imaging at 3T
PURPOSE: Cranial nerve imaging with 3T MRI commonly uses 3D fast imaging employing steady-state acquisition (3D-FIESTA); however, this sequence has limitations in achieving higher spatial resolution and provides poor tissue contrast between cranial nerves and adjacent vascular structures. We evaluated 3D T2-weighted imaging (T2-CUBE) with deep learning-based reconstruction (DLR) for cranial nerve visualization, focusing on the trochlear nerve, the smallest cranial nerve with the longest intracranial course, and compared it with T2-CUBE without DLR and 3D-FIESTA. METHODS: Ten healthy male volunteers (age, 23-40 years; mean age, 32 years) underwent T2-CUBE with and without DLR, and 3D-FIESTA at 3T. Two neuroradiologists independently evaluated trochlear nerve visualization in 4 anatomical segments (origin from the midbrain, cisternal, tentorial, and anterior portion of its cavernous segments) using a 3-point scale, and SNR of the pons (SNRPONS) and cerebrospinal fluid (SNRCSF) were calculated. RESULTS: T2-CUBE with DLR achieved a 100% visualization across all trochlear nerve segments and demonstrated significantly better visualization than both T2-CUBE without DLR and 3D-FIESTA (P < 0.025). T2-CUBE without DLR showed 67.5%-100% visualization across the 4 segments. 3D-FIESTA showed 32.5%-80% visualization of the origin from the midbrain, cisternal, and tentorial segments, with no visualization of the cavernous segments. DLR increased SNRPONS and SNRCSF by factors of 1.8-2.5 (SNRPONS: 14.1 vs 5.7; SNRCSF: 31.8 vs 17.5, respectively; P < 0.001). T2-CUBE with DLR demonstrated significantly higher SNRPONS than 3D-FIESTA (14.1 vs 6.4, P < 0.001), while SNRCSF was comparable (31.8 vs 36.3, P = 0.20). CONCLUSION: T2-CUBE with DLR at 3T provided a significantly better trochlear nerve visualization than T2-CUBE without DLR and 3D-FIESTA. This technique may extend beyond the trochlear nerve to other cranial nerves and to the evaluation of neurovascular compression in the cistern, with the potential to become the new standard for cisternal imaging.
3 May 2026
Read appraisal →Physics in medicine and biology
FLIP-HEDOS: a patient-specific blood dose quantification model during radiotherapy treatments
Objective.To introduce FLIP-HEDOS, a blood dose modeling framework that combines patient-specific and generic circulatory data to enable individualized blood dose estimation during radiotherapy (RT) treatments and to overcome limitations of non-personalized approaches.Approach.Patient-specific images of vasculature were obtained from patients with thorax-abdomen or head-neck tumors treated with proton or photon RT. Large vessels from the tumor location were segmented, and their blood flow velocities were extracted from phase-contrast MRI. Movement of blood particles (BPs) was simulated using a Lagrangian approach along patient-specific flowlines in these vessels, forming the FLIP method. This circulation was incorporated into a standardized model (HEDOS) to form a closed-loop whole-body model that computes the spatiotemporal BP distribution and associated blood dose accumulation, accounting for dose delivery dynamics, dose rate and beam-on/beam-off times. Cardiac output (CO) was individualized to match each patient blood flow characteristics. FLIP-HEDOS retains the modular HEDOS architecture, extending its large vessel compartments with the FLIP approach patient-specific vessels represented as new compartments. Remaining compartments were adapted to preserve the standardized connectivity of the circulatory network.Main Results.The study quantified mean blood dose values and tracked the number of times BPs entered patient-specific compartments during dose delivery. Simulations were performed across a range of inter-beam intervals, and the three models FLIP (patient-specific vessels), HEDOS (generic whole-body), and FLIP-HEDOS (integrated model) were compared with respect to their resulting blood dose values. The comparison shows that when patient-specific information is incorporated in blood dose assessment, BPs experience higher radiation doses.Significance.FLIP-HEDOS proves that standardized vascular data can be effectively combined with patient-specific parameters to achieve physiologically realistic modeling of circulating blood. By enabling individualized CO and blood flow dynamics, the framework adapts to the anatomical and hemodynamic characteristics of each patient.
3 May 2026
Read appraisal →PloS one
Deep feature optimization using fusion of multiple self-supervised learning approaches and filter-based feature selection for lung cancer histopathology classification
One of the most deadly illnesses in the world is lung cancer, and increasing survival rates require early detection. Lung cancer diagnostics from the imaging modalities is always subjective, and this paves the way for deep learning assisted computer aided techniques. Still, the accuracy of such a technique is the major concern. This research work attempts to enhance lung cancer diagnostics from histopathological images using a previously unaddressed combination of multiple self-supervised learning techniques, filter-based feature selection, and Vision Graph Convolutional Networks. The main contribution lies in optimized feature fusion, and it brings complementary strengths of three different self-supervised learning approaches- contrastive alignment, redundancy reduction, and semantic grouping. The first step involves extracting the key features from histopathological images using a custom Convolutional Neural Network. Three complementary self-supervised learning methods - Deep Cluster, Bootstrap Your Own Latent, and Simple Framework for Contrastive Learning of Visual Representations - are then used to refine each of these features separately. Following optimization, the improved features are combined, and the most significant features from the combined set are found and preserved using a filter-based feature selection technique called Minimum Redundancy Maximum Relevance. Vision Graph Convolutional Neural Network is used as a classifier. Initial experiments are carried out on the 691 histopathological images extracted from the publicly available LungHist700 dataset. This dataset contains histopathological images under three categories: 151 normal subjects, 280 lung adenocarcinoma subjects, and 260 lung squamous cell carcinoma subjects. The proposed approach provided a balanced accuracy of 97% while the plain Vision Graph Convolution Neural Network offers only an 86% balanced accuracy score. Further validation is performed using two more publicly available datasets, namely LC25000 and TCGA UT datasets. The experimental results demonstrate the enhanced lung cancer prediction performance of the proposed approach in all three datasets.
2 May 2026
Read appraisal →Physics in medicine and biology
Few shots transfer learning for universal SPECT denoising across diverse acquisition protocols
Objective.Accelerated single photon emission computed tomography (SPECT) imaging, achieved by reducing either the number of projection angles or the acquisition time per angle, enhances clinical workflow efficiency but introduces elevated noise. This study aims to develop and validate a universal DL-based reconstruction framework that effectively generalizes across diverse, clinically-realistic SPECT acceleration protocols by overcoming the data scarcity challenge.Approach.SPECT bone scans from 103 patients were acquired under a standard scan (60 views, 12 s/view, 60v12s) followed by a fast scan using one of five acceleration protocols (60v6s, 60v3s, 30v12s, 30v6s, 16v12s). AU-Net-based reconstruction framework was implemented using three strategies: (1) single model, trained on individual acceleration protocol), (2) base model, trained on aggregated datasets from five acceleration protocols, and (3) transfer model, fine-tuned from the base model for protocol-specific optimization. Pixel-level accuracy and structural similarity were assessed using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) metrics, and maximum pixel value of lesions. Clinical evaluation of image quality, radionuclide detail, artifacts, and diagnostic confidence was conducted using a 5-point system.Main results.Quantitative evaluation showed the transfer model achieved better PSNR and SSIM across all protocols (highest 48.02 PSNR and 0.9918 SSIM in 30v6 s protocol). Qualitative analysis confirmed enhanced structural fidelity. Clinical evaluations rated the transfer model highest across metrics, with scores of 4.667 ± 0.508 (image quality, 30v12 s), 4.800 ± 0.250 (radionuclide detail, 60v6 s), 1.150 ± 0.173 (artifact reduction, 60v6 s), and 4.800 ± 0.250 (diagnostic confidence, 60v6 s), surpassing full-scan results in most cases.Significance. The proposed transfer learning (TL) framework effectively addressed data scarcity and improved reconstruction performance across diverse SPECT acceleration scenarios. The adoption of a TL strategy mitigates data scarcity by utilizing shared features and fine-tuning for specific protocols. The framework demonstrates potential for integration into fast SPECT workflows, facilitating reliable use across diverse imaging scenarios.
2 May 2026
Read appraisal →Biomedical physics & engineering express
BiFUSE-UNet: a novel dual-encoder enhanced TransUNet for polyp segmentation
Polyps are abnormal protrusions that develop on the mucosal surface of the human body, arising from hyperproliferation or aberrant differentiation of mucosal tissue. While some polyps carry a risk of malignant transformation, timely intervention can halt this progression. Accurately segmenting polyp contours in medical images poses significant challenges, primarily due to the high heterogeneity in polyp appearance and morphology, blurry boundaries with background interference, and class imbalance coupled with data scarcity. To address these challenges, this paper presents BiFuse-UNet, a novel dual-encoder TransUNet architecture built upon a bidirectional fusion mechanism. Building upon the original TransUNet architecture, BiFuse-UNet incorporates several key modules. Firstly, a multi-axis attention module (MaxViT block) is integrated to form a dual-encoder structure, which extracts multi-scale local textural features across different dimensions and captures long-range dependencies through self-attention. Secondly, a Merge fusion module is designed, incorporating a Coordinate Attention module and the SaE block from SENetV2, to enhance the model's perception of local and detailed features. Finally, a Biformer module is introduced to strike an efficient balance between global semantics and local details, thereby improving the segmentation accuracy for subtle polyps and early-stage polyps with atypical morphology. Extensive evaluations on the Kvasir-SEG and CVC-ClinicDB dataset demonstrate that BiFuse-UNet achieves state-of-the-art segmentation performance, outperforming seven existing methods on key metrics such as the Dice coefficient, Recall, Precision, and IoU. The proposed model effectively mitigates the challenges of morphological heterogeneity, blurry boundaries, and background interference, leading to highly accurate pixel-level segmentation.
27 Apr 2026
Read appraisal →Physics in medicine and biology
Confounding factors in the generalisation of synthetic CT to diagnostic spine MRI
Objective. Generalisation of synthetic CT (sCT) generation to diagnostic MRI data in the spine faces many challenges, particularly when aiming at accurate visualisation of pathologies for orthopaedic treatment management. In this study we assessed the effect of potential confounding factors on the performance of sCT generation from diagnostic MR images.Approach. Paired spinal diagnostic MR and CT scans from two centres (51 patients) were collected retrospectively, spanning multiple spinal pathologies and regions. Each patient's dataset contained 3D T1-/T2-weighted and 2D T1- and T2-weighted scans. 3D U-Nets were trained per centre on the 3D MR data and CT using a mean absolute error (MAE) loss. The performance was assessed on sCTs derived from original 2D and 3D data and from simulated 2D data with varying orientation (axial/coronal/sagittal) and slice spacing (1.1-6.6 mm). The sCTs were compared to the CTs using MAE, peak signal-to-noise ratio, and dice similarity coefficient.Main results. We identified MR resolution, orientation and contrast as confounding factors for sCT generation from diagnostic MR. Additionally, CT noise and CT-to-MR registration were identified to influence the performance evaluation. Performance degraded with larger slice spacings and significant differences between orientations occurred more often at larger slice spacings. The effect of slice spacing and orientation was stronger inside the vertebrae than outside. While the networks performed better overall on contrasts more similar to the training contrasts, they demonstrated promising performance when trained on 3D T2w images and tested on 2D T1w images. Finally, higher CT noise levels resulted in worse sCT performance metrics, reflected in a significant correlation between noise and MAE.Significance. This work demonstrates that generalisation of sCT generation to diagnostic spine MRI data is hampered primarily by MRI acquisition related aspects including MR resolution, orientation and contrast, which provides guidance for future work on generalisable sCT from diagnostic MR.
27 Apr 2026
Read appraisal →Biomedizinische Technik. Biomedical engineering
Advanced deep learning framework for breast cancer detection using digital breast tomosynthesis images
OBJECTIVES: Early and accurate detection of breast cancer is critical for improving patient survival outcomes. This study proposes a robust deep learning-based framework for breast cancer detection using digital breast tomosynthesis (DBT) images, leveraging both single-slice and multi-slice inputs. METHODS: The proposed work includes image normalization, resizing and Laplacian Pyramid Enhancement (LPE). Various features were extracted and fused in different combinations. To retain the most discriminative features Exhaustive Feature Selection (EFS) is used. A hybrid model integrated using ResNet V2, MobileNet V3 and Inception V3+ for classification. Finally ensemble learning with XGBoost was applied and Hyperparameters were optimized using a grid search strategy. RESULTS: The hybrid model with multi-slice DBT inputs achieved better improvements with accuracy, sensitivity, specificity and area under the curve (AUC). While applying LPE, feature fusion and EFS substantially enhanced the hybrid model's diagnostic performance. CONCLUSIONS: The findings demonstrate the strong potential of advanced methods in enhancing performance. Future research work will focus on integrating this pipeline with clinical decision support systems, multi-center datasets and extending to other breast imaging modalities.
26 Apr 2026
Read appraisal →PloS one
LHW-Net: An ensemble-based machine learning framework for brain tumor classification
The classification of brain tumors is an unsolved problem associated with heterogeneity of tumors and fluctuations in imaging conditions. In this work, the investigation introduces a powerful novel framework, named LHW-Net that combines handcrafted features called local binary patterns (LBP), histogram of oriented gradients (HOG), and wavelet transform (WT). Within the LHW-Net framework, the extracted features are utilized in different machine learning classifiers, such as K-Nearest Neighbors (KNN), Random Forest (RF) and Support Vector Classifier (SVC). The results of the individual classifiers are further combined using probabilistic score fusion approach to improve classification performance. The effectiveness and robustness of the proposed work are validated by the achieved experimental results on commonly accepted benchmark datasets.
24 Apr 2026
Read appraisal →PloS one
Controlled comparative study of YOLOv8-Pose, YOLOv11-Pose, and Detectron2 for vertebrae detection and keypoint estimation
Accurate vertebrae detection with precise keypoint localization is essential for medical image analysis and clinical applications, such as spinal alignment assessment. This study presents a controlled, task-driven comparison of four pose-based deep learning models for vertebrae detection and keypoint estimation: YOLOv8n-Pose, YOLOv11n-Pose, and Detectron2 (Keypoint R-CNN) with ResNet-50 and ResNet-101 backbones. A single-class vertebrae dataset with bounding box annotations and four anatomically meaningful keypoints per vertebra was evaluated under consistent training settings. Performance was assessed in terms of keypoint localization accuracy, detection precision, inference speed, and vertebrae-specific prediction behavior, including detection completeness and duplicate detections, which are critical for clinical usability. Results show that YOLOv11n-Pose provides the best balance between keypoint accuracy and inference efficiency, while YOLOv8n-Pose achieves the fastest inference with competitive performance. Detectron2-based models exhibit lower keypoint accuracy, slower inference, and frequent duplicate predictions that obscure correct anatomical landmarks. Additional experiments indicate that larger YOLO variants (particularly YOLOv8l-Pose) improve accuracy, albeit at increased computational cost. Overall, the results demonstrate that model capacity or architectural recency alone does not guarantee superior performance in vertebrae keypoint detection, highlighting the importance of anatomy-aware model selection for spine imaging applications.
23 Apr 2026
Read appraisal →PloS one
DSCANet: Integrating dual encoder and spatial cross-attention for polyp segmentation
Colonoscopy is a crucial clinical procedure for detecting colorectal polyps, which are strongly associated with the development of colon cancer. This endoscopic technique plays a vital role in both cancer prevention and early diagnosis. Accurate and efficient polyp segmentation is critical for enhancing the diagnostic reliability and clinical utility of colonoscopy. However, achieving precise segmentation presents significant challenges, primarily due to the diversity of polyps in their size and shape, coupled with poorly defined boundaries between polyps and surrounding tissues. To address these challenges, we propose a novel segmentation network, named DSCANet, which is a dual-branch encoder-structured network designed to efficiently fuse body and edge features for high-precision medical image segmentation. DSCANet integrates four key modules: a dual-branch encoder, a spatial cross-attention (SCA) module, a bipolar fusion (BF) module, and a flexible axis-attention (FAA) module. The dual-branch encoder consists of separate body and edge encoders, which extract respective features independently. The SCA module bridges the semantic gap between the two encoders' features. The BF module fuses the shallowest and deepest features, while the FAA module assists the decoder in extracting semantic information from high-level features. DSCANet achieved superior performance on multiple colorectal polyp segmentation datasets. The code is available at https://github.com/Shantdst/DSCANet.
23 Apr 2026
Read appraisal →Scientific reports
A hybrid dual-stream CNN framework with dynamic data augmentation and improved Manta Ray Foraging Optimization for robust glaucoma detection.
A progressive neurological condition, glaucoma is one of the main causes of irreversible blindness in the globe. Early detection is crucial to preventing irreversible vision loss however conventional diagnostic methods are often time-consuming, and heavily reliant on clinical expertise. This study presents an innovative deep learning framework designed to automate glaucoma detection while addressing key challenges such as data imbalance, image quality inconsistencies, and the need for accurate feature extraction. The proposed framework begins with a dedicated preprocessing pipeline that enhances image clarity, and isolates clinically relevant regions of interest. To handle class imbalance, a novel hybrid data augmentation strategy is introduced, combining geometric transformations with adaptive Gaussian noise injection that adjusts intensity according to image characteristics, thereby simulating realistic clinical variability. At its core, the framework leverages a dual-stream CNN architecture that integrates DenseNet121 for structural feature extraction and ResNet50 for texture representation. A lightweight channel-wise attention mechanism is then introduced to selectively emphasize clinically significant channels while suppressing redundant features, thereby balancing efficiency and interpretability. To optimize model performance while reducing computational overhead, an Improved Manta Ray Foraging Optimization (IMRFO) algorithm is employed. IMRFO enhances the standard MRFO with Partial Centroid Opposition-Based Learning (PCOBL) to dynamically fine-tune hyperparameters, including augmentation settings and transfer learning configurations. Experimental validation was conducted on four public benchmark datasets, ACRIMA, Drishti-Gs, ORIGA, and RIM-ONE-DL, demonstrating the framework's superior performance across all evaluation metrics. The model achieved 100.00% accuracy, precision, recall, and AUC on both ACRIMA and Drishti-Gs, with losses of 0.003 and 0.001, respectively. On ORIGA, it reached 99.70% accuracy, 99.80% precision, 99.30% recall, and 99.50% AUC (loss = 0.01). On RIM-ONE-DL, the model scored 99.90% accuracy, 99.70% precision, 99.50% recall, and 99.70% AUC (loss = 0.006). These findings confirm the framework's robustness and clinical applicability for effective glaucoma screening.
19 Apr 2026
Read appraisal →Medical science monitor : international medical journal of experimental and clinical research
Use of Deep Learning Models in the Diagnosis of Proptosis Through Orbital Magnetic Resonance Imaging
BACKGROUND Proptosis is a common manifestation of orbital disease; however, current diagnostic tools, such as the Hertel exophthalmometer and manual radiological measurements, have limited reproducibility and are observer-dependent. More objective, automated approaches are needed. In this single-center retrospective study, orbital magnetic resonance imaging (MRI) examinations from 521 participants (261 with proptosis, 260 controls) were analyzed. Proptosis was defined on MRI using interzygomatic line-based distance criteria. Three-dimensional convolutional neural network models based on DenseNet121, DenseNet169, DenseNet264, and ResNet50 architectures were trained on volumetric orbital MRI data. MATERIAL AND METHODS Data were divided into training, validation, and test sets, and 5-fold cross-validation with early stopping was used to optimize and validate model performance. Diagnostic performance was assessed using accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. RESULTS DenseNet121 achieved the best overall performance, with mean accuracy of 95.0%, AUC of 0.986, sensitivity of 92.7%, and specificity of 96.9% across 5-fold cross-validation. CONCLUSIONS To the best of our knowledge, prior artificial intelligence studies in orbital imaging have primarily focused on CT-based measurements, radiomics approaches, or thyroid-associated orbitopathy assessment rather than end-to-end 3-dimensional deep learning analysis of orbital MRI volumes. In this context, the present study explores a volumetric MRI-based deep learning framework for automated proptosis detection, emphasizing patient-level classification and explainable model interpretation.
19 Apr 2026
Read appraisal →PloS one
LungNet: Leveraging state-space models with SE-enhanced skip connections for precise CT-based lung lesion segmentation
Lung cancer, which accounted for 2.48 million new cases and 1.82 million deaths worldwide in 2022, continues to be the most lethal cancer across the globe, underscoring the urgent demand for more advanced diagnostic tools. Although computed tomography (CT) imaging has long been central to lung cancer detection, the heterogeneous and complex characteristics of lung lesions make accurate segmentation particularly challenging. Current deep learning methods face a critical bottleneck: convolutional neural networks (CNNs) often struggle to capture long-range dependencies due to limited receptive fields, while Transformer-based architectures incur prohibitive computational costs when processing high-resolution CT volumes. Furthermore, standard skip connections in traditional UNet models frequently introduce redundant noise, leading to the dilution of subtle lesion features. To address these specific technical gaps, we introduce a novel deep learning framework that integrates Mamba state-space models with an improved UNet architecture. In this design, to mitigate feature redundancy, Squeeze-and-Excitation networks are embedded into skip connections, while auxiliary losses are introduced to address the degradation of shallow features and capture fine-grained lesion features across varied morphologies. Such a framework not only accommodates the intricate differences in lesion size, shape, and spatial distribution but also achieves a balance between global context modeling and linear computational efficiency. By uniting the local feature extraction strengths of convolutional layers with the long-range dependency modeling power of state-space models, our approach achieves more precise delineation of lung lesion boundaries. Extensive experiments conducted on multiple datasets provide compelling evidence of the method's effectiveness: it attains state-of-the-art segmentation accuracy and demonstrates significant promise for enhancing early detection, ongoing disease monitoring, and treatment planning in lung cancer patients. This advancement delivers a robust solution for the inherently complex task of lung lesion segmentation. Moreover, because lung cancer treatment is costly and insurance coverage plays a decisive role in distributing expenses, the study's outcomes also carry considerable implications for the insurance sector.
17 Apr 2026
Read appraisal →ACS nano
Deep Learning Enables Pixel-Level Nanoparticle Distribution Mapping in Routine Histological Sections by Integrating Cancer Associated Fibroblasts Features
The efficient accumulation and uniform distribution of nanomedicine within tumors are critical for achieving therapeutic outcomes. However, conventional medical imaging technologies struggle to efficiently and accurately detect nanoparticles (NPs) distribution, especially in resolving their cellular-level spatial heterogeneity. Existing deep-learning-based predictive models focus on tumor cell density and vascular distribution but do not address the complex spatial relationship between cancer-associated fibroblasts (CAFs) and drug distribution. This study presents NanoNet, a deep-learning framework that leverages fibroblast activation protein (FAP) immunostaining to spatially characterize CAFs and predict NPs distribution at high resolution. NanoNet achieved high predictive accuracy (ICC = 0.963, R2 = 0.9849) by transforming tumor section images into pixel-level NPs distribution maps. The FAP channel contributed substantially to predictive accuracy, indicating its important role in guiding NPs behavior. This study provides a spatially resolved predictive framework that enables pixel-level predictions of NPs distribution from conventional histological sections, with potential applications for optimizing nanomedicine design and personalized nanomedicine.
15 Apr 2026
Read appraisal →Scientific reports
Region guided mask R-CNN with Haralick ResNet fusion for accurate coronary artery disease detection in computed tomography angiography images
Coronary Computed Tomography Angiography (CCTA) is a non-invasive imaging technique used to visualize the coronary arteries and diagnose coronary artery disease. It provides detailed 3D images which helps in identifying blockages, stenosis and plaques. However, CCTA images often suffer from noise, low contrast, and motion artifacts, which complicate accurate segmentation and analysis. The proposed work addresses the challenges in detecting and classifying coronary artery disease (CAD) from CCTA images. For accurate coronary artery segmentation, Region-Guided Mask R-CNN (RG-Mask R-CNN) combines region growing for initial region identification with Mask R-CNN for precise instance segmentation. This hybrid approach reduces the complexity of the structures in CCTA images and provides better segmentation accuracy. For feature extraction, Haralick-ResNet Fusion (HRF) captures the texture information using Haralick features and ResNet for high-level feature extraction to handle the difficulty of subtle variation differentiation in CCTA images. Finally, a Deep Convolutional Network (DeepConvNet) combines deep learning and convolutional layers to classify CAD efficiently based on the features extracted. This integrated framework offers an accurate solution (98.3%) for CAD detection from CCTA images by overcoming common issues such as noise and complex arterial structures, while improving diagnostic performance.
15 Apr 2026
Read appraisal →Biomedical physics & engineering express
A multi-task cross-attention strategy to segment and classify polyps
Polyps are the main biomarkers for diagnosing colorectal cancer. Their early detection and accurate characterization during colonoscopy procedures rely on expert observations. Nevertheless, such a task is prone to errors, particularly in morphological characterization. This work proposes a multi-task representation capable of segmenting polyps and stratifying their malignancy from individual colonoscopy frames. The approach employs a deep representation based on multi-head cross-attention, refined with morphological characterization learned from independent maps according to the degree of polyp malignancy. The proposed method was validated on the BKAI-IGH dataset, comprising 1200 samples (1000 white-light imaging and 200 NICE samples) with fine-grained segmentation masks. The results show an average IoU of 83.5% and a recall of 94%. Additionally, external dataset validation demonstrated the model's generalization capability. Inspired by conventional expert characterization, the proposed method integrates textural and morphological observations, allowing both tasks, polyp segmentation and the corresponding malignancy stratification. The proposed strategy achieves the state-of-the-art performance in public datasets, showing promising results and demonstrating its ability to generate a polyp representation suitable for multiple tasks.
14 Apr 2026
Read appraisal →European radiology experimental
Impact of deep learning image reconstruction on ADC quantification and histogram metrics: a phantom study
OBJECTIVE: Recently, deep learning (DL)-based reconstruction methods have been introduced into clinical magnetic resonance imaging (MRI) systems to enhance image quality and reduce acquisition time. However, their effects on apparent diffusion coefficient (ADC) maps remain unclear. We investigated whether DL-based image reconstruction influences ADC quantification and histogram-based ADC metrics using a calibrated diffusion-weighted imaging (DWI) phantom. MATERIALS AND METHODS: A phantom containing vials with known ADC values was scanned on a 3-T system using full (fFOV) and reduced (rFOV) field-of-view DWI sequences. Each acquisition was performed using conventional (DL-OFF) and three DL-based strength levels (low, medium, high). Median ADC values were analyzed for repeatability (coefficient of variation (CV)) and accuracy. Histogram changes and first-order radiomic features were assessed using the Wasserstein distance, Friedman, and Wilcoxon tests. RESULTS: ADC estimates showed high repeatability (CV 0.1-1.2%) and good accuracy (deviation -2 to 7%) across all DL levels and sequences. DL reconstruction progressively reduced histogram dispersion, particularly in high-ADC vials. Wasserstein distances increased with DL strength, confirming a progressive effect on ADC value distributions, while median ADC values remained unchanged. Entropy and interquartile range decreased significantly (p < 0.001), whereas kurtosis and skewness increased, with differences showing less stable and sequence-dependent statistical significance. CONCLUSION: DL-based reconstruction maintained accurate and repeatable ADC quantification while reducing the dispersion of ADC values. The effect was more evident for high-ADC regions and the rFOV sequence, resulting in narrower distributions of ADC values. Further investigations comparing different DL-based solutions are warranted to assess the generalizability of these findings in clinical settings. RELEVANCE STATEMENT: Over the past decade, ADC histogram analysis has proven valuable for quantifying tumor heterogeneity, differentiating tumor grade, and evaluating early treatment response. Deep learning reconstruction narrows ADC distributions and reduces dispersion, supporting its potential in oncologic DWI, while highlighting the need for patient-based validation studies. KEY POINTS: DL reconstruction preserved ADC accuracy in both full FOV and reduced FOV DWI. ADC repeatability remained high across DL levels for both DWI sequences. Histogram dispersion progressively reduced across DL levels, particularly in high-ADC vials. Entropy and interquartile ranges decreased progressively with increasing DL strength.
14 Apr 2026
Read appraisal →Biomedical physics & engineering express
The impact of data consistency on deep learning models for nasopharyngeal cancer organ auto-segmentation
Purpose. To investigate how annotation consistency influences deep learning-based auto-contouring performance for organs-at-risk (OARs) in nasopharyngeal cancer radiotherapy.Methods. We evaluated CT scans from 1,301 nasopharyngeal carcinoma patients: 65 contoured by Physician A, 76 by Physician B, and 1,160 by heterogeneous multi-physician teams. Three cohorts (50 samples each for Physicians A/B; 1,000 for multi-physician) with standardized U-Net training protocols generated Models A, B, and C. Model C underwent physician-specific fine-tuning. Performance was quantified via Dice similarity coefficients (DSC) against ground-truth contours across 14 critical OARs.Results. Each model achieved peak accuracy on physician-matched test data. Critically, small-consistency models (A/B) outperformed large-heterogeneous Model C on target cohorts (Model A: 0.777 versus Model C's 0.743 on Test A; Model B: 0.806 versus 0.765 on Test B). Physician-specific fine-tuning closed institutional data gaps, boosting Model C's DSC to 0.795 (+7.08% versus original) on Test A and 0.814 (+6.32%) on Test B-surpassing both original Model C and dedicated small-data models (Model A: +2.39%; Model B: +0.97%).Conclusion. Annotation consistency supersedes dataset scale as the primary performance driver for OAR auto-contouring. Small high-consistency datasets enable optimal native model accuracy, whereas fine-tuning large pre-trained models with targeted physician data closes domain adaptation gaps and delivers state-of-the-art segmentation, advancing precision radiotherapy for head and neck oncology..
11 Apr 2026
Read appraisal →PloS one
MaxGRNet: A multi-axis vision transformer with improved generalization for eye disease classification using explainable AI with insertion-deletion operations on fundus images
Eye diseases, including diabetic retinopathy (DR), glaucoma, and cataracts, represent a major global health concern and can lead to severe visual impairment or blindness if not identified in a timely manner. This study proposes a novel eye disease classification framework based on a multi-axis vision transformer (MaxViT) applied to color fundus images with Explainable Artificial Intelligence (XAI) techniques to enhance model transparency. The proposed architecture integrates transformer-based attention mechanisms with Global Response Normalization (GRN)-based multi-layer perceptron (MLP) layers to capture complex spatial and contextual relationships within fundus images effectively. The model was evaluated on a publicly available eye disease classification dataset using a five-fold cross-validation strategy to assess its robustness and generalization. The experimental results show that the proposed approach consistently outperforms conventional Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) variants, including ResNet50, Swin-T, MaxViT-T, and ViT-B16. The model achieved a macro-averaged test accuracy, precision, and recall values of 96.75%, 96.70%, and 96.80%, respectively, with paired statistical t-tests confirming that these improvements were statistically significant. Rigorous preprocessing techniques were employed to improve data consistency, and XAI-based visual explanations provided insights into the model's decision-making process, supporting interpretability in ophthalmic image analysis. Overall, the proposed MaxViT-based framework is robust and computationally feasible for research-oriented evaluation approaches for automated fundus image classification, highlighting the potential of advanced transformer architectures for future decision-support and research-oriented ophthalmic applications.
10 Apr 2026
Read appraisal →Scientific reports
Research on lung nodule detection in X-ray plain films based on improved YOLOv12 model
To improve automatic lung nodule detection in chest X-ray images, this study proposes an improved YOLOv12-based detection framework by integrating space-to-depth convolution (SPDConv), a dynamic upsampling module (DySample), and a lightweight feature aggregation module (VoVGSCSP). SPDConv enhances spatial information preservation during feature extraction, DySample replaces conventional upsampling to improve multi-scale feature fusion, and VoVGSCSP strengthens feature representation while reducing computational redundancy. The optimized YOLOv12-DSV model was trained and evaluated using a publicly available chest X-ray dataset with annotated lung nodules from the Roboflow platform, with performance assessed through five-fold cross-validation and external testing. Experimental results show that the proposed model achieved an mAP50 of 0.735 and an mAP50-95 of 0.426, outperforming the original YOLOv12 model (mAP50: 0.704; mAP50-95: 0.411). In addition, the proposed model reduced the number of parameters from 2.52 to 2.21 M, decreased FLOPs from 6.0 to 5.2 G, and increased inference speed from 97.6 to 107.8 FPS. These results indicate that the proposed YOLOv12-DSV model improves detection accuracy while reducing computational cost, achieving a more favorable balance between detection performance and model complexity for lung nodule localization in chest X-ray images.
10 Apr 2026
Read appraisal →Physics in medicine and biology
End-to-end testing of an online adaptive radiotherapy workflow for liver tumors using a dynamic Thoraco-abdominal anthropomorphic phantom.
Objective.Adaptive radiotherapy (ART) requires robust end-to-end (E2E) testing tools capable of reproducing organ deformation, physiological motion, and multimodal imaging properties. TheToraco-abdominalAnthropomorphic Phantom withMotion forAdaptiveRadiotherapy (TAM-ARa) is a dynamic anthropomorphic phantom developed to support the design, validation, and quality assurance of online ART workflows.Approach.The phantom was constructed with anatomically realistic bone, lung, and abdominal organ models fabricated from tissue equivalent materials. Modular components allowed simulation of interfractional and intrafractional anatomical variations, including ventilator driven respiratory motion, abdominal deformation, and variable gastric filling. CT, CBCT, and 3T MRI scans were acquired to assess imaging performance. Two E2E tests of an online adaptive intensity-modulated radiotherapy (IMRT) workflow for liver tumors were performed on a Varian Ethos (ETHOS, Varian, USA) system. Multiple deformation scenarios were investigated, including a static reference configuration, two abdominal and one thoracic deformation for ionization chamber (IC) dosimetry, and a static plus abdominal deformation scenarios for radiochromic film dosimetry.Main results.The TAM-ARa phantom demonstrated realistic radiological characteristics, with CT Hounsfield units and MRI relaxation times closely matching reportedin vivovalues. The modular architecture allowed reproducible and controlled abdominal deformations, while setup and dosimeter placement were completed within minutes. In E2E tests for adaptive IMRT of liver tumors, excellent agreement was achieved between measured and planned dose distributions after the treatment plan was adapted. IC E2E measurements showed that online adaptation consistently restored accurate dose delivery under deformation, reducing deviations to below 3%. Independent film measurements confirmed sub-millimeter geometric accuracy and full planning target volume coverage after adaptation, demonstrating effective compensation of deformation-induced errors.Significance.The TAM-ARa phantom provides a versatile and reproducible platform for multimodal E2E testing and validation of online ART workflows. Employing its realistic anatomical design, modular structure, and motion capabilities, TAM-ARa was successfully applied for E2E tests to validate a workflow for adaptive IMRT of liver tumors.
8 Apr 2026
Read appraisal →International ophthalmology
Retinal image-based cardiovascular risk prediction using AI-CRS: a multi-modal deep learning framework
This study presents AI-CRS, an AI-driven deep learning framework for cardiovascular risk assessment using retinal images. By combining convolutional neural networks (CNNs), attention mechanisms, and vasculature segmentation, AI-CRS overcomes the limitations of traditional methods, such as reliance on handcrafted features and single-modality data. The model integrates multi-modal information, fusing vasculature segmentation maps with raw image data to capture subtle vascular changes indicative of cardiovascular diseases. Experimental validation demonstrates that AI-CRS outperforms conventional diagnostic techniques in cardiovascular risk stratification, achieving superior sensitivity, specificity, and accuracy. The model excels in detecting early-stage disease and subtle vascular anomalies, and it shows strong generalizability across diverse datasets, patient demographics, and varying image qualities. Beyond cardiovascular risk, AI-CRS also shows promise in assessing systemic health conditions like hypertension and diabetes, demonstrating the broader utility of retinal imaging in health monitoring. The attention-driven architecture enhances model interpretability, providing clinicians with visual explanations of disease-associated vascular patterns, which is essential for clinical adoption. AI-CRS offers a non-invasive, scalable solution that can be integrated into routine clinical practice, supporting early diagnosis, personalized care, and population-wide screening. Its automated analysis reduces the need for manual feature extraction and subjective interpretation, streamlining workflows and improving efficiency. Ultimately, AI-CRS represents a significant step toward precision medicine by enabling timely interventions, reducing healthcare costs, and improving patient outcomes through non-invasive digital biomarkers.
7 Apr 2026
Read appraisal →PloS one
A robust deep learning approach for impulse noise filtering using hybrid auto-encoder with fuzzy median filter
De-noising convolutional neural networks (DnCNNs), are a powerful nonlinear mapping models in image processing for impulse noise removal. During training and validation, a set of 12 standard testing images is used to evaluate model performance. DnCNNs demonstrate strong capability in classification of impulse noise with excellent results. To evaluate de-noising performance, a suitable noise ratio should be added so that most appropriate DnCNN model can be used for impulse noise detection. This research proposes an effective image restoration technique that integrates DnCNN and an autoencoder with a fuzzy median filter to detect and eliminate high-density impulse noise. The proposed deep learning de-noising technique used to classify noisy and clean pixels, and result are presented in different metrics such as accuracy, FPR, FNR and f1 score. Further, to remove impulse noise an auto-encoder with fuzzy median filter are used that then reconstructs the clean image based with parametric values. Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), are used to assess our methodology, it is compared to conventional impulse noise filtering techniques, experimental results indicate a significant improvement in image quality. Based on the final de-noised images, this research contributes to developing deep learning-based, de-noising techniques that enhance image restoration quality while preserving image details and essential features.
5 Apr 2026
Read appraisal →Hematology (Amsterdam, Netherlands)
Machine learning in the prediction of liver iron concentration and iron chelation therapy adjustment
OBJECTIVES: Iron chelation therapy titration is managed by hematologists who monitor iron levels and adjust medication dosages to achieve patient outcomes. This study developed a machine learning (ML) algorithm to predict the liver iron concentration (LIC) and chelation therapy adjustments. METHODS: This retrospective, single-centre cohort study included adult patients who underwent FerriScan MRI to obtain LIC for chelation therapy monitoring. Using an XgBoost-based ML framework, the proximal-time model (PTM) utilised clinical/drug, laboratory and MRI data features from one visit prior to the target visit, whereas the all-time model (ATM) utilised the data from all prior visits. RESULTS: 94 patients with 892 consecutive visits between January 2008 and November 2018 were included in this study. We assessed the prediction capabilities of the PTM and ATM in LIC, changes to chelation drug type and dosage changes. The PTM model was superior to the ATM model in all the experiments. When drug features were excluded, the CLICT model for predicting patient iron overload status improved to an AUROC of 0.83 [95% CI 0.75-0.91] for PTM; compared to an AUROC of 0.73 [95% CI 0.66-0.80] when drug features were included.For predicting changes in chelation type, the CLICT model showed AUROC of 0.83 [95% CIs 0.77-0.89] for PTM. There is high concordance of the agreement of hematologists with ML in not changing the chelation drug type. DISCUSSION/CONCLUSION: The ML model is a step toward creating a clinical decision support system tool for the prediction of LIC and iron chelation therapy adjustment in patients with haemoglobinopathies or hemolytic anemias.
5 Apr 2026
Read appraisal →Diagnostic and interventional imaging
Deep learning-based image reconstruction significantly improves image quality of MRI examinations of the orbit at 3 Tesla.
PURPOSE: The purpose of this study was to assess the benefit of a deep learning-based image reconstruction (DLBIR) for improving image quality in orbital magnetic resonance imaging (MRI) at 3 Tesla (T). MATERIALS AND METHODS: Seventy-one patients (48 women and 23 men) with a mean age of 52 ± 19.5 (standard deviation [SD]) years (age range: 7-90 years) who underwent MRI examination of the orbit at 3 T between January and June of 2024, were included in the study. Coronal T2-weighted MR images obtained in 70 patients and post-contrast fat-saturated (FS) coronal T1-weighted MR images obtained in 25 patients, were reconstructed with and without DLBIR, resulting in four imaging sets. Two radiologists independently and blindly measured the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the optic nerves on the four imaging sets. Image quality and orbital abnormalities were assessed using a standardized 5-point Likert scale. Comparisons between MR images obtained with and without DLBIR were performed using Wilcoxon test for ordinal and quantitative variables and McNemar test for paired binary data. RESULTS: SNR and CNR of coronal T2-weighted MR images were significantly greater using DLBIR (26.67 ± 9.03 [SD], and 14.87 ± 10.31 [SD], respectively) than without DLBIR (18.91 ± 7.28 [SD], and 9.78 ± 8.47 [SD], respectively) (P < 0.001). There were no differences in SNR and CNR between post-contrast FS T1-weighted images obtained with DLBIR (85.56 ± 63.13 [SD], and 64 ± 41.38 [SD], respectively) and those obtained without DLBIR (91.36 ± 48.49 [SD], and 43.25 ± 20.4 [SD], respectively) (P = 0.35, and P = 0.14, respectively). Qualitatively, good-to-excellent image quality was obtained more frequently with DLBIR than without DLBIR for T2-weighted and post-contrast FS T1-weighted images with respect to optic nerve sharpness (67 % vs. 16 %, and 8 % vs. 0 %, respectively), brain sharpness (90 % vs. 6 %, and 68 % vs. 4 %, respectively), and overall image quality (73 % vs. 1 % and 36 % vs. 0 %, respectively) (all P ≤ 0.001). No significant differences in the detection rates of orbital abnormalities were found between MR images obtained with and without DLBIR, including optic nerve hyperintensity (34 % vs. 31 %, respectively; P = 0.16) and optic nerve atrophy (33 % for both) on T2-weighted images, and optic nerve enhancement on post-contrast FS T1-weighted images (16 % for both). CONCLUSION: DLBIR significantly improves image quality of MRI examinations of the orbit at 3 T, without losing clinically relevant information.
4 Apr 2026
Read appraisal →Physiological measurement
StructEIT: realistic 3D EIT model generation from CT scans for deep learning applications
Objective.Artificial intelligence has significantly improved image reconstruction quality across various medical imaging modalities. However, its application in electrical impedance tomography (EIT) reconstruction remains limited, mainly due to the absence of comprehensivein vivodatasets that incorporate realistic anatomical geometries and conductivity distributions. This limitation constrains the development of supervised and data-driven reconstruction methods.Approach.To address this bottleneck, we developed StructEIT, an integrated EIT modeling framework for generating anatomically and biophysically realistic EIT simulation models. The framework incorporates three key components: (1) a structure extraction module, which automatically processes human CT scans to extract body contours and organ boundaries, thereby providing high-fidelity spatial geometry for three-dimensional finite element modeling; (2) a surface electrode attachment module, which enables flexible and accurate placement of electrodes on irregular body surfaces, supporting diverse configurations and ensuring precise definition of the electrode-tissue interface; and (3) a tissue property assignment module, which establishes frequency-dependent conductivity models for multiple organs, enabling physiologically realistic conductivity values across tissues.Main results and Significance.By bridging the gap between CT imaging and EIT, StructEIT facilitates flexible, realistic, and scalable generation of high-resolution EIT datasets. Using this framework, we constructed Chest-EIT, a thoracic EIT simulation dataset comprising over 1400 publicly available CT cases, with multiple electrode configurations provided for each case.
4 Apr 2026
Read appraisal →IEEE transactions on medical imaging
EviVLM: When Evidential Learning Meets Vision Language Model for Medical Image Segmentation
The disparity between image and text representations, often referred to as the modality gap, remains a significant obstacle for Vision Language Models (VLMs) in medical image segmentation. This gap complicates multi-modal fusion, thereby restricting segmentation performance. To address this challenge, we propose Evidence-driven Vision Language Model (EviVLM)-a novel paradigm that integrates Evidential Learning (EL) into VLMs to systematically measure and mitigate the modality gap for enhanced multi-modal fusion. To drive this paradigm, an Evidence Affinity Map Generator (EAMG) is proposed to collect complementary cross-modal evidences by learning a global cross-modal affinity map, thus refining modality-specific evidence embedding. An Evidence Differential Similarity Learning (EDSL) is further proposed to collect consistent cross-modal evidences by performing Bias-Variance Decomposition on differential matrix derived from bidirectional similarity matrices between image and text evidence embeddings. Finally, the subjective logic is used for mapping the collected evidences to opinions, and the Dempster-Shafer's theory based combination rule is introduced for opinion aggregation, thereby quantifying the modality gap and facilitating effective multi-modal integration. Experimental results on three public medical image segmentation datasets validate that the proposed EviVLM can achieve state-of-the-art performance. Code is available at: https://github.com/QingtaoPan/EviVLM.
4 Apr 2026
Read appraisal →European radiology
Attention-based deep learning network for predicting World Health Organization meningioma grade and Ki-67 expression based on magnetic resonance imaging
OBJECTIVES: Preoperative assessment of World Health Organization (WHO) meningioma grading and Ki-67 expression is crucial for treatment strategies. We aimed to develop a fully automated attention-based deep learning network to predict WHO meningioma grading and Ki-67 expression. MATERIALS AND METHODS: This retrospective study included 952 meningioma patients, divided into training (n = 542), internal validation (n = 96), and external test sets (n = 314). For each task, clinical, radiomics, and deep learning models were compared. We used no-new-Unet (nn-Unet) models to construct the segmentation network, followed by four classification models using ResNet50 or Swin Transformer architectures with 2D or 2.5D input strategies. All deep learning models incorporated attention mechanisms. RESULTS: Both the segmentation and 2.5D classification models demonstrated robust performance on the external test set. The segmentation network achieved Dice coefficients of 0.98 (0.97-0.99) and 0.87 (0.83-0.91) for brain parenchyma and tumour segmentation. For predicting meningioma grade, the 2.5D ResNet50 achieved the highest area under the curve (AUC) of 0.90 (0.85-0.93), significantly outperforming the clinical (AUC = 0.77 [0.70-0.83], p < 0.001) and radiomics models (AUC = 0.80 [0.75-0.85], p < 0.001). For Ki-67 expression prediction, the 2.5D Swin Transformer achieved the highest AUC of 0.89 (0.85-0.93), outperforming both the clinical (AUC = 0.76 [0.71-0.81], p < 0.001) and radiomics models (AUC = 0.82 [0.77-0.86], p = 0.002). CONCLUSION: Our automated deep learning network demonstrated superior performance. This novel network could support more precise treatment planning for meningioma patients. KEY POINTS: Question Can artificial intelligence accurately assess meningioma WHO grade and Ki-67 expression from preoperative MRI to guide personalised treatment and follow-up strategies? Findings The attention-enhanced nn-Unet segmentation achieved high accuracy, while 2.5D deep learning models with attention mechanisms achieved accurate prediction of grades and Ki-67. Clinical relevance Our fully automated 2.5D deep learning model, enhanced with attention mechanisms, accurately predicts WHO grades and Ki-67 expression levels in meningiomas, offering a robust, objective, and non-invasive solution to support clinical diagnosis and optimise treatment planning.
4 Apr 2026
Read appraisal →Yonsei medical journal
Deep Learning for Automated Tumor Segmentation of Rectal Cancer on T2-Weighted Magnetic Resonance Images
PURPOSE: To develop a deep learning (DL)-based automated segmentation model for rectal cancer on T2-weighted (T2W) magnetic resonance (MR) images. MATERIALS AND METHODS: A total of 458 patients who underwent baseline rectal MR imaging were retrospectively enrolled. An experienced radiologist labeled the tumor on each slice of the selected T2W axial images covering the entire tumor mass, and tumor volume was measured during labeling. Another radiologist labeled the rectum on the same images. Attention U-Net was trained on the T2W images from the training dataset to classify each voxel as tumor or non-tumor. Segmentation with or without rectum guidance was compared using evaluation metrics, including the Dice similarity coefficient (DSC), in the test dataset. RESULTS: The tumor segmentation DSC without rectum guidance was 73.35% (71.97-74.71); that with the predicted rectum was 71.79% (70.53-73.05); and that with the overlaid rectum was 75.52% (74.32-76.62), which was the highest among the three segmentation models. Tumor volume showed a positive correlation with tumor segmentation accuracy with the predicted rectum (r=0.333, p=0.001; overlaid rectum, r=0.319, p=0.002) and without rectum guidance (r=0.219, p=0.036). CONCLUSION: A DL-based automated segmentation model can predict RC with over 70% accuracy, which was further improved with overlaid rectum, and larger tumor volume.
4 Apr 2026
Read appraisal →Journal of biophotonics
Cell-MICS: Detecting Immune Cells With Label-Free Two-Photon Autofluorescence and Deep Learning
Multiphoton imaging has been widely used for deep-tissue imaging. Although its label-free, metabolic contrast is ideal for investigating inflammation, the label-free two-photon induced autofluorescence is often regarded as less specific compared to conventional antibody markers. In this work, we investigate the potential for multiphoton imaging with computational specificity (MICS) by training a convolutional neural network on images of different immune cells. A low-complexity squeezeNet architecture was able to achieve reliable immune cell classification results (0.89 ROC-AUC, 0.95 PR-AUC for binary classification between T cells and neutrophils; 0.689 F1 score, 0.697 precision, 0.748 recall for multi-class classification between six isolated cell types). Perturbation tests confirmed that the model was not confused by the extracellular environment and that 2P-AF from NADH and FAD is equally important for the classification. In the future, deep learning could provide computational specificity for specific immune cells in unstained tissues, with great potential for label-free in vivo endomicroscopy.
4 Apr 2026
Read appraisal →AJNR. American journal of neuroradiology
Multimodal CT Perfusion-Based Deep Learning for Predicting Stroke Lesion Outcomes in Complete and No Recanalization Scenarios
BACKGROUND AND PURPOSE: Predicting the final location and volume of lesions in acute ischemic stroke is crucial for clinical management. While CTP is routinely used for estimating lesion outcomes, conventional threshold-based methods have limitations. We developed specialized outcome-prediction deep learning models that predict infarct core in successful reperfusion cases and the combined core-penumbra region in unsuccessful reperfusion cases. MATERIALS AND METHODS: We developed single-modal and multimodal deep learning models using CTP parameter maps to predict the final infarct lesion on follow-up DWI. Using a multicenter data set from multiple sites, we developed deep learning models and evaluated them separately for patients with complete recanalization (successful reperfusion [CR], n = 350) and no recanalization (unsuccessful reperfusion [NR], n = 138) after treatment. The CR model was designed to predict the infarct core region, while the NR model predicted the expanded, hypoperfused tissue encompassing both the core and penumbra regions. Five-fold cross-validation was performed for robust evaluation. RESULTS: The multimodal 3D nnU-Net model demonstrated superior performance, achieving mean Dice scores of 35.36% in patients with CR and 50.22% in those with NR. This model substantially outperformed the current clinically used method, providing more accurate outcome estimates than the conventional single-technique threshold-based measures, which yielded Dice scores of 15.73% and 39.71% for CR and NR groups, respectively. CONCLUSIONS: Our approach offered both successful reperfusion and unsuccessful reperfusion estimations for potential treatment outcomes, enabling clinicians to better evaluate treatment eligibility for reperfusion therapies and assess potential treatment benefits. This advancement facilitates more personalized treatment recommendations and has the potential to substantially enhance clinical decision-making in acute ischemic stroke management by providing more accurate tissue outcome predictions than conventional single-technique threshold-based approaches.
4 Apr 2026
Read appraisal →Japanese journal of radiology
Super-resolution deep learning reconstruction improves brain MRI quality and detection of metastases.
PURPOSE: Accurate identification of brain metastases is critical for determining prognosis and guiding treatment. Deep learning reconstruction (DLR) enhances MRI quality by reducing noise, while super-resolution DLR (SR-DLR) may further improve spatial resolution and lesion detectability. To evaluate SR-DLR versus conventional DLR in detecting and visualizing brain metastases on postcontrast T1-weighted brain MRI. MATERIALS AND METHODS: This retrospective study included 47 consecutive patients who underwent postcontrast 3D whole-brain T1-weighted MRI between July and December 2024. Images were reconstructed using both SR-DLR and DLR. Three independent readers evaluated metastatic lesion detection and rated image quality. Subjective assessments included lesion visibility, visibility of normal structures, sharpness, noise, and overall image quality. Objective metrics-full width at half maximum (FWHM), edge rise distance (ERD), edge rise slope (ERS), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR)-were also measured. Statistical tests included jackknife alternative free-response receiver operating characteristic (JAFROC) analysis, Wilcoxon signed-rank test, McNemar's test, and paired t-tests, with significance threshold of p < 0.050. RESULTS: A total of 117 brain metastases were detected in 47 patients (mean age, 59 years ± 18; 27 men). SR-DLR demonstrated significantly better lesion detection performance than DLR (mean figure of merit = 0.842 vs. 0.797; p = 0.042). Subjective image quality ratings favored SR-DLR for lesion and structure visibility, sharpness, noise, and overall quality in most cases. Objectively, SR-DLR yielded lower FWHM (1.2 mm vs. 1.9 mm; p < 0.001), higher ERS (791.3 mm- 1 vs. 645.3 mm- 1; p = 0.013) indicating enhanced sharpness as well as improved CNR (27.5 vs. 24.9; p < 0.001) compared to DLR. CONCLUSION: Compared to DLR, SR-DLR significantly enhances brain MRI quality and improves detection of metastatic lesions.
4 Apr 2026
Read appraisal →Developmental cognitive neuroscience
Deep learning in fetal, infant, and toddler neuroimaging research.
Artificial intelligence (AI) is increasingly being integrated into everyday tasks and work environments. However, its adoption in medical image analysis has progressed more slowly due to high clinical stakes, limited availability of labeled data, and substantial variability in imaging protocols and population. These challenges are further pronounced in the field of fetal, infant, and toddler (FIT) neuroimaging, where datasets are especially scarce and subject to large amounts of anatomical variability. However, deep learning (DL), a specific method within machine learning, which is itself a subfield of AI, has emerged as a powerful framework to adapt to the challenges of medical image analysis. This review is written for the broad FIT research community, including clinicians, neuroscientists, and develop mental scientists who may not have formal training in AI. To make the material accessible, we provide a concise overview of DL concepts before reviewing a selected, and non-exhaustive, list of applications of DL in FIT neuroimaging, including structural image analysis, enhancement of data acquisition, modeling of cognitive and perceptual processes, and automated video tagging. In closing, we discuss best practices for data curation, ongoing challenges, and opportunities for future research.
4 Apr 2026
Read appraisal →European radiology
Deep learning models for radiography body-part classification and chest radiograph projection/orientation classification: a multi-institutional study.
OBJECTIVES: Large-scale radiographic datasets often include errors in labels such as body parts or projection, which can undermine automated image analysis. Therefore, we aimed to develop and externally validate two deep-learning models-one for categorising radiographs by body part, and another for identifying projection and rotation of chest radiographs-using large, diverse datasets. MATERIALS AND METHODS: We retrospectively collected radiographs from multiple institutions and public repositories. For the first model (Xp-Bodypart-Checker), we included seven categories (Head, Neck, Chest, Incomplete Chest, Abdomen, Pelvis, Extremities). For the second model (CXp-Projection-Rotation-Checker), we classified chest radiographs by projection (anterior-posterior, posterior-anterior, lateral) and rotation (upright, inverted, left rotation, right rotation). Both models were trained, tuned, and internally tested on separate data, then externally tested on radiographs from different institutions. Model performance was assessed using overall accuracy (micro, macro, and weighted) as well as one-vs.-all area under the receiver operating characteristic curve (AUC). RESULTS: In the Xp-Bodypart-Checker development phase, we included 429,341 radiographs obtained from Institutions A, B, and MURA. In the CXp-Projection-Rotation-Checker development phase, we included 463,728 chest radiographs from CheXpert, PadChest, and Institution A. The Xp-Bodypart-Checker achieved AUC values of 1.00 (99% CI: 1.00-1.00) for all classes other than Incomplete Chest, which had an AUC value of 0.99 (99% CI: 0.98-1.00). The CXp-Projection-Rotation-Checker demonstrated AUC values of 1.00 (99% CI: 1.00-1.00) across all projection and rotation classifications. CONCLUSION: These models help automatically verify image labels in large radiographic databases, improving quality control across multiple institutions. KEY POINTS: Question This study examines how deep learning can accurately classify radiograph body parts and detect chest projection/orientation in large, multi-institutional datasets, enhancing metadata consistency for clinical and research workflows. Findings Xp-Bodypart-Checker classified radiographs into seven categories with AUC values of over 0.99 for all classes, while CXp-Projection-Rotation-Checker achieved AUC values of 1.00 across all projection and rotation classifications. Clinical relevance Trained on over 860,000 multi-institutional radiographs, our two deep-learning models classify radiograph body-part and chest radiograph projection/rotation, identifying mislabeled data and enhancing data integrity, thereby improving reliability for both clinical use and deep-learning research.
4 Apr 2026
Read appraisal →Clin Obstet Gynecol
Artificial Intelligence in Obstetrics: Current Applications, Opportunities, and Clinical Implementation Challenges
Artificial intelligence is transforming obstetric practice through applications in diagnostic imaging, risk prediction, and clinical decision-making. Deep learning algorithms have achieved diagnostic accuracy comparable to that of experienced clinicians. However, gaps persist between algorithmic capability and clinical implementation. Critical challenges include limited external validation and algorithmic bias. This review examines current AI applications in obstetrics across multiple clinical domains: automated fetal biometry, structural anomaly detection, prediction of pregnancy complications, and intrapartum fetal surveillance. It highlights persistent technical, ethical, and implementation barriers. Key recommendations include multicenter validation across diverse populations, explainable AI approaches, and creating strong regulatory frameworks.
1 Mar 2026
Read appraisal →