Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 15 appraisals
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 →European radiology experimental
Automated identification of MRI series using a hierarchical modular machine-learning pipeline
OBJECTIVE: The volume and diversity of large MR imaging datasets require efficient automated labelling tools for cataloguing MR series, as manual annotation is impractical and costly. However, relying on DICOM header fields alone is unreliable because sequence descriptors are heterogeneous and locally defined, frequently missing or incorrect, and may be altered or removed during anonymisation. MATERIALS AND METHODS: We developed an AI-based modular model to classify MR series. The pipeline comprises five sequential classifiers (Family, Weighting, Fat Suppression, Contrast, and Others) and was trained and tested on 18,181 MRI series from the multicentre PRIMAGE repository. The dataset was split by patient into 80% training/validation and 20% testing; within the training/validation subset, five-fold cross-validation was used. With the exception of contrast classification, all modules used DICOM tag-based machine learning models (CatBoost/Random Forest), while the Contrast classifier incorporated image analysis using a pretrained single-slice ResNet-50. Ethical approval for the study was obtained. RESULTS: Accuracy was 0.994 for Weighting, 0.984 for Family, 0.959 for Fat suppression and 0.958 for Others; the Contrast classifier recorded 0.841. Overall, the end-to-end classification yielded a weighted F1 of 0.849 (CI: 0.837-0.861) and an accuracy of 0.853 (CI: 0.841-0.865). CONCLUSION: The proposed approach provides a reliable and scalable solution for labelling large, heterogeneous MRI datasets across multiple anatomical regions. The pipeline achieved excellent performance for Weighting and Family classification, solid performance for Fat Suppression and 'Others'. However, Contrast classification remains the main limitation and warrants further refinement and/or additional modules. RELEVANCE STATEMENT: Reliable, automated MRI sequence labelling enables faster, reproducible cohort selection and protocol harmonisation in large archives, supporting downstream clinical research and AI tools while reducing manual curation effort and error. KEY POINTS: Multicentre PRIMAGE cohort (18,181 MR series) enabling evaluation in a heterogeneous setting. Machine learning model combining DICOM metadata and image features. High accuracy in series classification: up to 0.994 on key tasks. A scalable pipeline reduces manual annotation workload for radiologists.
30 May 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 →Computer assisted surgery (Abingdon, England)
An automatic detection model for spread through air spaces in postoperative pathological sections based on deep learning in NSCLC
Spread through air spaces (STAS) is recognized as an aggressive pattern of invasion in lung cancer and has been associated with poorer survival outcomes. However, STAS is frequently overlooked or misdiagnosed during routine pathological diagnosis. We analyzed 129 pathological slides from 91 STAS-positive patients with non-small cell lung cancer (NSCLC). An artificial intelligence framework consisting of a tumor region segmentation algorithm and an object detection algorithm was developed. The segmentation algorithm was first used to isolate the main tumor region, followed by STAS detection using an improved object detection model. The segmentation module achieved a Jaccard similarity of 0.846. The proposed object detection algorithm demonstrated superior performance with a precision of 0.738, recall of 0.747 and average precision (AP) of 0.784 and F1 score of 0.742, outperforming other object detection methods. Both segmentation and detection results met diagnostic requirements, with higher accuracy observed in lung adenocarcinoma (LUAD) than in lung squamous cell carcinoma (LUSC). In external validation, the model achieved a precision of 0.670, recall of 0.705 and AP of 0.732 and F1 score of 0.687. Additionally, the AI-derived STAS counts showed substantial agreement with pathologists (interclass correlation coefficient = 0.703). Kaplan-Meier survival analysis revealed that the number of STAS events was significantly associated with disease-free survival (DFS) in stage I LUAD (p < .05). We propose a deep learning-based artificial intelligence framework for automated STAS detection and quantification in digital whole-slide images of NSCLC. This model holds promising potential to assist pathologists in achieving more comprehensive and accurate diagnosis of STAS.
25 May 2026
Read appraisal →Philosophical transactions. Series A, Mathematical, physical, and engineering sciences
Unexpected benefits of self-modelling in neural systems
Self-models have been a topic of interest for decades in human cognition and more recently in machine learning. Yet what benefits do self-models confer? Here we show that when an artificial network learns to predict its internal states as an auxiliary task, the network becomes simpler, more regularized and more parameter efficient. To test the hypothesis of self-regularizing through self-modelling, we used a range of network architectures performing three classification tasks across two modalities. In all cases, adding self-modelling caused a significant reduction in network complexity. The reduction was observed in two ways. First, the distribution of weights was narrower when self-modelling was present. Second, a measure of network complexity, the real log canonical threshold (RLCT), was smaller when self-modelling was present. These results support the hypothesis that self-modelling has a restructuring effect, reducing complexity and increasing parameter efficiency. This self-regularization may help explain some of the benefits of self-models reported in recent machine learning literature, as well as the adaptive value of self-models to biological systems. In particular, these findings may shed light on the possible interaction between the ability to model oneself and the ability to be more easily modelled by others in a social or cooperative context. This article is part of the theme issue 'World models in natural and artificial intelligence'.
17 May 2026
Read appraisal →JMIR mental health
Large Language Models and Their Applications in Mental Health: Scoping Review
BACKGROUND: Large language models (LLMs) are poised to transform mental health care, offering advanced capabilities in diagnosis, prognosis, and decision support. Since their inception, numerous mental health-focused LLMs have emerged in the scientific literature, reflecting the growing interest in leveraging these models across various clinical applications. With a broad range of models available, diverse optimization strategies, and multiple use cases, reviewing the current landscape is critical to understanding where future impact lies. OBJECTIVE: This study aimed to conduct a scoping review investigating the use of LLMs in mental health across diagnostic, prognostic, and decision support tasks. METHODS: We screened 3121 papers from PubMed, Scopus, and Web of Science for studies published between January 2023 and October 2025, using terms related to LLM and mental health. After removing duplicates, 2 reviewers (MCL and WWBG) independently screened the studies, with a third (JJK) to resolve conflicting opinions. We extracted and synthesized information on the models, use cases, datasets, and adaptation methods from selected papers. RESULTS: In total, 41 papers were selected. Many studies included evaluations on OpenAI's GPT series applications: GPT-4 (24 studies, 58.5%) and GPT-3.5 (16 studies, 39%). Others included Bidirectional Encoder Representations from Transformers-derived models (9 studies, 22%), LLaMA (8 studies, 19.5%), and RoBERTa-derived models (6 studies, 14.6%). While all studies initially applied out-of-the-box LLMs, several adapted them through few-shot learning or fine-tuning to better align with specific research goals. The most common use case was in diagnostics (31 studies, 75.6%), while the most common target condition was depression (11 studies, 26.8%). While many studies reported superior performance of LLMs, only a minority of studies (13 studies, 31.7%) validated LLM performance against clinician assessments using real patient data, with the majority relying on proxy outcomes such as clinical vignettes, examination questions, or social media posts. CONCLUSIONS: Despite rapid growth and diversity of LLM applications in mental health, the field remains nascent and exploratory. Future developments must emphasize consistent model adaptation procedures to ensure safety and clinical workflow alignment. Models must also be evaluated on robust evaluation criteria by using standardized protocols and real clinical outcome measures.
16 May 2026
Read appraisal →Physics in medicine and biology
ReaderAdaptNet: modeling reader variability in breast imaging with reader-specific embeddings
Objective.Inter-reader variability remains a major challenge in breast imaging interpretation, particularly for ordinal classification tasks such as breast density and background parenchymal enhancement (BPE). These visual assessments are prone to inconsistencies due to subjective interpretation, limiting the reliability of AI-based models trained on aggregated or noisy labels. This work aims to account for this variability rather than collapse it.Approach.We propose ReaderAdaptNet, a reader-adaptive network that explicitly models inter-reader variability through reader-specific embeddings. A novel two-stage deep learning framework underpins the training process of the model. The first stage jointly learns discriminative image features and low-dimensional embeddings representing individual annotation styles, enabling personalized classification. The second stage allows for embedding calibration using a small set of labeled examples, enabling rapid adaptation to new readers or consensus definitions without retraining the full model. The method is designed for multi-reader datasets and is evaluated on two breast classification tasks: breast density classification from full-field digital mammography and BPE classification from contrast-enhanced mammography.Main results.Results demonstrate that reader embeddings improve both individual and consensus-level performance, and calibrated embeddings enable flexible, low-cost personalization. Reader-specific embeddings of dimension 32 improved the mean classification accuracy across readers, increasing from 76.4% to 84.4% for breast density, and from 65.1% to 72.1% for BPE, compared to the proposed baseline method.Significance.By disentangling stable image features from reader-specific decision tendencies, ReaderAdaptNet provides a parameter-efficient, interpretable route topersonalization(alignment to a given reader) orunification(alignment to an institutional standard) under real-world variability.
16 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 →Artificial intelligence in medicine
Learning with less: A survey of deep learning in medical imaging under varying supervision levels
The rapid evolution of deep learning has significantly advanced the field of medical image analysis. However, despite these achievements, further progress is hindered by the scarcity of large, well-annotated datasets. To overcome this limitation, recent years have seen a growing focus on developing methods that leverage varying levels of supervision as alternatives to fully supervised learning. Instead of relying solely on complete annotations, these approaches employ partial annotations or other supervision strategies to train deep learning models. This paper presents a systematic review of deep learning methods for medical image analysis under different levels of supervision. To this end, we categorize these methods based on the level of supervision they rely on, encompassing categories such as no supervision, inexact supervision, incomplete supervision, inaccurate supervision, and only limited supervision. We further divide these categories into finer subcategories. For example, we categorize inexact supervision into multiple instance learning and learning with weak annotations. Similarly, we categorize incomplete supervision into semi-supervised learning, active learning, and domain-adaptive learning and so on. Furthermore, we systematically summarize commonly used datasets for different types of supervision in deep learning for medical image analysis and explore future research directions to conclude this survey.
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 →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 →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
SCGAN-generated skin-tone-balanced extension of the HAM10000 dataset for fairer dermatological AI
This dataset presents a synthetic collection of skin lesion images generated using a tone-conditioned Generative Adversarial Network (GAN) to model skin tone variation in dermatological datasets, motivated by the limited representation of darker skin tones in commonly used benchmarks. The dataset includes 10,000 images, at a resolution of 128×128 pixels, equally divided between medium and dark skin tones, with lesion class labels assigned to follow the class distribution of the original HAM10000 dataset. To quantitatively characterize the generated data, we report Fréchet Inception Distance (FID) scores and a baseline classification experiment comparing models trained on HAM10000 alone versus HAM10000 augmented with the proposed synthetic images. This work provides a controlled synthetic extension intended to support fairness-aware experimentation, data augmentation, and downstream evaluation in dermatological machine learning.
20 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 →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 →