Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 54 appraisals
Journal of pediatric surgery
Interpretable deep learning model for pediatric strangulated small bowel obstruction on CT: A multicenter study
OBJECTIVE: To develop and validate a deep learning-based multi-instance learning (MIL) model that integrates CT imaging and clinical data to improve the accuracy of discriminating between strangulated small bowel obstruction (StSBO) from simple small bowel obstruction (SiSBO) in pediatric patients. MATERIALS AND METHODS: This multicenter retrospective study, conducted between January 2018 and June 2024, enrolled hospitalized pediatric patients aged 1-14 years with a diagnosis of small bowel obstruction. We developed the clinical, multi-instance learning, and combined models based on CT and clinical features. Model performance was evaluated using receiver operating characteristic (ROC) analysis, while SHapley Additive exPlanations (SHAP) interpreted feature contributions. We further assessed whether MIL-assisted diagnosis could enhance physician accuracy in diagnosing StSBO. RESULTS: The study sample comprised 168 patients (mean age, 6.36 ± 3.97, 118 men). Ascites and closed-loop sign were identified as independent predictors of StSBO on multivariate analysis (both p< 0.05). The MIL model achieved the area under the curve (AUC) of 0.86 (95%CI 0.70-1.00), p= 0.01 in the external test cohort. The combined model showed the highest diagnostic performance (AUC 0.87, 95%CI 0.72-1.00, p= 0.01) in the external test cohort, with MIL-derived features showing predominant importance in SHAP analysis. Both junior and experienced radiologists and surgeon demonstrated improved diagnostic performance with MIL assistance, showing AUC increases of 16%, 2%, and 20%, respectively. CONCLUSIONS: The MIL model performed well in diagnosing StSBO, and clinical data integration improved its performance. As a decision support tool, the model may aid risk stratification and facilitate timely escalation of care in pediatric StSBO management.
3 Aug 2026
Read appraisal →Seminars in musculoskeletal radiology
Artificial Intelligence (Pattern Recognition) in Musculoskeletal Imaging: The Future or Hype?
Artificial intelligence and automated pattern recognition, in particular, have been described as the next frontier in musculoskeletal imaging. However, as the initial hype phase transitions into clinical reality, an essential evaluation of these technologies is required. This narrative review examines the dichotomy between a potential future in which artificial intelligence offers unprecedented efficiency in automatizing multiple tasks in musculoskeletal radiology, from fracture detection, automated segmentation, to automated reporting, versus the hype, characterized by deep learning models that lack generalizability across different scanner vendors and patient populations. We explore the black box nature of deep learning and the ethical implications of automation. By analyzing current barriers to deployment, including workflow integration and regulatory hurdles, this article argues that although artificial intelligence holds transformative potential for musculoskeletal radiology, its success depends on moving beyond narrow diagnostic tasks toward robust multi-institutional validation, to evolve from a speculative trend into an essential clinical copilot tool.
2 Aug 2026
Read appraisal →Biomedical physics & engineering express
BMSNet: a boundary-guided multi-scale polyp segmentation network
Accurate polyp segmentation in colonoscopy images is essential for computer-aided diagnosis and early colorectal cancer screening. However, this task remains challenging due to large variations in polyp size, shape, and texture, as well as blurred boundaries and low contrast between polyps and surrounding mucosa. To address these challenges, we propose boundary-guided multi-scale polyp segmentation network (BMSNet), a boundary-guided multi-scale network for accurate polyp segmentation. First, an Scale-Aware Modulation Meets Transformer-T backbone initialized with officially released pretrained weights is adopted to extract multi-level features efficiently. Then, a boundary prediction module is introduced to generate preliminary boundary priors through bidirectional feature interaction and differentiable Canny-based auxiliary supervision. Based on the generated boundary information, we design a boundary-guided feature enhancement module to refine feature representations. This module combines frequency-domain enhancement, multi-scale feature interaction, and boundary-guided modulation. Finally, a feature fusion unit is introduced to progressively integrate multi-scale features and generate the final segmentation prediction. Extensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed method. On ClinicDB, Kvasir, ColonDB, ETIS, and CVC-300, BMSNet achieves Dice scores of 0.943, 0.920, 0.828, 0.825, and 0.913, respectively. These results indicate that BMSNet achieves competitive segmentation performance with moderate computational complexity under the evaluated experimental setting.
28 July 2026
Read appraisal →PloS one
A novel dynamic Optuna hybrid Harris Hawks Optimization approach for classification of CAD
Coronary Artery Disease (CAD) is a leading cause of mortality worldwide and is primarily associated with atherosclerotic plaque formation, resulting in coronary artery stenosis. The accurate prediction of CAD using deep learning models is often constrained by the limitations of conventional optimization techniques, including premature convergence and limited adaptability of the models. To address these challenges, this study proposes a Dynamic Optuna Hybrid Harris Hawks Optimization (Dynamic Optuna H-HHO) framework to enhance the performance of deep learning-based CAD prediction models. The proposed approach integrates dynamic parameter adjustment, adaptive escape energy mechanisms, and Optuna-based hyperparameter tuning. The framework was applied to optimize several deep-learning classifiers, including ResNet-50, VGG-16, InceptionV3, and MobileNet, using a coronary artery stenosis dataset. The performance was evaluated through a comparative analysis with models optimized using the conventional Hybrid Harris Hawks Optimization (H-HHO) algorithm. The experimental results indicate that the proposed Dynamic Optuna H-HHO framework consistently improves the predictive accuracy across all evaluated models. InceptionV3 achieved the highest accuracy of 97.9%, followed by MobileNet with 97.6%, compared with the maximum accuracy of 82.46% obtained using traditional HHO-based optimization. By combining adaptive optimization strategies with automated hyperparameter tuning, the proposed framework provides a robust and scalable solution for improving the accuracy of coronary artery disease prediction.
20 July 2026
Read appraisal →Physics in medicine and biology
Four-dimensional on-beam computed tomography reconstruction using projection-difference images
Purpose.Accurate verification of delivered dose requires anatomical information during the treatment beam delivery. However, conventional cone-beam computed tomography (CBCT) is typically acquired prior to beam delivery and fails to account for intrafractional tumor shift/drift. This study aims to develop a 4D on-beam computed tomography (4D OBCT) framework to capture anatomical states during irradiation.Materials and methods.A thoracic phantom with rigid tumor motion was used to develop a conformal arc treatment plan. During delivery, the target was displaced by 1 cm superior to its planned position to simulate a baseline shift, and megavoltage transmission data were acquired using an electronic portal imaging device. Transmission data corresponding to the planned tumor position were obtained from the planning CT (pCT) via Monte Carlo simulation. Volumetric difference images were reconstructed from the differences between the measured and simulated transmission data for each phase. Subsequently, the images were superimposed onto the corresponding phases of the 4D pCT to obtain 4D OBCT. For evaluation, we compared 4D OBCT with 4D pCT (ground-truth) and 4D CBCT which were acquired after the target displacement. A dosimetric performance was evaluated by recalculating 4D dose distributions on the three images and comparing in dose-volume histograms and gamma indices.Results.4D OBCT achieved lower root mean square error than 4D CBCT in both the lung (58.46 HU vs 74.37 HU) and chest wall regions (23.78 HU vs 66.84 HU). This quantitative accuracy was also reflected in the dosimetric comparison. Under 2%/2 mm and 1%/1 mm criteria, 4D OBCT achieved gamma pass rates of 99.76% and 99.08%, respectively, whereas 4D CBCT yielded 72.50% and 57.77%, respectively.Conclusions.The proposed 4D OBCT framework successfully reconstructed accurate phase-resolved images of the target during irradiation. By providing high-resolution volumetric images that reflect intrafractional motion, 4D OBCT could provide a robust solution for precise post-delivery dosimetric verification and adaptation.
12 July 2026
Read appraisal →Korean journal of radiology
Agentic Artificial Intelligence for the Automated Generation of Accurate Summary Podcasts of Radiology Research Papers
OBJECTIVE: To evaluate whether a custom agentic artificial intelligence (AI) pipeline can overcome the limitations of general-purpose large language model tools, when compared with a generic commercial tool (Google NotebookLM [NBLM]), for generating podcast-style summaries of radiology research articles. MATERIALS AND METHODS: Twenty-two PDF-format original research articles published in the April 2025 issue of Radiology were processed using our Programmable, Phoneme-Aware PDF-to-Podcast Pipeline (P5) and NBLM to generate 44 audio episodes. P5 utilizes a multi-agent workflow for script generation, quality assurance, pronunciation enhancement, and audio synthesis. Four radiologists from a pool of 25 (7 generalists and 18 specialists) were randomly assigned to evaluate each blinded audio episode, yielding 176 total evaluations. The primary outcomes were the number of hallucinations (factual errors) per episode and the percentage of hallucination-free episodes. Secondary outcomes included the number of inappropriate statements, mispronunciations, and flow disruptions; the composite quality score (Quality Assessment of Educational Podcasts [QAEP]); the key results coverage score; and overall listener preference. Data were analyzed using generalized linear mixed models. RESULTS: The P5 method produced significantly fewer hallucinations per episode compared with NBLM (mean, 0.32 vs. 0.93; P < 0.001) and a higher proportion of hallucination-free episodes (71.6% [63/88] vs. 56.8% [50/88]; P = 0.013), consistently across generalists and specialists. P5 demonstrated significantly fewer mispronunciations (mean, 0.11 vs. 1.62; P < 0.001) and flow disruptions (mean, 0.19 vs. 1.06; P < 0.001) per episode, as well as higher mean QAEP composite scores (4.62 vs. 4.28; P < 0.001), compared with NBLM, consistently across generalists and specialists. Raters preferred P5 over NBLM in 72.7% (64/88) of comparisons (P = 0.003). CONCLUSION: Our custom agentic AI pipeline generated podcast-style summaries of radiology research articles with significantly higher quality and greater listener preference than the generic commercial tool.
3 July 2026
Read appraisal →Experimental eye research
Artificial intelligence applications in OCT and OCTA for diabetic retinopathy: A systematic review
PURPOSE: Diabetic retinopathy (DR) is a leading cause of vision impairment worldwide. Optical coherence tomography (OCT) and OCT angiography (OCTA) provide detailed retinal imaging, enabling early detection of microvascular changes. This study aims to systematically review artificial intelligence (AI), particularly deep learning (DL), applications for DR detection and analysis using OCT and OCTA images. METHODS: A comprehensive literature search was conducted across PubMed, Web of Science, Scopus, IEEE Xplore, and Embase for studies published up to March 2026. A total of 1007 articles were identified, of which 175 studies met the inclusion criteria following the PRISMA study selection process. RESULTS: DL-based approaches consistently demonstrated superior performance compared to traditional machine learning (ML) methods, with reported AUC values typically ranging from 0.90 to 0.99 across classification and segmentation tasks. Convolutional neural networks (CNNs), Vision Transformers (ViTs), and encoder-decoder architectures such as U-Net showed strong performance in detecting key DR biomarkers, including microaneurysms, macular edema, and neovascularization. However, performance variability was observed depending on dataset size, imaging modality, and annotation quality. CONCLUSIONS: AI-driven analysis of OCT and OCTA images offers significant potential for automated DR detection. Despite promising results, challenges such as limited public datasets, lack of cross-institutional validation, and model interpretability remain. Future research should focus on multimodal integration, explainable AI, and large-scale validation to enhance clinical applicability.
2 July 2026
Read appraisal →Medical image analysis
A review of deep learning-based Unsupervised Anomaly Detection in brain MRI
The manual assessment of brain Magnetic Resonance Imaging (MRI) scans can be labor-intensive and time-consuming for radiologists. Deep Learning methods have demonstrated the potential to aid this process. However, their effectiveness relies on the availability of large, annotated data sets. Unsupervised Anomaly Detection (UAD) presents a promising alternative, offering the potential to identify and localize anomalies without per-pixel annotations. Instead, a normative distribution is learned using healthy data, enabling the identification of abnormalities as deviations. This allows UAD methods to detect abnormalities that were unseen during training. This appealing feature has led to numerous studies proposing innovations and novel approaches. In this work, we provide a review of the literature and systematically collect and compare the proposed approaches. We observe that UAD has made significant advancements in brain MRI analysis. However, individual approaches are often evaluated in different contexts, i.e., changes in acquisition parameters, pre- and post-processing, and anomaly scoring. This variability makes it challenging to assess which models perform best, underscoring the need for comprehensive comparative studies concerning the specific context of MRI scans. Our collection, featuring public data sets, research studies, and open implementations, is available at our GitHub repository https://github.com/FinnBehrendt/Unsupervised-Anomaly-Detection-in-Brain-MRI.
1 July 2026
Read appraisal →Progress in biomedical engineering (Bristol, England)
Advancement of deep learning models with whole slide image in diagnosis, subtyping and prognosis for glioma.
Gliomas are heterogeneous primary central nervous system tumors with diverse molecular subtypes and variable prognosis. A paradigm shift in histopathology is underway as samples are digitized into whole-slide images (WSIs). Deep learning (DL) shows great potential for glioma diagnosis, subtyping, grading, and prognosis. This review summarizes recent advances in WSI-based DL models, covering data preprocessing, model architectures, and translational performance. Model evolution has progressed from CNNs (e.g. ResNet, capturing local features via convolution) to Transformers (e.g. Vision transformer, modeling global dependencies via self-attention), hybrid architectures, and large language model. Typical pipelines include quality control, stain normalization, patch extraction, feature integration, and prediction. Public datasets such as The Cancer Genome Atlas and Clinical Proteomic Tumor Analysis Consortium serve as key resources. On various datasets (public and institutional), a glioma-specific model ResNet-50 achieved AUC 0.983 for subtype classification; a Vision Transformer model reached AUC 0.960 for molecular typing; a hybrid model, ROAM, attained AUC 0.990; and the pan-cancer hybrid model called CHIEF achieved AUC 0.9397 for diagnosis. Attention heatmaps and Shapley Additive exPlanations provide interpretability by linking model outputs to histologic regions. Novel multimodal fusion integrates genomic, proteomic, radiologic, and clinical data to enhance prediction and uncover biological relevance. Future research may focus on enhancing model generalizability and predictive accuracy via advanced architectures, developing lightweight models for resource-limited settings to facilitate clinical translation.
20 June 2026
Read appraisal →Journal of medical systems
MEL-IA: An Interoperable AI System for Multimodal Skin Lesion Classification in Hospital Settings
UNLABELLED: Skin cancer represents a growing healthcare challenge, motivating the development of AI sytems that can support lesion assessment while integrating efficiently into clinical workflows. This study presents MEL-IA (MobilE skin Lesion dIAgnosis), an interoperable AI system designed for automated skin lesion classification and full integration with hospital information systems. METHODS: The system combines dermatoscopic images and structured clinical metadata using an EfficientNet-B4 - based multimodal model trained on the ISIC 2019 BCN_20000 and MSK subsets. Internal robustness was assessed through stratified 5-fold cross-validation, and external generalization was evaluated on the independent HAM10000 dataset. Interoperability and operational performance were validated through deployment in a real hospital environment using HL7, DICOM, PACS, and HIS/RIS systems. RESULTS: We evaluated internal robustness using stratified five‑fold cross‑validation, with global accuracy of 0.86, macro F1-score of 0.85, and AUC values above 0.97. The external evaluation showed that the model generalizes well under different acquisition conditions, achieving an accuracy of 0.84 and balanced accuracy of 0.65, with sensitivities of 0.60 for melanoma 0.72 for basal cell carcinoma. The multimodal model outperformed the image model baseline across all classes. Technical deployment confirmed full interoperability, > 99% successful study integration, and near real time processing. CONCLUSIONS: MEL-IA provides multimodal skin lesion classification capabilities and integrates smoothly into hospital infrastructures based on standards. The results demonstrate technical feasibility, interoperability, and operational viability for deployment within dermatology oriented clinical workflows. Further clinical validation studies are warranted to assess the system's impact on diagnostic decision-making and patient outcomes.
14 June 2026
Read appraisal →Progress in orthodontics
Evaluation of artificial intelligence-based cephalometric analysis compared to the manual method.
BACKGROUND: The application of artificial intelligence (AI) to cephalometry has aroused growing interest due to its potential to optimize clinical workflows and improve efficiency. This study aimed to compare the differences and reproducibility of cephalometric measurements obtained using three methods: manual, WebCeph™ and AngelAligner™. METHODOLOGY: 30 patients were analysed, evaluating 9 measurements generated by the three methods at two different timepoints (T1 and T2). RESULTS: WebCeph™ presented significant differences as compared to the manual method in 7 of the 9 measurements (SNA, SNB, ANB, U1-APg, L1-APg, GoGn-SN, U1-L1), while AngelAligner™ displayed significant differences with the manual method in 4 measurements (SNA, U1-APg, L1-APg and U1-L1). Significant differences were found between the two AI-based systems in 7 of the 9 measurements. Reproducibility was high at all angles with the manual method, except for U1-L1, NLA, and LFH. The ICC was 1 for all measurements for both AI-based systems, indicating optimal reproducibility. When grouping by malocclusions, however, WebCeph™ presented greater differences with the manual method than AngelAligner™. CONCLUSION: AI-based methods offered greater reproducibility than the manual method, but still require human supervision.
31 May 2026
Read appraisal →European radiology experimental
Towards an accessible, centralised, searchable database for AI courses in Europe: the Artificial Intelligence in Medical Imaging and Radiation Oncology Education (AIMIROE) project.
OBJECTIVE: Artificial intelligence (AI) is transforming medical imaging and radiation oncology, yet limited understanding and access to education hinder adoption. This study, led by the European Society of Medical Imaging Informatics (EuSoMII) in collaboration with the European Federation of Radiographer Societies (EFRS), aimed to create an accessible, centralised, searchable database including all AI courses in Europe. MATERIALS AND METHODS: An electronic survey was developed to collect data on European AI course characteristics, such as format, delivery, content, target audience and European Qualifications Framework (EQF) level. This was disseminated via purposive sampling through social media and mailing lists of the EuSoMII and the EFRS between September 2024 and January 2025. Quantitative data were analysed using descriptive statistics and visual representations using Python Seaborn and Geopandas. RESULTS: This study identified 29 AI courses in Europe. Of them, 53.6% were offered by universities. Courses targeted radiographers (59%), medical physicists (52%), and radiologists (41%), mainly at EQF level 7 (44.4%). Most courses were standalone (65.6%) and online (55.1%), while 41.3% were free of charge. English was the primary language of delivery (79%). CONCLUSIONS: Different AI courses across Europe offer some entry-level knowledge but are often short in duration. Expanding formats, building practical competencies, providing multilingual access, and European-wide reach are essential for meaningful, practical, and equitable AI integration. RELEVANCE STATEMENT: With the scaling-up of AI adoption in medical imaging and radiation oncology, there is a variety of AI education provisions currently available. Accessing these options via an open, centralised, regularly updated database enables people to make an informed decision about their training and practise safely and meaningfully. KEY POINTS: We identified 29 different AI European courses varying in language, content, and delivery. Many clinical practitioners and researchers are unaware of these resources. We need a centralised database for customising AI learning choices and guiding future course design.
31 May 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 →Scientific reports
Automated RECIST tumor response classification through prompt-guided large language models
This study investigates whether an entirely offline, general-purpose large language model (LLM) can reliably automate the classification of routine radiology reports according to RECIST (Response Evaluation Criteria in Solid Tumors) guidelines, focusing on how different prompting strategies support accurate, privacy-preserving tumor response assessment without additional model fine-tuning. An offline, in-house implementation of LLaMA-3.3 (70B) was used to classify real-world CT imaging reports from oncology patients. Reports were authored following RECIST-structured reporting but had outcome labels programmatically withheld prior to processing. Three prompting strategies-zero-shot, few-shot, and chain-of-thought prompting-were tested to guide the model in assigning RECIST categories: Baseline (BL), Complete Response (CR), Partial Response (PR), Stable Disease (SD), and Progressive Disease (PD). Model outputs were benchmarked against original expert labels using accuracy, precision, recall, and F1 scores. Across all tested prompting strategies, the LLaMA-3.3 model achieved strong classification performance. The best results were obtained with chain-of-thought prompting, reaching micro F1 scores of 0.81 across all RECIST categories. Overall model predictions aligned well with human expert assessments. Operating entirely offline within hospital infrastructure, the system preserved full compliance with stringent data privacy requirements. Prompt-driven large language models can accurately classify tumor response categories from real-world radiology reports in a scalable, reproducible, and privacy-preserving manner. Offline LLM deployment, combined with optimized prompting strategies, offers a promising approach for automating structured oncology report interpretation, potentially enhancing consistency and efficiency in clinical decision support workflows.
29 May 2026
Read appraisal →Journal of medical Internet research
Explainable AI in Cancer Imaging: Scoping Review of Methods, Modalities, and Clinical Integration
BACKGROUND: Explainable artificial intelligence (xAI) is increasingly used in medical imaging to enhance transparency, clinical interpretability, and trust in artificial intelligence (AI)-assisted diagnostics, particularly in oncology. Evidence on how explainability is implemented, validated, and reported in cancer imaging remains fragmented. OBJECTIVE: This scoping review aimed to systematically map research applying xAI methods to radiologic cancer imaging, summarize methodological and clinical trends, and identify persistent gaps in validation and integration. METHODS: We conducted a structured search of PubMed and Scopus (final search executed on October 20, 2025), covering studies published from 2017 to December 2024. Eligible peer-reviewed articles using machine learning or deep learning were analyzed with a focus on xAI components. Data from 371 studies were extracted into predefined categories covering cancer type, imaging modality, AI model, xAI method, terminology, validation, code availability, and decision support system integration. RESULTS: Studies focused primarily on breast (112/371, 30.2%), lung (87/371, 23.5%), and brain (56/371, 15.1%) cancers, with prostate, thyroid, and liver cancers also represented. The primary imaging modalities were computed tomography (139/371, 37.5%) and magnetic resonance imaging (104/371, 28%). Deep learning was used in 70.1% (260/371) of studies, classical machine learning in 18.1% (67/371), hybrid pipeline methods for 10% (37/371), and emerging concept-, prototype-, or causal-based approaches accounted for 1.9% (7/371) of studies. Post hoc xAI methods were dominant (305/371, 82.2%), with visualization (163/371, 53.4%), and feature relevance (111/371, 36.4%) as the most common subcategories. Hybrid post hoc or inherent approaches comprised 12.1% (45/371) and intrinsically interpretable methods 5.7% (21/371). Data sources were mostly public (149/371, 40.2%) or mixed (100/371, 26.9%); 22.9% (85/371) used private institutional datasets, and 7.8% (29/371) did not report data sources. Among validated studies, expert or user-based validation was most common (104/193, 53.9%), followed by mixed methods (74/193, 38.3%), while quantitative metrics (10/193, 5.2%) and clinical knowledge-based (8/193, 4.1%) validation remained rare. Only 17.5% (65/371) of studies provided code and 12.1% (45/371) reported decision support system integration, with few achieving actual clinical deployment. CONCLUSIONS: This scoping review maps xAI implementation across multiple cancer imaging modalities, revealing methodological inconsistency and insufficient validation. Most research emphasizes visualization over quantitative interpretability, and few models are reproducible or clinically implemented. These findings provide an evidence base for researchers, clinicians, and regulators to prioritize standardization of xAI reporting, quantitative validation, and user-centered frameworks to advance trustworthy AI in oncology imaging.
22 May 2026
Read appraisal →Dental press journal of orthodontics
Accuracy of cephalometric landmark identification on artificial intelligence-based software: a comparative study
INTRODUCTION: Cephalometry has long been a cornerstone of orthodontic diagnosis and treatment planning. While traditionally performed manually, cephalometric analysis has increasingly shifted toward semi-automated or fully automated computer-based software, due to advances in technology. OBJECTIVE: This study aims to evaluate and compare the accuracy of anatomical landmarks identified manually by orthodontists using digital tools versus those identified automatically by various artificial intelligence (AI)-based cephalometric software. METHODS: A total of 25 lateral cephalometric radiographs were selected for analysis. Manual landmark identification was independently performed by two orthodontists using the digital cephalometric software NemoCeph. These were compared to the automatic landmark identifications generated by two AI-based programs: CephNinja and WebCeph™. A total of 17 anatomical landmarks, 15 hard tissue and 2 soft tissue points, were evaluated. The 17 cephalometric landmarks selected in this study are among the most commonly used and reliably identifiable reference points in the literature. To facilitate direct visual comparison, landmark positions derived from all three methods were overlaid using semi-transparent image layers. One-way ANOVA and intraclass correlation coefficient (ICC) tests were conducted to assess inter-method differences. The threshold for statistical significance was set at p < 0.05 and p < 0.01. RESULTS: Statistically significant differences were observed at the porion, basion, gonion, and the apexes of the upper and lower incisors (p < 0.05). Compared to the manual tracings in NemoCeph, AI-based software generally placed these landmarks in more distal and superior positions. Among the AI tools, CephNinja showed greater similarity to NemoCeph for the gonion and basion points, while WebCeph™ was closer at the incisor apexes. CONCLUSION: Although AI-based cephalometric software offers rapid and reproducible assessments, certain anatomical landmarks - particularly porion, basion, gonion, and the apexes of the upper and lower incisors - demonstrate statistically significant positional discrepancies, compared to manual tracings. Clinicians should be cautious when relying exclusively on automated analyses for diagnostic or treatment planning purposes, since AI struggles with certain landmarks, due to algorithmic limitations.
21 May 2026
Read appraisal →The British journal of dermatology
Artificial intelligence-based smartphone application for skin cancer detection: a prospective diagnostic accuracy study
BACKGROUND: Smartphone applications ('apps') that use artificial intelligence (AI) for skin cancer diagnosis may help with early detection. However, prospective studies assessing real-world performance are scarce. OBJECTIVES: To evaluate in an independent prospective study the diagnostic accuracy of a widely used skin cancer detection app and the effect of photographic conditions in a user-representative cohort of patients consulting for lesions of concern. METHODS: Between 1 February 2021 and 30 June 2023, participants presenting at an early-access consultation with a lesion of concern were consecutively enrolled. Following dermatological assessment, each lesion was photographed using the app. The app's convolutional neural network (CNN) generates a binary risk output and advice. A teledermatology review was performed in a subset of cases. The diagnostic accuracy of the CNN combined with teledermatology review was compared with final clinical or histopathological diagnosis. Performance was tested under different photographic conditions (angle, lighting, user) and with different smartphone models. The trial was registered with ClinicalTrials.gov (NCT05246163). RESULTS: A total of 1458 participants with 1904 lesions of concern were included. Of these, 185 (9.7%) were skin cancers, with 32 melanomas. Image capture was unsuccessful in 16.6% of lesions (n = 317/1904), despite optimal conditions. For successfully captured lesions, the CNN achieved a sensitivity of 82.5% and specificity of 76.8% for skin cancer detection. In-app teledermatology review was available for 65.7% (n = 1042/1587) of images. Combined CNN and teledermatology review resulted in increased specificity (86.8%; P < 0.001), with a sensitivity of 75.3%. Overall sensitivity in melanocytic lesions was lower than in nonmelanocytic lesions (71.0% vs. 76.3%; P < 0.001), whereas specificity was higher (90.0% vs. 84.9%; P < 0.001). A substudy demonstrated low image-capture success in user hands (28.9%) and variation in diagnostic performance between smartphone models. CONCLUSIONS: This independent, prospective study evaluated diagnostic performance of a widely used skin cancer detection app in a user-representative cohort. The findings highlight the importance of independent clinical validation of AI-based healthcare tools in real-world settings. Skin cancer is one of the most common cancers that affects people worldwide. Finding cancer early is important, especially for melanoma, a serious form of skin cancer. There is potential for smartphone apps to use AI to help people identify skin spots that may need medical attention. In this study, we evaluated a widely used app called SkinVision. People presenting with a worrisome spot at a dermatology appointment were invited to take part in this study. Altogether, 1458 people were included. The skin area of concern was assessed with the app. Using AI, the app produced a risk score. It also gave a final score. The final score was the AI risk score that may be combined with dermatology advice delivered remotely (‘teledermatology’). We compared the app scores with the doctor’s final diagnosis. We also studied the impact of things such as photo angle, lighting, type of smartphone and the person who took the photo. We found that over 16% of worrisome skin areas could not be captured by the app under ideal conditions. If participants were asked to take the photo themselves, this increased to over 70%. The app correctly identified skin cancers in 83% of cases but gave false alarms in 23%. The in-app teledermatology review reduced the number of false alarms, but this increased the number of missed skin cancers. Finally, we found important differences between different smartphones. Our findings highlight the importance of trialling apps like this with real patients in real-life setting before they are fully introduced to the public.
20 May 2026
Read appraisal →Biomedical physics & engineering express
Oral squamous cell carcinoma diagnosis with time and frequency domain features from optical coherence tomography A-scan signals
Optical coherence tomography (OCT) is extremely useful in the screening and detection of oral cancers. But various challenges, such as its subjectivity, operator dependence in interpretation, and lack of quantitative outcomes, have been delaying its adoption in clinical decision-making. The objective of this research is to quantify various A-scan features embedded in the OCT signal through advanced signal processing techniques and machine learning algorithms.Ex vivoimaging of biopsied oral tissues (normal mucosa, carcinomain situ(CIS), well-differentiated, and poorly differentiated oral squamous cell carcinoma) was performed using a spectral domain OCT system. Our A-scan dataset consisted of representative 1D signals obtained from different regions of the tissue bed. A set of 12 time- and 8 frequency-domain features were computed on each A-scan, and ten machine learning models were evaluated for binary and multi-label classification. LightGBM achieved the highest performance in both binary (accuracy: 0.8847,F1: 0.8878, AUC: 0.9539) and multi-label classification (accuracy: 0.8248,F1: 0.8201, AUC: 0.964). LightGBM was selected as the final model based on its superior and consistent performance across both classification paradigms. Our proof-of-concept feasibility study demonstrates good accuracy for differentiating oral mucosal tissues, highlighting the biomarker signature of A-scans.
19 May 2026
Read appraisal →Computers in biology and medicine
AI unleashed: A meta-analysis transforming radiological insights in diagnosing abdominal infections
INTRODUCTION: Artificial intelligence (AI)-based imaging modalities are next-generation diagnostic devices for abdominal infections that promise to provide enhanced diagnostic speed and accuracy. This systematic review and meta-analysis critically analyze the diagnostic accuracy of AI-assisted imaging modalities, including for appendicitis, pneumoperitoneum, and cholecystitis, to present a balanced estimate of their clinical utility. METHODS: A systematic literature search was undertaken in PubMed, Scopus, and Cochrane databases to search for studies that assessed the diagnostic accuracy of AI-based imaging modalities for abdominal infections. Eleven studies were included based on the inclusion criteria, and data were pooled for analysis. Diagnostic performance was measured by estimating sensitivity, specificity, likelihood ratios, diagnostic odds ratio (DOR), and area under the curve (AUC) using a random-effects model. Subgroup analyses were done to investigate the effect of infection type on diagnostic accuracy. RESULTS: AI-assisted imaging demonstrated an overall sensitivity of 0.891 (95% CI: 0.824-0.944) and specificity of 0.860 (95% CI: 0.784-0.922) for diagnosing abdominal infections. Subgroup analysis revealed that AI-aided computed tomography (CT) exhibited a sensitivity of 0.902 (95% CI: 0.850-0.948), while ultrasound (US) showed a sensitivity of 0.864 (95% CI: 0.792-0.922). The highest AUCs were observed for pneumoperitoneum (0.985) and appendicitis (0.947), underscoring AI's robust diagnostic capabilities across multiple pathologies. CONCLUSION: AI-imaging modalities significantly enhance diagnostic accuracy for abdominal infections, particularly for appendicitis and pneumoperitoneum. This meta-analysis underscores AI's clinical potential, though future research should prioritize multicenter studies to validate AI models' generalizability and ensure consistent performance across diverse healthcare settings.
17 May 2026
Read appraisal →JMIR medical informatics
Multimodal Fusion of Echocardiogram Images and Electronic Medical Records for Heart Disease Screening: Retrospective Algorithm Development and Validation Study
BACKGROUND: Echocardiography is a fundamental imaging modality for the diagnosis of heart disease (HD), but its interpretation remains operator-dependent and lacks standardized, data-driven decision support. Although artificial intelligence has improved image-based diagnosis, the added value and interpretability of integrating routinely collected electronic medical records (EMRs) with echocardiogram (ECHO) for large-scale screening remain underexplored. OBJECTIVE: This study aims to develop an explainable artificial intelligence framework that integrates multimodal data, including ECHO images and EMRs, with derived temporal and clinical features, to enhance the screening and interpretability of heart disease diagnosis. METHODS: In this retrospective single-center study, we analyzed 26,936 patients with linked ECHO and EMR records (1470 with clinically confirmed HD and 25,466 without). We constructed cross-sectional EMR features (demographics, comorbidities, and registration history) and extracted view-specific image features with a simplified Inception-v3 backbone. Five modality-specific feature vectors (4 imaging views + EMR) were concatenated and input to an extreme gradient boosting classifier within a patient-level stratified 5-fold cross-validation framework. To mitigate class imbalance during training, non-HD cases in the training folds were randomly downsampled to a 1:3 HD:non-HD ratio; held-out test folds preserved the original prevalence. Post hoc explainability was provided via Grad-CAM heatmaps for images and Shapley additive explanations analysis for EMR features. RESULTS: For the primary binary task (presence vs absence of HD), the multimodal model achieved an AUC of 0.8147 (SD 0.009) on held-out test folds, compared with 0.7785 (SD 0.014) for the image-only baseline, and 0.7343 (SD 0.009) for the EMR-only baseline. At the clinically selected decision threshold, sensitivity, specificity, positive predictive value, and negative predictive value were 84.6% (SD 1.6%), 72.4% (SD 2%), 15% (SD 2.3%), 98.8% (SD 0.3%) for the multimodal model, versus 82.3% (SD 1.9%), 73.5% (SD 2%), 15.2% (SD 2.1%), 98.6% (SD 0.3%) (image-only), and 77.5% (SD 2.1%), 67% (SD 2.3%), 11.9% (SD 1.7%), 98.1% (SD 0.4%) (EMR-only). Grad-CAM visualizations qualitatively highlighted anatomically and physiologically meaningful regions (eg, valvular structures and abnormal flow jets), while Shapley additive explanations analysis identified age, registration years, sex, and hypertension among the top EMR contributors-findings that align with established cardiovascular risk factors. CONCLUSIONS: Integrating echocardiographic images and EMR data in an explainable multimodal framework yields improved and clinically plausible screening performance for heart disease. Future work should focus on external multicenter validation, quantitative assessment of visual explanations against expert annotations, and prospective studies assessing clinical use and workflow integration.
17 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 →Clinical oral investigations
Classification of occlusal caries severity using spectrophotometric CIELAB measurements and machine learning algorithms
OBJECTIVES: This study evaluated whether spectrophotometric CIELAB color measurements can support machine-learning (ML) classification of occlusal caries severity based on ICDAS categories. METHODS: Three hundred extracted human teeth were visually classified using ICDAS (0-4) and served as labels for supervised learning. Five occlusal sites per tooth were measured using a spectrophotometer under standardized conditions. Site-level CIELAB values were transformed into engineered color features and aggregated at the tooth level. Teeth were categorised as sound (ICDAS 0), initial carious lesions (1-2), and moderate carious lesions (3-4). Five ML models: Random Forest (RF), XGBoost, CatBoost, Multilayer Perceptron (MLP), and a Deep Sets architecture were trained. Performance was evaluated using accuracy, balanced accuracy (BA), and macro-F1, with additional binary analyses for early (ICDAS 1-4 vs. 0) and operative lesion detection (ICDAS 3-4 vs. 0-2), based on the included ICDAS range. Learning curve analysis was performed to evaluate the effect of training data size on model performance. RESULTS: Deep Sets achieved the highest multiclass performance (BA = 0.89) followed by the MLP (BA = 0.72). Tree-based models demonstrated lower performance overall. For early lesion detection, all models showed high sensitivity (SE) but reduced specificity (SP). At the operative level, the MLP achieved 100% SE and moderate SP. Learning-curve analysis showed that neural models benefited most from increased training data, whereas tree-based models showed limited improvement. CONCLUSION: Under controlled in-vitro conditions, spectrophotometric CIELAB measurements enabled machine-learning classification of occlusal caries severity. Deep sets achieved the highest overall performance, supporting the potential of color-based approaches as adjunctive tools for caries assessment. CLINICAL RELEVANCE: Spectrophotometric tooth color measurements may support machine learning classification of occlusal caries severity under controlled conditions, using an image-independent approach. Such color-based methods may complement visual assessment by providing a more standardized and reproducible evaluation of enamel changes following further validation.
16 May 2026
Read appraisal →Computers in biology and medicine
Artificial intelligence for metabolic dysfunction-associated steatotic liver disease diagnosis: A systematic review
Metabolic dysfunction-associated steatotic liver disease (MASLD) poses significant diagnostic challenges. We conducted a systematic review of artificial intelligence (AI) applications in MASLD diagnosis, focusing on predictive modeling, screening, imaging analysis, and therapeutic assistance. Following PRISMA guidelines, we searched five databases (Web of Science, SciFinder, Elsevier ScienceDirect, EBSCOhost, Springer) for studies published January 2020-January 2026. Of 1247 initially identified records, 191 studies met inclusion criteria. AI models demonstrate high diagnostic performance across multiple modalities. Deep learning analysis of US, CT, and MRI enables automated steatosis quantification and fibrosis staging with accuracy comparable to expert interpretation. AI integration with multi-omics data facilitates novel biomarker discovery and therapeutic target identification. Despite these advances, model generalisability, standardization, and clinical integration remain challenges. AI holds substantial potential to transform MASLD diagnostics, but prospective validation and implementation strategies are needed before widespread clinical adoption.
16 May 2026
Read appraisal →Nature communications
A deep learning ECG model for identification and localization of occlusion myocardial infarction.
Rapid identification and localization of an acute coronary occlusion are vital to prevent myocardial damage, yet reliance on ST-segment ECG criteria misses many acute occlusion myocardial infarctions (OMI) and triggers unnecessary acute angiographies. Here, we present a trained and validated deep learning model using 540,372 emergency ECGs paired with definitive catheterization outcomes. The model has a C-statistic of ≥0.95 for OMI and ≥0.87 for non-OMI infarctions and can localize culprit lesions in the three main coronary branches, which can guide the angiographer. Performance is similar across age, sex, and ECG hardware subgroups. Obviating dependence on ST-elevations and troponins, this model for the identification and localization of OMI has the potential to shorten the time to reperfusion of an acute coronary occlusion and save resources. Because human oversight of OMI detection on the ECG is limited, randomized clinical trials with patient-relevant outcomes are warranted.
15 May 2026
Read appraisal →JMIR medical informatics
Adverse Pregnancy Outcomes in Women With Immune Abnormalities: Machine Learning Model Development and Validation Using First-Trimester Sonographic Features
BACKGROUND: The maintenance and progression of pregnancy rely on immune homeostasis at the maternal-fetal interface. However, pregnancy complicated by autoimmune abnormalities can disrupt this balance and significantly increase the risk of adverse pregnancy outcomes (APOs). OBJECTIVE: This study aimed to (1) develop an interpretable predictive tool for APOs in patients with immune abnormalities and (2) interpret the models using Shapley additive explanations (SHAP) values. METHODS: This study retrospectively analyzed clinical data from 288 patients with autoimmune abnormalities at Yichang Central People's Hospital between 2019 and 2024. Feature selection was performed using both the Boruta algorithm and Least Absolute Shrinkage and Selection Operator regression to identify optimal predictive factors associated with APOs. Nine machine learning models were developed and subsequently underwent a comprehensive comparative evaluation of their predictive performance, leading to the identification of the optimal predictive model. SHAP values were generated to provide interpretable insights into model predictions. RESULTS: A total of 288 patients were included in the study, 124 (43.06%) of whom had APOs. The extreme gradient boosting algorithm was shown to be the optimal model after a comparison of 9 different models utilizing various metrics. The SHAP analysis showed that crown-rump length at 6+0 to 8+6 weeks, the number of other drugs, the number of complications during pregnancy, gestational sac volume at 6+0 to 8+6 weeks, and yolk sac diameter change at 6+0 to 8+6 weeks were the key predictive factors affecting APOs. CONCLUSIONS: The study developed an interpretable predictive tool for APOs in patients with immune abnormalities, which may assist clinicians in making early intervention decisions.
14 May 2026
Read appraisal →Physiological measurement
A teacher-student deep learning framework for enhanced clinical screening of heart failure from 12-lead electrocardiograms
Objective.Identifying heart failure (HF) from electrocardiograms (ECGs) is challenging due to the lack of definitive features. This study aims to develop a deep learning-based clinical decision support system, CTTSnet, for accurate and automated HF screening using ECGs.Approach.We propose a novel teacher-student framework, CTTSnet, which synergistically integrates a transformer as the teacher and a convolutional neural network as the student. This architecture combines global context-awareness with efficient local feature extraction. The model was trained and evaluated on a large-scale, real-world dataset of 27 018 ECGs and further validated on two external public datasets to assess generalizability.Main results.CTTSnet achieved an AUROC of 0.941 on the primary clinical dataset, surpassing strong baselines. The model attained an accuracy of 87.3%, with a recall of 88.6% and specificity of 85.8%. External validation on two additional public datasets further confirmed the model's generalizability.Significance.CTTSnet provides a scalable, high-accuracy tool for HF triage from routine ECG recordings. Its potential clinical value lies in assisting early identification of patients at risk of HF, thereby helping reduce missed diagnoses and facilitating timely referral for further evaluation.
14 May 2026
Read appraisal →PloS one
Using deep learning to identify inherited retinal diseases based on wide-field retinal imaging data
OBJECTIVE: To evaluate a novel image-based deep learning method for the automated identification of inherited retinal diseases (IRDs) and to explore the feasibility of predicting selected causative gene groups using a multimodal analysis of wide-field fundus autofluorescence (FAF) and pseudocolor fundus (pCF) images. DESIGN: The method was evaluated using a retrospective dataset of patient studies containing FAF and pCF images, as well as genetic tests for IRD. PARTICIPANTS: Patients with confirmed IRD for which both wide-field FAF and pCF images and genetic tests for IRD performed at Hadassah University Medical Center were included. The dataset consisted of 409 patients (330 patients with IRD with the 25 most commonly affected genes in our population and patients without IRD, and 79 patients without IRD). METHODS: Nine EfficientNet-V2-m convolutional neural networks were trained for the following three classification tasks: a binary IRD vs. non-IRD classification, and classification into two groups of five causative genes (Groups 1 and 2). For each task, three models were trained on the FAF images only, the pCF images only, and both the FAF and pCF images. The performance of the models was then evaluated and compared using 5-fold cross-validation. MAIN OUTCOME MEASURES: Accuracy, precision, F1 scores, AUC, and confusion matrices. RESULTS: The multimodal classification models that were trained on both the FAF and pCF images yielded the best results. The binary classification model had a mean (±SD) accuracy of 0.95 ± 0.01, a mean precision of 0.92 ± 0.01, and a mean F1 score of 0.90 ± 0.02. The Group 1 classification model had a mean accuracy of 0.92 ± 0.03, a mean precision of 0.93 ± 0.03, and a mean F1 score of 0.89 ± 0.03. Finally, the Group 2 classification model had a mean accuracy of 0.85 ± 0.03, a mean precision of 0.87 ± 0.04, and a mean F1 score of 0.83 ± 0.04. CONCLUSIONS: Our results indicate that determining whether a patient has IRD can be performed with high accuracy within this retrospective cohort based on FAF and pCF images using image-based deep learning classifiers. This image-based approach may assist clinicians during the patient's initial visit by providing decision support prior to genetic testing. It may also help prioritize patients for genetic workup, particularly in settings in which genetic testing is not readily available. Further prospective and external validation is required before clinical implementation.
13 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 →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 →Journal of medical systems
Interpretable Machine Learning to Anticipate the Diagnostic Yield of EEG in the Emergency department. The EMINENCE study.
INTRODUCTION: Emergent electroencephalography (emEEG) is increasingly employed in the emergency department (ED) for evaluating altered consciousness and seizure-related conditions, yet standardized criteria guiding its use remain limited. METHODS: We retrospectively analyzed 1,018 patients (mean age 66 ± 20 years; 48.4% female) undergoing emEEG at the ED of the Careggi Teaching Hospital (Florence, Italy) in 2023. Clinical, anamnestic, and neuroimaging data available at admission were used to train supervised machine-learning (ML) models. We evaluated tree-based ensembles (Random Forest and XGBoost) to predict abnormal and epileptiform emEEG, as well as confirmation or refutation of initial diagnosis. Ground-truth labels were derived from a multidisciplinary expert team including neurologists, neurophysiopathologists and intensivists. Model performance was assessed with 5 × 5 nested cross-validation, receiver operating characteristic (ROC) analysis, balanced accuracy, decision-curve analysis, and Shapley Additive Explanations (SHAP) interpretability. RESULTS: Abnormal emEEG occurred in 691 cases (67.9%), epileptiform activity in 192 patients (18.9%). emEEG ruled out the initial diagnostic suspicion in 514 cases (50.5%) and confirmed it in 188 cases (18.5%). Best performance was obtained with Random Forest for abnormal emEEG (AUC 0.79, 95% CI: 0.76-0.82) and diagnosis rule-out (0.84, 0.81-0.86), and with XGBoost for epileptiform emEEG (0.82, 0.78-0.85) and diagnosis confirmation (0.82, 0.79-0.85). Performance varied by initial diagnostic suspicion, but subgroup-stratified analyses showed overall consistent patterns. Key predictive features included altered consciousness, prior brain lesions, antiseizure therapy, and seizure-related presentations. Interpretability analyses revealed seizure-centric features drove confirmation, while systemic or nonspecific features favored refutation. CONCLUSIONS: Interpretable ML models using only admission data can predict emEEG outcomes and anticipate their diagnostic contribution, supporting triage and decision-making in emergency neurology without replacing clinical judgment. Models and explanations were easily usable on a freely-accessible website ( www.emergencyeeg.com ), where tools return probabilistic outputs for all four prediction tasks together with per-patient explanation plots, enabling transparent and reproducible use. CLINICAL TRIAL NUMBER: Not applicable.
9 May 2026
Read appraisal →Journal of biomedical informatics
Model utility and explainability in federated learning - A case study in healthcare using fundus oculi datasets
OBJECTIVE: Introduce a case study for Federated Learning (FL) in healthcare, addressing challenges posed by patient privacy and limited large-scale datasets. Our goal is to assess the features learned by FL methods in a simulated, diverse setting that emphasizes realistic data heterogeneity, and to analyze the learned representations for their medical relevance using both local and global explainability techniques. METHODS: Six fundus oculi datasets were combined to simulate a diverse federated learning environment, representing heterogeneous data conditions. We evaluated three established FL methods against centrally trained models, assessing both predictive performance and the learned representations. Specifically, explainability techniques were employed to examine the features learned by the models, and local explanations were evaluated against attention maps annotated by ophthalmologists. Robustness against common biases in fundus datasets was also assessed. RESULTS: Our study found improvements in model utility (up to 9.97%) with FL methods compared to isolated training. Analysis of learned representations revealed that federated models predominantly learn the vertical cup-to-disc ratio, a crucial feature for glaucoma diagnosis, and demonstrated robustness against common biases. High agreement was observed between local explanations and ophthalmologist-annotated attention maps. CONCLUSION: This study demonstrates the benefits of FL systems in a healthcare scenario, providing a case study for evaluating federated systems beyond idealized benchmarks. Our findings highlight the potential of FL to not only improve model utility in privacy-sensitive medical domains but also to learn medically relevant features instead of spurious correlations.
9 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 →Open heart
Relative apical sparing of left ventricular longitudinal strain for the diagnosis of cardiac amyloidosis: a systematic review and meta-analysis
BACKGROUND: Cardiac amyloidosis (CA) is underdiagnosed due to non-specific clinical and echocardiographic features. This systematic review and meta-analysis evaluated the diagnostic accuracy of the left ventricular relative apical sparing (RELAPS) pattern on speckle-tracking echocardiography. METHODS: A literature search of PubMed, Scopus and Cochrane was conducted through August 2025. RELAPS was defined as the average apical longitudinal strain divided by the sum of average basal and average middle left ventricular longitudinal strains. Diagnostic accuracy was assessed using a bivariate random-effects model. Subgroup analyses by CA subtype and echocardiogram software, and meta-regression for clinical variables, were performed. Optimal cut-offs were determined by maximising the Youden index. Analyses were conducted in R V.4.4.1. RESULTS: Forty-one studies (3473 CA; 4525 comparators) were included. Pooled analysis yielded an area under the receiver operating characteristic curve (AUC-ROC) of 0.818, with sensitivity 65.9% (95% CIs 59.2% to 72.0%) and specificity 83.1% (CI 78.5% to 86.8%). The diagnostic oods ratio (DOR) was 9.54 (CI 7.41 to 12.10). In subtype analyses, immunoglobulin light chains (AL-CA) showed AUC-ROC 0.876, sensitivity 59.8% (CI 41.1% to 76.0%) and specificity 92.7% (CI 83.6% to 96.9%); transthyretin (ATTR-CA) showed AUC-ROC 0.822, sensitivity 63.8% (CI 51.2% to 74.8%) and specificity 84.4% (CI 76.7% to 89.9%), with no significant differences between subtypes (sensitivity p=0.760; specificity p=0.172). Three echocardiography software systems were evaluated. GE EchoPAC (26 studies) achieved an AUC-ROC of 0.822 (sensitivity 70.4% (CI 64.1% to 76.0%), specificity 81.2% (CI 75.7% to 85.7%)). Philips QLAB (four studies) performed comparably (AUC-ROC 0.904; sensitivity 67.6% (CI 31.2% to 90.6%), specificity 92.7% (CI 73.6% to 98.3%)). In contrast, TomTec (four studies) had an AUC-ROC of 0.806, but its sensitivity (31.1% (CI 9.4% to 66.3%)) was significantly lower than GE EchoPAC's (p<0.001), despite a similar specificity (93.6% (CI 78.8% to 98.3%)). CONCLUSION: RELAPS provides moderate diagnostic accuracy for identifying CA, with good specificity and modest sensitivity. Performance varies by analysis software and threshold, underscoring the need for standardised measurement and prespecified cut-offs.
27 Apr 2026
Read appraisal →Physics in medicine and biology
MRDT-GAN: generative adversarial network with multi-scale residual dense transformer generator for low-dose CT denoising
Objective.Low-dose computed tomography (LDCT) reduces radiation exposure but introduces noise and artifacts that degrade diagnostic quality. Existing deep learning-based denoising methods still face challenges such as over-smoothing, loss of fine structures, and uneven contrast. This study aims to develop an LDCT denoising framework that enhances noise suppression while preserving anatomical details and structural fidelity.Approach.We propose a multi-scale residual dense transformer generative adversarial network (MRDT-GAN). In the generator, we adopt the MRDT Block as the core unit, which introduces multi-scale strategy into residual dense network to reduce over-smoothing and preserve fine details, and also patching transformer block to capture long-range dependencies, mitigating distortions caused by localized receptive fields in convolutional neural network-based approaches. A hybrid attention module is also introduced in the generator to process spatial, frequency, and contrast information, enabling the network to focus on critical regions for noise suppression, improve contrast uniformity, and maintain texture consistency. In the discriminator, we adversarially explore differences on global, pixel, and also sub-scale between denoised LDCT and normal dose CT to better capture structural variations, reduce local noise and distortions, and ensure more realistic texture reconstruction while minimizing artifacts.Main results.We validate MRDT-GAN on both the NIH-AAPM-Mayo Clinic LDCT dataset and a real-world dataset. Experimental results indicate that MRDT-GAN achieves superior denoising performance compared with existing methods, effectively preserves details, enhances visual quality, and achieves a better balance between noise suppression and structural integrity.Significance.MRDT-GAN provides an effective and generalizable LDCT denoising solution that balances noise reduction with fine-detail preservation. By integrating MRDT modeling, hybrid attention mechanisms, and multi-difference adversarial learning, the framework offers improved clinical applicability and supports high-quality image reconstruction for downstream diagnostic tasks.
27 Apr 2026
Read appraisal →La Medicina del lavoro
Artificial Intelligence in Occupational Health Surveillance: Evaluating AI-Assisted ILO Classification of Radiographs of Pneumoconioses
BACKGROUND: Pneumoconioses remain an important occupational health issue, particularly in low- and middle-income countries. The International Labour Organization (ILO) Classification standardizes chest radiograph interpretation but requires trained readers and is affected by inter-reader variability. This study evaluated whether generative multimodal artificial intelligence (AI) models can approximate ILO-based diagnostic reasoning. METHODS: Eighty-two chest radiographs from the official NIOSH B Reader syllabus were analysed using four AI systems (GPT-4o, GPT-5, MedGemma-4B, MedGemma-27B). Each image was evaluated with a standardized prompt based on the 2022 revised ILO guidelines using deterministic settings. Model outputs were mapped to ILO codes and compared with the official answer keys of the ILO Standard Radiograph Set used for B Reader training and examination. Performance metrics included balanced accuracy, sensitivity, specificity, precision, and Matthews correlation coefficient (MCC). Bootstrap 95% confidence intervals, McNemar's test, and Cohen's κ assessed performance variability and agreement. RESULTS: All four AI models showed moderate diagnostic performance, with balanced accuracy ranging from 60.8% to 70.3%. Sensitivity remained limited (35.5%-54.9%), while specificity was consistently high (84.6%-86.2%). MedGemma-27B performed best for small opacities, GPT-5 for pleural abnormalities and for technical quality. Large opacities and rare findings were systematically under-detected. Statistical comparisons showed significant differences between models, although agreement patterns were broadly similar. CONCLUSION: All AI models partially followed structured ILO radiographic criteria but did not achieve expert-level performance, confirming that they cannot replace certified B Readers. Larger, real-world datasets are needed to assess their potential clinical utility as supportive tools in occupational health surveillance programs.
25 Apr 2026
Read appraisal →The British journal of ophthalmology
Multimodal artificial intelligence for retinal detachment diagnosis using fundus imaging and patient questionnaires
BACKGROUND/AIMS: This study aimed to develop a multimodal artificial intelligence (AI) system that integrates fundus imaging and patient questionnaire data to achieve clinician-level diagnostic accuracy for diagnosing retinal detachment (RD). METHODS: Ultra-widefield fundus images and comprehensive patient questionnaires were collected from patients with RD and healthy controls at Tsukazaki Hospital. A multimodal model was developed using the Contrastive Language-Image Pretraining framework to classify RD cases, alongside separate image-only and questionnaire-only models for comparison. Per-image and per-subject analyses were conducted to assess model performance. RESULTS: The multimodal model outperformed single-modal models in per-image and per-subject assessments. It achieved accuracy, recall and F1 scores of 0.899±0.054, 0.902±0.043 and 0.902±0.048 in the per-image analysis and 0.893±0.071, 0.949±0.044 and 0.873±0.074 in the per-subject analysis, respectively. The AI model's overall diagnostic accuracy was slightly lower than that of human clinicians; however, it exhibited a higher recall rate, indicating improved detection of true RD cases. CONCLUSION: Integrating fundus imaging with patient questionnaire data significantly improves AI-based RD diagnosis. Future research needs to focus on expanding the dataset and refining the questionnaire design to further improve model performance.
25 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 →Archives of disease in childhood
How accurate is point-of-care ultrasound for detecting paediatric appendicitis? A systematic review and meta-analysis.
OBJECTIVES: Primary: To evaluate the diagnostic accuracy of point-of-care ultrasound (POCUS) in the diagnosis of appendicitis in paediatric patients presenting with suspected appendicitis.Secondary: To investigate how the use of POCUS affects length of stay in the paediatric emergency department and CT utilisation. METHODS: We searched MEDLINE, EMBASE and Cochrane Central Register of Controlled Trials (CENTRAL) up until February 2025 for studies involving patients presenting to the paediatric emergency department with suspected appendicitis, who underwent POCUS. Studies were assessed for risk of bias using the Quality Assessment Tool for Diagnostic Accuracy Studies framework. The primary outcome of interest was the ability of POCUS to detect appendicitis in terms of sensitivity and specificity. RESULTS: Eight studies were included encompassing 993 patients. Calculated pooled sensitivity was 85.6% (95% CI 68.9% to 94.1%) with a specificity of 90.2% (95% CI 86.5% to 93.0%). There was a wide range of reported sensitivities for POCUS, ranging from 53% to 100%. There was less variation in reported specificities that ranged from 82% to 95.2%. The studies analysed were of variable quality with the reference standard and flow and timing being the main areas subject to bias. CONCLUSIONS: Based on the findings of this systematic review and meta-analysis, it is reasonable to conclude that POCUS can be used as an effective tool to diagnose paediatric appendicitis. With a moderately high sensitivity, it may not be recommended to be used to exclude appendicitis based on POCUS findings alone. Further prospective studies evaluating the reliability of POCUS in excluding appendicitis, either alone or when combined with a clinical prediction rule, are required. Secondary findings from this study demonstrate that POCUS may result in a reduced length of stay in the paediatric emergency department and lead to a reduction in CT utilisation, though insufficient data were available to draw firm conclusions in relation to this.
19 Apr 2026
Read appraisal →Journal of medical Internet research
Error Detection in Emergency Radiology Reports Using a Large Language Model: Multistage Evaluation Study
BACKGROUND: Emergency radiology requires highly accurate reporting under time constraints; yet, increasing workloads raise the risk of errors. While large language models (LLMs) show potential for proofreading in general radiology, their performance in emergency settings and non-English contexts remains unclear. OBJECTIVE: We aim to evaluate the performance of a domain-optimized LLM, DeepSeek-R1, for identifying errors in Chinese emergency radiology reports, with comparison against assessments by board-certified radiologists. METHODS: We compiled 7435 emergency reports (dataset 1; radiography, computed tomography, and magnetic resonance imaging) collected from November 2024 to April 2025. In stage 1, a total of 5 LLMs were evaluated using 200 reports. The best model, DeepSeek-R1, proceeded to stages 2 and 3, where 0-shot and few-shot learning were tested on a separate set (n=100). Model performance was compared against 12 radiologists. Stage 4 validated real-world utility on 800 verified reports. RESULTS: In subdataset 1, under stress-testing conditions, DeepSeek-R1 achieved a higher error detection rate in the few-shot setting than in the 0-shot setting (84.4% vs 60.9%, P=.003). Its performance exceeded that of radiology residents (84.4% vs 51.6% and 53.1%, respectively; both P<.05) and showed no statistically significant difference compared with senior radiologists and attending radiologists (84.4% vs 68.8%-93.8%, P=.26 to ≥.99). Compared with residents, DeepSeek-R1 detected more critical omissions (100% vs 25% and 50%; both P<.05) and other errors (92% vs 33% and 33%; both P=.02). In dataset 2, collected from independent institutions, DeepSeek-R1 achieved a detection rate of 95% under the few-shot setting. Reading time was shorter than that of human readers (92 vs 109 s). In real-world validation, DeepSeek-R1 identified 117 true reporting errors, yielding a positive predictive value of 56.5%. CONCLUSIONS: DeepSeek-R1 holds promise for improving quality control in emergency radiology reports. Its performance and efficiency support its use as an assistive proofreading tool in real-world radiology workflows.
15 Apr 2026
Read appraisal →Biomedical physics & engineering express
Efficient machine learning models leveraging DCE-MRI morphological and dynamic features allow accurate breast lesion classification.
We propose an ensemble learning approach to classify malignant versus benign breast lesions by leveraging morphological and dynamic features derived from Magnetic Resonance Images (MRI). The analysis was performed on 164 breast lesions of the publicly available 'Advanced MRI Breast Lesions' dataset from The Cancer Imaging Archive, containing T2-weighted and Dynamic Contrast-Enhanced (DCE)-MRI sequences, along with the segmentation masks of suspicious lesions. After extracting radiomic features using the Pyradiomics Python package, we computed dynamic features from DCE-MRI kinetic curves, which describe the contrast agent wash-in and wash-out. These features have been defined as the derivatives of image intensity measures, such as mean and standard deviation, computed within the masks on the five DCE-MRI time steps. We trained and evaluated an eXtreme Gradient Boosting (XGBoost) classifier, experimenting with different combinations of features and segmentation masks in a stratified 5-fold cross-validation. The best model trained on T2-weighted MRI morphological features achieved an Area Under the Curve (AUC) score of 0.83 ± 0.04 (95% C.I. 0.826-0.835) on the independent test set consisting of 20 lesions, while the model using only dynamic features performed an AUC of 0.91 ± 0.03 (95% C.I. 0.907-0.915). Despite being obtained on a small test sample, these results show the potential of features derived from DCE images for breast lesions classification.
10 Apr 2026
Read appraisal →Physics in medicine and biology
SSI-Net: a hybrid physics-constrained deep learning framework for quantitative ultrasound speed-of-sound reconstruction
Objective.Quantitative ultrasound tomography faces challenges in reconstructing speed-of-sound (SoS) distributions due to the ill-posed nature of the inverse problem and the computational complexity of full-waveform inversion. This study aims to develop an efficient and physically consistent deep learning framework for accurate SoS mapping from ultrasound signals.Approach. We propose speed of sound inversion network (SSI-Net), a dual data-physics-driven framework that integrates a bidirectional gated recurrent unit encoder for temporal feature extraction, a U-Net decoder for high-resolution spatial mapping, and a physics-constraint module based on an exact, differentiable finite-difference time-domain solver of the nonlinear Westervelt equation. The network model is trained through a joint optimization strategy that minimizes both data-fidelity loss and physics-informed residual loss, with a two-phase training schedule to ensure stable convergence.Main results.SSI-Net was validated across simulated, tissue-mimicking phantom, andin vivomouse datasets. It achieved superior reconstruction accuracy, with peak signal-to-noise ratio improvements of 7.4%-18.9% over state-of-the-art methods, while reducing the physics-based residual by 25.6%-57.5%. The framework demonstrated strong generalization to different transducer geometries (ring and linear arrays) and maintained stable performance across a ten-fold frequency range. Inference was completed within 109.4 ms per sample.Significance.By embedding an exact wave solver into a trainable architecture, SSI-Net combines the representational capacity of deep learning with strict physical consistency. It provides a robust, efficient, and clinically translatable solution for quantitative SoS imaging, offering a promising tool for tissue characterization in diagnostic ultrasound.
9 Apr 2026
Read appraisal →Journal of medical Internet research
Initial Insights Into an Institutional Secure Large Language Model for Magnetic Resonance Imaging Examination Requests: Retrospective Study
BACKGROUND: Incomplete clinical details on magnetic resonance imaging (MRI) examination requests (MERs) can lead to suboptimal protocol selection. An institutional secure large language model (sLLM) with access to manually retrieved salient data from the electronic medical record (EMR) may improve request completeness and protocol accuracy across multiple MRI subspecialties. OBJECTIVE: The objective of this study was to compare clinician MERs with sLLM-augmented MERs for information quality and to evaluate the protocoling accuracy of the sLLM versus board-certified radiologists across body, musculoskeletal, and neuroradiology MRI. METHODS: This retrospective study included 608 random outpatient MRI examinations performed between September 2023 and July 2024 (body 206, musculoskeletal 203, neuroradiology 199). The cohort comprised 528 patients (mean 51.2 years, SD 19.2; range 4-93; n=279, 52.8% women, n=249, 47.2% men). MERs without EMR access were excluded. A privately hosted Anthropic Claude 3.5 model (temperature 0) augmented each MER with manually retrieved salient EMR data and, via rule-based parsing, mapped the extracted elements onto predefined institutional criteria to recommend region or coverage and contrast use. Two experienced radiologists established a consensus reference standard. Two board-certified general radiologists (Rad 3 and Rad 4) and the sLLM were compared with this standard. Clinical information quality was graded using the Reason-for-Exam Imaging Reporting and Data System (RI-RADS). Interrater reliability was quantified with Gwet AC1. Paired accuracies were compared with the McNemar test to determine whether there was a statistically significant difference. RESULTS: Interreader agreement for RI-RADS was almost perfect for sLLM-augmented MERs (AC1 0.97, 95% CI 0.94-0.99) and moderate for clinician MERs (AC1 0.43, 95% CI 0.34-0.52). Limited or deficient clinical information (RI-RADS C/D) fell to 0% to 0.7% (0/608 to 4/608) with sLLM augmentation vs 4.1% to 20.4% (25/608 to 124/608) for clinician MERs. Overall protocol accuracy was 93.1% (566/608; 95% CI 89.6-96.6) for the sLLM, 91.4% (556/608; 95% CI 87.6-95.3) for Rad 3, and 92.1% (560/608; 95% CI 88.4-95.8) for Rad 4 (sLLM vs Rad 3 P=.23 vs Rad 4 P=.40). Region or coverage accuracy was similar (sLLM: 579/608, 95.2%; Rad 3: 585/608, 96.2%; Rad 4: 573/608, 94.2%; P=.46 and P=.36). Contrast decisions were more accurate using the sLLM at 94.4% (574/608; 95% CI 91.3-97.5) vs Rad 3 at 92.1% (560/608; 95% CI 88.4-95.8; P=.027) and were not significantly different to Rad 4 at 92.9% (565/608; 95% CI 89.4-96.4; P=.16). Subspecialty analyses showed similar patterns, with the sLLM outperforming Rad 4 for musculoskeletal MRI contrast decisions (96.6% vs 91.1%; P=.006) and matching readers elsewhere. Manual review indicated that sLLM improvements arose from EMR details not listed on the MER (infection/inflammation, tumor history, prior surgery). No clinically significant hallucinations were identified in a manual review of discordant cases. CONCLUSIONS: Across body, musculoskeletal, and neuroradiology MRI, sLLM-augmented examination requests improved clinical context and enhanced contrast selection while demonstrating accuracy comparable to general radiologists for region or coverage. Integrating sLLMs into routine vetting workflows may reduce manual workload in protocol selection for more efficient, standardized protocoling.
8 Apr 2026
Read appraisal →Pulmonology
Comparative diagnostic performance of advanced endoscopic techniques in suspected sarcoidosis: A systematic review and network meta-analysis
BACKGROUND: While various endoscopic techniques are used for minimally invasive tissue sampling in suspected sarcoidosis, comprehensive evidence comparing their diagnostic efficacy is limited, and the optimal approach remains unclear. RESEARCH QUESTION: Which endoscopic technique provides the greatest diagnostic advantage for sarcoidosis? STUDY DESIGN AND METHODS: We systematically searched 3 databases for studies published up to 22 April 2025. A network meta-analysis was conducted within a Bayesian framework, with diagnostic yield as the primary outcome. RESULTS: A total of 35 studies were included. Compared to TBLB, both EBUS-guided intranodal forceps biopsy (EBUS-IFB) (RR 3.50, 2.14-5.84) and EBUS-guided transbronchial mediastinal cryobiopsy (EBUS-TMC) (RR 3.52, 1.97-6.48) showed higher diagnostic yield. SUCRA rankings placed EBUS-IFB (0.912) and EBUS-TMC (0.910) at the top, followed by EBUS-guided core needle biopsy (EBUS-CNB) (0.587), EUS-FNA (0.582), and EBUS-TBNA (0.560), EUS-B-FNA (0.498). Conventional techniques, including cTBNA (0.279), TBLB (0.168), and EBB (0.001), ranked lowest. All procedures had a favourable safety profile. CONCLUSIONS: In centres with appropriate expertise, EBUS-IFB and EBUS-TMC may offer superior diagnostic performance for sarcoidosis. However, given the low certainty of evidence, these conclusions should be interpreted with caution. Clinical decisions should be individualised, considering the patient's condition when determining the final diagnostic strategy.
8 Apr 2026
Read appraisal →Journal of neurology
Accuracy of muscle ultrasonography in detecting fasciculations for the diagnosis of amyotrophic lateral sclerosis: a systematic review and meta-analysis
PURPOSE: This systematic review and meta-analysis aims to evaluate the diagnostic accuracy of muscle ultrasonography in detecting fasciculations for the diagnosis of amyotrophic lateral sclerosis (ALS). METHODS: Following PRISMA-DTA guidelines, we systematically searched PubMed, Embase, Cochrane Library, Ovid Medline, Sinomed, Web of Science, CNKI and VIP for studies published up to July 8, 2025 that evaluated muscle ultrasonography to detect fasciculations for ALS diagnosis. The study protocol was registered in PROSPERO (CRD420251057866). Studies were screened using predefined inclusion and exclusion criteria and data were extracted. Risk of bias was assessed with QUADAS-2. Statistical analyses (Stata 16.0 and R 4.5.1 with the "midas," "metandi," and "mada" packages) were used to calculate pooled sensitivity (Sen), specificity (Spe), positive likelihood ratio (LR+), negative likelihood ratio (LR-), and diagnostic odds ratio (DOR). We constructed forest plots, hierarchical summary receiver operating characteristic (HSROC) curves, summary ROC (SROC) curves and calculated the area under the SROC curve (AUC). Univariate meta-regression and subgroup analyses explored sources of heterogeneity. Publication bias was assessed using Deeks' funnel plot asymmetry test. Fagan nomograms were also used to illustrate the changes from pre-test to post-test probability and to enhance clinical interpretability. RESULTS: Thirteen studies involving 1176 participants met the inclusion criteria. Muscle ultrasonography for fasciculation detection in ALS yielded a pooled sensitivity of 0.87 (95% CI 0.83-0.91) and specificity of 0.91 (95% CI 0.86-0.94). The pooled LR+ was 9.81 (95% CI 6.25-15.40) and LR- was 0.14 (95% CI 0.10-0.19), with a DOR of 70.03 (95% CI 41.72-117.56). The area under the SROC curve was 0.94 (95% CI 0.91-0.95). Meta-regression identified scan duration as a primary factor influencing diagnostic accuracy, with scan durations ≥ 30 s associated with higher sensitivity but relatively lower specificity. Deeks' funnel plot showed no significant asymmetry (p = 0.61), indicating no notable publication bias. Fagan nomograms showed that, at a pre-test probability of 30%, the post-test probability increased to 81% after a positive MUS result and decreased to 6% after a negative result. CONCLUSION: Muscle ultrasonography demonstrates good pooled diagnostic accuracy for detecting fasciculations in ALS and may serve as a useful adjunct to electrodiagnostic evaluation. Scan duration appears to significantly affect the diagnostic performance, with longer scanning improving sensitivity at the cost of reduced specificity. We speculate that prolonged scanning may be more useful in clinical scenarios where fasciculations are subtle or atypical, whereas shorter scanning may be sufficient when fasciculations are already readily apparent. Nevertheless, further large-scale prospective studies are needed to validate standardized scanning protocols and to better define the clinical role of MUS in ALS diagnostic pathways.
7 Apr 2026
Read appraisal →Cancer research
Multicenter Histology Image Integration and Multiscale Deep Learning Support Machine Learning-Enabled Pediatric Sarcoma Classification
UNLABELLED: Pediatric sarcomas present diagnostic challenges due to their rarity and diverse subtypes, often requiring specialized pathology expertise and costly genetic tests. To overcome these barriers, we developed a computational pipeline leveraging deep learning methods to accurately classify pediatric sarcoma subtypes from digitized histology slides. To ensure classifier generalizability and minimize center-specific artifacts, a dataset comprising 867 whole-slide images (WSI) from three medical centers and the Children's Oncology Group was collected and harmonized. Multiple convolutional neural network and vision transformer (ViT) architectures were systematically evaluated as feature extractors for SAMPLER-based WSI representations, and input parameters, such as tile size combinations and resolutions, were tested and optimized. The analysis showed that advanced ViT foundation models (UNI and CONCH) significantly outperformed earlier approaches, and incorporating multiscale features enhanced classification accuracy. The optimized models achieved high performance, distinguishing rhabdomyosarcoma (RMS) from non-RMS soft-tissue sarcomas (NRSTS) with an AUC of 0.969 and differentiating RMS subtypes (alveolar vs. embryonal) with an AUC of 0.961. Additionally, a two-stage pipeline effectively identified scarce Ewing sarcoma images from other NRSTS (AUC = 0.929). Compared with conventional transformer-encoder architectures used for WSI representations, these SAMPLER-based classifiers were three orders of magnitude faster to train, despite operating entirely without a graphical processing unit. This study highlights that digital histopathology paired with rigorous image harmonization provides a powerful solution for pediatric sarcoma classification. SIGNIFICANCE: An approach pairing a multi-institutional dataset with a published pipeline extends imaging-based diagnostic capabilities and creates the potential for accurate, rapid cancer diagnoses across resource-limited and remote settings. This article is part of a special series: Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI .
5 Apr 2026
Read appraisal →European journal of pediatric surgery : official journal of Austrian Association of Pediatric Surgery ... [et al] = Zeitschrift fur Kinderchirurgie
Latest Developments in Artificial Intelligence and Machine Learning Models in General Pediatric Surgery
Artificial intelligence (AI) and machine learning (ML) models rapidly transform health care with applications ranging from diagnostic image interpretation, predictive modeling, personalized treatment planning, real-time intraoperative guidance, and outcome prediction. However, their implementation in general pediatric surgery remains limited due to the rarity and complexity of pediatric surgical conditions, small and heterogeneous datasets, and a lack of formal AI training and competencies among pediatric surgeons.This narrative review explores the current landscape of AI and ML applications in general pediatric surgery, focusing on five key conditions: appendicitis, necrotizing enterocolitis, Hirschsprung's disease, congenital diaphragmatic hernia, and biliary atresia. For each, we summarize recent developments, including the use of AI in image analysis, diagnostic support, prediction of disease severity and outcome, postoperative monitoring, and histopathological evaluation. We also highlight novel tools such as explainable AI models, natural language processing, and wearable technologies.Recent findings demonstrate promising diagnostic and prognostic capabilities across multiple conditions. However, most AI/ML models still require external validation and standardization. The review underscores the importance of collaborative, multicenter research based on joint datasets as well as targeted AI education for pediatric surgeons to fully explore the benefits of these technologies in clinical practice.AI and ML offer significant potential to improve pediatric surgical care, but broader implementation will require multicenter collaboration, a robust dataset, and targeted AI education for pediatric surgeons.
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 →Annals of oncology : official journal of the European Society for Medical Oncology
Deep learning discriminates thymic epithelial tumors' histological subtypes using digital pathology.
BACKGROUND: Thymic epithelial tumors (TETs) are rare malignancies that pose significant diagnostic challenges due to their heterogeneous histological patterns and substantial interobserver variability in classification. Despite standardized World Health Organization (WHO) classification criteria, diagnostic concordance remains suboptimal, particularly in nonexpert settings, where second-opinion reviews lead to diagnostic reclassification in up to 57% of cases. Deep learning may offer a tool to reduce diagnostic variability and improve the consistency of histological classification. MATERIALS AND METHODS: We trained a deep learning-based model using hematoxylin-eosin (H&E) whole-slide images from The Cancer Genome Atlas as a training dataset. The model incorporated a novel hierarchical loss function designed to reflect clinically relevant tumor groupings based on treatment strategies and patient outcomes. We validated the model on 112 consecutive cases from the University of Chicago, with diagnoses confirmed by an expert thoracic pathologist. Model performances were evaluated using both a three-group hierarchical scheme and the six-class WHO classification. RESULTS: In the clinically relevant hierarchical three-group classification (As: A + AB; Bs: B1 + B2 + B3; and thymic carcinoma), the model achieved an accuracy of 91.1% with Cohen's κ = 0.859, indicating almost perfect agreement. In the six-class classification (A, AB, B1, B2, B3, and thymic carcinoma), the accuracy was 77.7% with κ = 0.716. The model demonstrated 100% sensitivity and 94.6% accuracy for thymic carcinoma detection. Notably, 60% of misclassifications occurred within the same clinical management group, thereby limiting their impact on therapeutic decision making. CONCLUSION: This deep learning model demonstrates strong potential as a diagnostic tool for TETs classification, particularly in settings with limited thoracic pathology expertise. The high sensitivity for thymic carcinoma detection and robust performance across different tissue processing conditions suggest its clinical applicability for improving diagnostic consistency and supporting pathological decision making in both specialized and nonspecialized settings.
4 Apr 2026
Read appraisal →Clinical & experimental optometry
Advancements in artificial intelligence for meibography: clinical utility and contributions to meibomian gland dysfunction management
Recent advances in artificial intelligence (AI) have enhanced the capabilities of meibography by enabling objective and quantitative assessment of meibomian gland structure. This review explores the clinical utility of AI-based meibography in the diagnosis and management of Meibomian Gland Dysfunction (MGD), with a focus on segmentation, morphological analysis, and disease staging. Developments in deep learning have enabled more precise gland feature extraction, including gland dropout, density, and tortuosity, supporting efforts towards standardised and reproducible clinical evaluation. Although not the focus of this review, insights from traditional image processing techniques are referenced to highlight potential areas of improvement in current AI models. Key issues such as limited modelling of regional gland variation, restricted dataset diversity, and lack of standardised image quality control are discussed. Although significant progress has been made, further work is needed to ensure AI-driven meibography tools are generalisable, interpretable, and suitable for broad clinical implementation.
4 Apr 2026
Read appraisal →Ultrasound in medicine & biology
From Recognition to Action: Integrating Deep Learning and Robotic Control in Transthoracic Echocardiography
Population aging has driven a rise in heart failure cases, increasing the clinical burden on cardiac diagnostics. As a first-line imaging method, transthoracic echocardiography (TTE) faces limitations due to operator dependence, patient variability, and workflow inefficiencies. Meanwhile, advances in artificial intelligence (AI) and robotic ultrasound systems offer new potential pathways toward automated diagnosis. This review examines the current landscape of AI-based image analysis and robotic-assisted echocardiography. It presents a detailed analysis of advancements in artificial intelligence (AI) applied to echocardiography and the evolution of robotic ultrasound systems, aiming to introduce a discussion on semantic-to-motion mapping. By synthesizing recent progress and outlining future directions, we can correctly recognize the current maturity level of artificial intelligence development in the field of ultrasound examination and prepare well for the subsequent work.
4 Apr 2026
Read appraisal →European radiology
Deep learning-based segmentation of T1 and T2 cardiac MRI maps for automated disease detection.
OBJECTIVES: Parametric tissue mapping enables quantitative cardiac tissue characterization but is limited by inter-observer variability during manual delineation. Traditional approaches relying on average relaxation values and single cutoffs may oversimplify myocardial complexity. This study evaluates whether deep learning (DL) can achieve segmentation accuracy comparable to inter-observer agreement, explores the utility of statistical features beyond mean T1/T2 values, and assesses whether machine learning (ML) combining multiple intensity-based statistical features enhances disease detection. MATERIALS AND METHODS: T1 and T2 maps were manually segmented. The test subset was independently annotated by two observers, and inter-observer variability was assessed. A DL model was trained to segment the left-ventricular blood pool and myocardium. Average (A), lower quartile (LQ), median (M), and upper quartile (UQ) were computed for the myocardial pixels and employed in classification by applying cutoffs or in ML. Dice similarity coefficient (DICE) and mean absolute percentage error evaluated segmentation performance. Bland-Altman plots assessed inter-user and model-observer agreement. Receiver operating characteristic analysis determined optimal cutoffs. Pearson correlation compared features from model and manual segmentations. F1-score, precision, and recall evaluated classification performance. Wilcoxon test assessed differences between classification methods, with p < 0.05 considered statistically significant. RESULTS: 144 subjects (mean age 42.2 years ± 16.1, 76 men) were split into training (100), validation (15) and evaluation (29) subsets. Segmentation model achieved a DICE of 85.4%, surpassing inter-observer agreement. Random forest applied to all features increased F1-score (92.7%, p < 0.001). CONCLUSION: DL facilitates segmentation of T1/ T2 maps. Combining multiple features with ML improves disease detection. KEY POINTS: Question Manual segmentation of myocardial T1/T2 maps is time-consuming and affected by inter-observer variability; relying on single cutoff values for diagnosis may oversimplify myocardial complexity. Findings Deep learning achieves segmentation accuracy within inter-observer agreement, while machine learning improves disease detection compared to singular cutoffs. Clinical relevance Automated segmentation and feature extraction from T1/T2 maps can enhance workflow efficiency, reduce inter-observer variability, and improve diagnostic consistency. The high recall of the machine learning model minimizes missed diagnoses, ensuring more reliable disease detection.
4 Apr 2026
Read appraisal →Otolaryngology--head and neck surgery : official journal of American Academy of Otolaryngology-Head and Neck Surgery
Multicenter Clinical Validation of an Artificial Intelligence Diagnostic Classification Model for Laryngoscopy Images
OBJECTIVE: To develop and externally validate a computer-aided diagnosis (CADx) model using artificial intelligence (AI) for classifying laryngeal lesions from laryngoscopy images into high-risk (HR), low-risk (LR). STUDY DESIGN: Retrospective multicenter development of a CADx model and external validation on independent cohorts. SETTING: Multicenter tertiary referral hospitals (Italy, India, China, Greece, and Spain). METHODS: Over 20,000 images derived from laryngoscopic examinations were retrieved. Images were annotated based on histopathology or expert consensus. A deep learning model was trained using an internal dataset and evaluated on 2 external datasets to assess generalizability. The CADx model classifies only images containing visible lesions, discriminating between LR and HR categories. Diagnostic performance was measured using standard metrics, including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Model performance was compared with physicians of varying expertise and ChatGPT-4o. RESULTS: The computer-aided diagnosis model achieved a similar performance across internal and external datasets in distinguishing HR from LR lesions, with accuracy/AUC of 0.90/0.89 internally, 0.85/0.85 on the Greek dataset, and 0.88/0.88 on the Spanish dataset. The model's accuracy was statistically noninferior to that of otolaryngologists and expert laryngologists, and superior to general practitioners and ChatGPT-4o. CONCLUSION: This is a large multicenter clinical validation of a CADx model for laryngeal endoscopy, demonstrating generalizability and performance comparable to clinicians in discriminating between LR and HR lesions. The model's success supports its potential role in augmenting diagnostic capabilities, especially in resource-limited settings. A prospective multicenter clinical trial is underway to assess real-world clinical implementation.
2 Apr 2026
Read appraisal →Ann Med
Artificial intelligence-assisted diagnosis of early gastric cancer: present practice and future prospects
Gastric cancer (GC) occupies the first few places in the world among tumors in terms of incidence and mortality, causing serious harm to human health, and at the same time, its treatment greatly consumes the health care resources of all countries in the world. The diagnosis of GC is usually based on histopathologic examination, and it is very important to be able to detect and identify cancerous lesions at an early stage, but some endoscopists' lack of diagnostic experience and fatigue at work lead to a certain rate of under diagnosis. The rapid and striking development of Artificial intelligence (AI) has helped to enhance the ability to extract abnormal information from endoscopic images to some extent, and more and more researchers are applying AI technology to the diagnosis of GC. This initiative has not only improved the detection rate of early gastric cancer (EGC), but also significantly improved the survival rate of patients after treatment. This article reviews the results of various AI-assisted diagnoses of EGC in recent years, including the identification of EGC, the determination of differentiation type and invasion depth, and the identification of borders. Although AI has a better application prospect in the early diagnosis of ECG, there are still major challenges, and the prospects and limitations of AI application need to be further discussed.
1 Dec 2025
Read appraisal →