Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 16 appraisals
Clinica chimica acta; international journal of clinical chemistry
Exploring clinical feasibility of zero-shot learning in a large language model for immunofixation electrophoresis image interpretation
INTRODUCTION: Immunofixation electrophoresis (IFE) is the gold standard method for detecting and typing monoclonal immunoglobulins, but its interpretation requires expert evaluation and remains susceptible to subjectivity and variability. Recent advances in multimodal large language models (LLMs) have enabled zero-shot visual reasoning without task-specific training. This study aimed to evaluate the feasibility and diagnostic performance of zero-shot multimodal LLMs for automated interpretation of IFE images. MATERIALS AND METHODS: A dataset of 487 immunofixation electrophoresis images representing ten semantic classes (AK, AL, GK, GL, K, L, LGL, MK, ML, and NONE) was retrospectively collected and annotated by expert laboratory specialists. Two multimodal LLMs, ChatGPT-5.2 and Gemini3 Pro, were evaluated using a zero-shot visual question answering (ZS-VQA) framework without additional training or fine-tuning. Each image was analyzed using two prompt configurations: a simple prompt and a detailed prompt with structured diagnostic guidance. Model performance was assessed using precision, recall, F1-score, and confusion matrix analysis. RESULTS: Gemini outperformed ChatGPT across most classes, particularly with detailed prompts, achieving high F1-scores in clinically relevant monoclonal categories such as MK (89.80), AK (84.75), and GK (71.54). Detailed prompting consistently improved recall and F1-scores, underscoring the importance of prompt design. In contrast, ChatGPT showed lower recall, F1-scores, and classification consistency. Both models performed less effectively in visually subtle patterns. CONCLUSIONS: Zero-shot multimodal LLMs, particularly Gemini, show promising potential for interpreting IFE images without task-specific training. However, performance variability and limitations in certain classes indicate that further optimization and validation are required before clinical implementation.
18 May 2026
Read appraisal →Histochemistry and cell biology
Artificial intelligence in small tissue biopsies: diagnostic applications, histochemical integration, and methodological challenges in surgical pathology.
Small tissue biopsies, including renal core biopsies, bone marrow trephines, gastrointestinal endoscopic samples, prostate needle cores, liver biopsies, and skin punch or shave specimens, are fundamental to contemporary diagnostic pathology. Their limited volume, focal sampling, and susceptibility to technical artifacts impose distinct interpretive challenges, often requiring semi-quantitative assessment within a restricted architectural context. Although artificial intelligence (AI) has rapidly expanded in digital pathology, most models have been developed using large surgical resection specimens, with comparatively limited attention to small biopsy material. This minireview examines current and emerging applications of AI in small biopsy diagnostics across renal, hepatic, gastrointestinal, hematopathological, dermatopathological, and urological pathology. Reported applications include glomerular segmentation and fibrosis quantification in renal biopsies; automated cellularity, blast detection, and fibrosis grading in bone marrow trephines; dysplasia and microorganism detection in gastrointestinal biopsies; quantitative steatosis and fibrosis assessment in liver samples; tumor detection and grading in prostate cores; and neoplastic pattern recognition in skin specimens. Despite encouraging performance in research settings, substantial barriers to routine clinical implementation remain, including limited dataset size, class imbalance, pre-analytical variability, inter-institutional heterogeneity, and insufficient external validation. We discuss methodological considerations relevant to diagnostic practice, including multi-institutional validation, stain normalization, multimodal integration with histochemistry and ancillary testing, explainability, and regulatory oversight. In the context of small biopsies, AI should be regarded as a quantitative adjunct to morphological interpretation rather than an autonomous diagnostic system. Careful integration within established histopathological workflows is essential to ensure reproducibility, safety, and clinical accountability.
15 May 2026
Read appraisal →Scientific reports
The impact of tissue detection on diagnostic artificial intelligence algorithms in prostate digital pathology
Tissue detection is a crucial first step in most digital pathology applications. By applying image segmentation algorithms, all tissue is delineated and background discarded from further analyses, improving both computational efficiency and analytical results. Details of the segmentation algorithm are rarely reported, and there is a lack of studies investigating the downstream effects of a poor segmentation algorithm. Disregarding tissue detection quality could jeopardize patient safety if diagnostically relevant parts of the specimen are excluded from analysis in clinical applications. This study aims to determine whether performance of downstream tasks is sensitive to the tissue detection method, and to compare the performance of a classical and an AI-based tissue detection approach. To this end, we trained an AI model for Gleason grading of prostate cancer in whole slide images (WSIs) using two different tissue detection algorithms: thresholding (classical) and UNet++ (AI). A total of 33,823 WSIs scanned on seven digital pathology scanners were used to train the segmentation AI model. The downstream Gleason grading algorithm was trained and tested using 70,524 WSIs from 13 clinical sites scanned on 13 different scanners. On the slides where tissue could be detected by both algorithms, no significant difference in overall Gleason grading performance was observed. There was a decrease from 118 (0.43%) to 24 (0.09%) fully undetected tissue samples when switching from thresholding-based tissue detection to AI-based, suggesting this AI model may be more reliable than the classical model for avoiding total failures on slides with unusual appearance. Moreover, tissue detection dependent clinically significant variations in AI grading were observed in 3.5% of malignant slides, highlighting the role of tissue detection for optimal clinical performance of diagnostic AI.
15 May 2026
Read appraisal →Endocrine pathology
Lineage Classification of Pituitary Neuroendocrine Tumors From Whole-Slide Images Using Attention-Guided Graph Representation Learning
Pituitary neuroendocrine tumors (PitNETs) are common sellar neoplasms and represent a major component of routine pituitary pathology. In the 2022 World Health Organization (WHO) Classification, transcription factor-defined lineage assignment is central to diagnosis. However, lineage-related morphologic information on routine hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) has not been systematically characterized. We developed an attention-guided graph neural network to predict PitNET lineage directly from H&E-stained WSIs and to identify regions prioritized for classification. Consecutive patients who underwent surgical resection for PitNETs at Beijing Tiantan Hospital between 2021 and 2025 were included. In the internal hold-out set, the five fold-specific models achieved a mean F1-score of 92.78% and a mean balanced accuracy of 94.84%. In a temporally independent validation cohort, the final model achieved an F1-score of 87.64% and a balanced accuracy of 89.48%. Performance varied across lineages and histomorphologic subtypes, and the most common error pattern was misassignment of PIT-1 and T-PIT tumors to SF-1. Attention maps predominantly highlighted tumor-rich regions. Quantitative cell-level morphometry supported lineage-associated patterns, including larger cell and nuclear sizes in PIT-1 tumors, more elongated nuclei in SF-1 tumors, and higher cellular density with reduced intercellular spacing in T-PIT tumors. In six cases with multiple synchronous PitNETs of distinct lineages, patch-level prediction maps corresponded closely to transcription factor immunohistochemistry. In a small exploratory subset with available DNA methylation data, methylation classifier results were concordant with model predictions in 11 of 15 cases and with routine histopathologic diagnoses in 8 of 15 cases. These findings indicate that routine H&E-stained WSIs contain learnable morphologic information related to PitNET lineage. Attention-guided spatial modeling provides an interpretable framework for characterizing lineage-associated patterns in PitNETs.
12 May 2026
Read appraisal →Biomedical physics & engineering express
SAM2HIPT: a hybrid deep learning framework integrating SAM2 and HIPT with joint loss optimization for immunohistochemical cell nucleus segmentation
Nucleus segmentation in immunohistochemistry images plays a critical role in cancer diagnosis and treatment assessment. However, existing methods remain limited in segmentation accuracy and boundary delineation due to staining heterogeneity, densely packed cell distributions, and complex background interference. To address these challenges, this paper proposes a two-stage nucleus segmentation framework, termed SAM2HIPT. In the first stage, the pre-trained Segment Anything Model 2 (SAM2) is employed to generate initial segmentation predictions for input images, wherein the image encoder is kept frozen to preserve the pre-trained visual representation capacity while the mask decoder is fine-tuned to adapt to the characteristics of the pathological image domain; local texture, morphological, and boundary information are extracted through visual feature encoding to produce initial nucleus segmentation masks and spatial prior representations. In the second stage, the Hierarchical Image Pyramid Transformer is introduced to refine the initial segmentation results, performing multi-scale, multi-level feature representation and fusion of morphological, textural, and spatial structural information through a hierarchical vision Transformer architecture, thereby enhancing nuclear structural representation and boundary consistency. To enable collaborative optimization across both stages, a joint loss function is designed to impose unified constraints on segmentation accuracy and feature representation. Evaluated on two public histopathological benchmark datasets, BCData and DeepLIIF, the proposed method achieves Dice coefficients of 0.92 and 0.91, respectively, and HD95 boundary error values of 1.05 pixels and 1.10 pixels, demonstrating superior segmentation performance and robustness over multiple state-of-the-art baseline methods.
11 May 2026
Read appraisal →Cancer immunology, immunotherapy : CII
Artificial intelligence-powered H&E-based quantification of spatial tumor-infiltrating lymphocyte distribution identifies prognostic immune niches in colorectal cancer
PURPOSE: The prognostic significance of tumor-infiltrating lymphocytes (TILs) in colorectal cancer (CRC) is well established; however, existing approaches inadequately capture their spatial distribution. We investigated the prognostic implications of TIL spatial distribution in CRC using an artificial intelligence (AI)-based method. METHODS: A total of 202 patients with stage II-III CRC were included. TIL densities in intratumoral (iTIL) and stromal (sTIL) regions were quantified using AI-based analysis of hematoxylin and eosin (H&E)-stained images. Based on proximity to the tumor-stromal border (TSB), TILs were subclassified into core iTIL, bounding iTIL, bounding sTIL, and outermost sTIL. Immunoscore was calculated from CD3+ and CD8+ T-cell densities in the tumor center and invasive margin. RESULTS: Correlations between AI-based and pathologist assessments (iTIL: r = 0.57; sTIL: r = 0.70) were comparable to inter-pathologist correlations (iTIL: r = 0.47; sTIL: r = 0.70). In univariate Cox regression analysis, bounding iTIL, bounding sTIL, and outermost sTIL were significantly associated with recurrence-free survival (RFS), whereas core iTIL was not. Composite TIL and TSB scores were developed by incorporating the prognostically significant regions. In multivariable analysis, the TIL score (p = 0.001), TSB score (p < 0.001), and Immunoscore (p < 0.001) independently predicted RFS. In microsatellite instability-high tumors, only the TSB score remained prognostically significant. CONCLUSION: AI-powered spatial analysis of TILs, particularly the TSB score, demonstrated prognostic performance comparable to conventional Immunoscore, thereby supporting the value of spatial immune profiling and AI-driven analysis of H&E-stained slides for improved risk stratification in CRC.
11 May 2026
Read appraisal →Cytopathology : official journal of the British Society for Clinical Cytology
Comparative Evaluation of Transfer Learning Models for Detecting Malignant Cells in Urinary Cytology
AIMS AND OBJECTIVES: In the present paper, we compared the efficiency of six transfer learning models to detect malignant cells in urine cytology. We also applied an ensemble learning with weighted soft voting to assess its importance in the diagnostic accuracy in urine cytology. MATERIALS AND METHODS: The voided urine samples of total 104 cases of urothelial cell carcinoma (UCC) and 86 cases with no malignancy (benign) were selected. All the 104 cases of UCC were histopathology proven high grade urothelial cell carcinoma (UCC). Urine with negative cytology reports were followed up clinically. There were 446 images of benign samples and 1369 images from malignant samples on 100 x objective. We applied six transfer learning models (DenseNet121, inception_v3, ResNet50, MobileNetV2, VGG16 and Xception) to detect malignant cells in urine. To compare the performance of different models, dynamic training optimization was done and each model was auto stopped after their maximum performance. In addition, an ensemble learning with soft voting was used including the top three models to enhance the diagnostic accuracy. RESULT: Xception transfer model showed the highest sensitivity (88.57%), accuracy (86.55%), precision (80.52%) and FI score (84.35%). It was the best performing model. The other two top performing models were InceptionV3 and ResNet50. The area under curve of receiver operating characteristic (AUCROC) was ≤ 90 in all the transfer learning models. The accuracy, sensitivity, specificity, precision, F1-Score and AUCROC of the ensemble transfer learning model were as follows: 92.10%, 95.41%, 85.51%, 91.23%, 93.24% and 0.977 respectively. CONCLUSIONS: First time, we evaluated a large number of transfer learning models in urine cytology to detect malignant cells. All the models showed high sensitivity, specificity and accuracy. In addition, the ensemble learning technique with soft voting showed remarkable superior performance than individual top three transfer learning models. The techniques transfer learning and ensemble models have high potential to use in routine screening of urine.
6 May 2026
Read appraisal →Medical image analysis
PH2ST: Prompt-guided hypergraph learning for spatial transcriptomics prediction in whole slide images
Spatial Transcriptomics (ST) reveals the spatial distribution of gene expression in tissues, offering critical insights into biological processes and disease mechanisms. However, the high cost, limited coverage, and technical complexity of current ST technologies restrict their widespread use in clinical and research settings, making obtaining high-resolution transcriptomic profiles across large tissue areas challenging. Predicting ST from H&E-stained histology images has emerged as a promising alternative to address these limitations but remains challenging due to the heterogeneous relationship between histomorphology and gene expression, which is affected by substantial variability across patients and tissue sections. In response, we propose PH2ST, a prompt-guided hypergraph learning framework, which leverages limited ST signals to guide multi-scale histological representation learning for accurate and robust spatial gene expression prediction. Extensive evaluations on two public ST datasets and multiple prompt sampling strategies simulating real-world scenarios demonstrate that PH2ST not only outperforms existing state-of-the-art methods, but also shows strong potential for practical applications such as imputing missing spots, ST super-resolution, and local-to-global prediction, highlighting its value for scalable and cost-effective spatial gene expression mapping in biomedical contexts.
5 May 2026
Read appraisal →Platelets
Neural network reveals platelet age from fluorescence microscopy images
Platelets are small, anucleate cells with a primary physiological role in vascular damage repair (hemostasis) and initiation of thrombus formation in response to vascular injury. Platelets circulate approximately 7-10 days, slowly undergoing age-related changes in molecular composition, morphology, activation capacity, function, and surface receptor density. As older platelets are associated with poor clinical outcome, no in vitro tests are available to predict platelet age, or to determine the fitness of platelet transfusion products. In this study, we developed a convolutional neural network model that could determine platelets' chronological age from confocal microscopic images. The model was trained using platelets stored in platelet-rich plasma up to 8 hours and using routine platelet concentrates up to 10 days. The model predicted chronological age of stored platelets with >97% accuracy. To test our model in vivo, we analyzed a cohort of patients with acute myeloid leukemia, experiencing thrombocytopenia due to chemotherapy. Our model could reliably distinguish in vivo between samples with younger and older platelets during the course of treatment. This study demonstrates the ability to predict platelets' chronological age both in vitro during storage and in vivo, which may impact clinical transfusion medicine and the diagnosis and treatment of patients with platelet disorders. What is the context?Platelets circulate approximately 7–10 days, slowly undergoing age-related changes in molecular composition, morphology, activation capacity, function, and surface receptor density. During storage, proinflammatory changes occur in platelets over time and platelet age may be associated with poorer clinical outcomes. Therefore, understanding platelet aging and developing novel tools to interrogate platelet age will be useful. However, no platelet tests are currently available that can predict platelet age accurately based on platelet characteristics without keeping record of the (storage) time.What is new?We developed a convolutional neural network model that could determine platelets’ chronological age from confocal microscopic images. The model was trained using platelets stored in platelet-rich plasma up to 8 hours and using routine platelet concentrates up to 10 days. The model predicted chronological age of stored platelets with very high accuracy. Furthermore, our model was in vivo tested in a cohort of patients with acute myeloid leukemia, experiencing thrombocytopenia due to chemotherapy. Our model could reliably distinguish between in vivo samples with younger and older platelets during the course of treatment.What is the impact?Our study demonstrates the utility of a convolutional neural network model using confocal images of platelets to determine storage time-associated platelet phenotype in vitro as well as physiologic platelets’ chronological age in vivo. This model may be useful in several clinical settings, including clinical transfusion medicine, testing the quality of stored platelet concentrates, and diagnosing and treatment of patients with platelet disorders, thrombocytopenia, or thrombosis.
18 Apr 2026
Read appraisal →Computers in biology and medicine
MIPHEI-ViT: Multiplex immunofluorescence prediction from H&E images using ViT foundation models
Histopathological analysis is a cornerstone of cancer diagnosis, with Hematoxylin and Eosin (H&E) staining routinely acquired for every patient to visualize cell morphology and tissue architecture. On the other hand, multiplex immunofluorescence (mIF) enables more precise cell type identification via proteomic markers, but has yet to achieve widespread clinical adoption due to cost and logistical constraints. To bridge this gap, we introduce MIPHEI (Multiplex Immunofluorescence Prediction from H&E Images), a U-Net-inspired architecture that leverages a ViT pathology foundation model as an encoder to predict mIF signals from H&E images using rich pretrained representations. MIPHEI targets a comprehensive panel of markers spanning nuclear content, immune lineages (T cells, B cells, myeloid), epithelium, stroma, vasculature, and proliferation. We train our model using the publicly available OrionCRC dataset of restained H&E and mIF images from colorectal cancer tissue, and validate it on five independent datasets: HEMIT, PathoCell, IMMUcan, Lizard and PanNuke. On OrionCRC test set, MIPHEI achieves accurate cell-type classification from H&E alone, with F1 scores of 0.93 for Pan-CK, 0.83 for α-SMA, 0.68 for CD3e, 0.36 for CD20, and 0.28 for CD68, substantially outperforming both a state-of-the-art baseline and a random classifier for most markers. Our results indicate that, for some molecular markers, our model captures the complex relationships between nuclear morphologies in their tissue context, as visible in H&E images and molecular markers defining specific cell types. MIPHEI offers a promising step toward enabling cell-type-aware analysis of large-scale H&E datasets, in view of uncovering relationships between spatial cellular organization and patient outcomes.
15 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 →Journal of biophotonics
Cell-MICS: Detecting Immune Cells With Label-Free Two-Photon Autofluorescence and Deep Learning
Multiphoton imaging has been widely used for deep-tissue imaging. Although its label-free, metabolic contrast is ideal for investigating inflammation, the label-free two-photon induced autofluorescence is often regarded as less specific compared to conventional antibody markers. In this work, we investigate the potential for multiphoton imaging with computational specificity (MICS) by training a convolutional neural network on images of different immune cells. A low-complexity squeezeNet architecture was able to achieve reliable immune cell classification results (0.89 ROC-AUC, 0.95 PR-AUC for binary classification between T cells and neutrophils; 0.689 F1 score, 0.697 precision, 0.748 recall for multi-class classification between six isolated cell types). Perturbation tests confirmed that the model was not confused by the extracellular environment and that 2P-AF from NADH and FAD is equally important for the classification. In the future, deep learning could provide computational specificity for specific immune cells in unstained tissues, with great potential for label-free in vivo endomicroscopy.
4 Apr 2026
Read appraisal →The Lancet. Oncology
Deep learning on histopathological images to predict breast cancer recurrence risk and chemotherapy benefit: a multicentre, model development and validation study
BACKGROUND: Genomic assays such as Oncotype DX have transformed adjuvant treatment selection for hormone receptor-positive, HER2-negative, early breast cancer but remain inaccessible to many patients because of high cost and logistical barriers. We aimed to develop and validate an artificial intelligence (AI) model that estimates Oncotype DX 21-gene recurrence scores directly from routine histopathology slides and clinicopathological variables. METHODS: In this multicentre, model development and validation study, a multimodal deep-learning model was trained on digital whole-slide images and clinical features using a foundation model pre-trained on 171 189 histopathology slides for predicting Oncotype DX recurrence score. We included slides from patients with hormone receptor-positive, HER2-negative, invasive breast cancers and without scanning artifacts and with at least 100 tissue tiles (1·6 mm2). The model was fine-tuned and validated on the TAILORx randomised trial (8284 patients after quality control). Prognostic and predictive performance was assessed in the TAILORx-test set and externally validated in six independent cohorts (Carmel, Haemek, and Sheba medical centres [Israel], the University of Chicago Medical Center [USA], the Australian Breast Cancer Tissue Bank [Australia], and the Cancer Genome Atlas Breast Invasive Carcinoma project [USA]). FINDINGS: In the TAILORx-test set (n=2407), the AI model classified 1097 (45·6%) patients as low risk, 1021 (42·4%) as intermediate risk, and 289 (12·0%) as high risk. For identifying high genomic-risk disease (recurrence score ≥26), the area under the curve (AUC) was 0·898 (95% CI 0·879-0·913). AI-based risk stratification was prognostic for recurrence-free interval (hazard ratio 2·61 [95% CI 1·68-4·04]), distant recurrence-free interval (2·88 [1·73-4·79]), and disease-free survival (1·32 [0·92-1·89]). Chemotherapy benefit was evident in premenopausal patients classified by AI as being at high risk (0·63 [0·46-0·86]) but absent in postmenopausal patients classified by AI as being at low risk (0·94 [0·78-1·12]). 151 (31·3%) clinically high-risk postmenopausal women (by MINDACT criteria) were reclassified as low AI risk with no chemotherapy benefit. Analysis on external cohorts (5497 patients) showed that the model is transferable to new data with high generalisability (recurrence score ≥26 AUC ranging from 0·858 to 0·903). INTERPRETATION: These findings show that AI applied to routine histopathology can serve as a practical and scalable tool for guiding chemotherapy decisions in hormone receptor-positive, HER2-negative, early breast cancer. This approach has the potential to reduce unnecessary chemotherapy and broaden access to precision oncology, particularly in resource-limited settings where genomic testing remains unavailable or unaffordable. FUNDING: Israel Innovation Authority (Kamin), Zimin Institute for Artificial Intelligence Solutions in Healthcare, Israel Precision Medicine Partnership program, and Israel Cancer Research Fund.
4 Apr 2026
Read appraisal →Clinical chemistry
Towards Clinical Integration of Deep Learning-Based Classification of Urinary Sediment Particles from Digital Microscopy Images: A Prospective Study
BACKGROUND: Urinalysis is a standard clinical test that includes the microscopic examination of urinary sediment to identify formed elements. Manual evaluation by laboratory technicians is time-intensive and subject to human error. Automated analysis using digital microscopy images presents a potential alternative. This study evaluates the integration of a deep learning approach to automatically classify urinary sediment images in the clinical laboratory, including independent prospective validation of its performance. METHODS: An annotated data set comprising 13 classes of urinary sediment elements was created from a database of Sysmex UD-10 digital microscope images. An EfficientNet-based model was trained and tested across three experimental scenarios to evaluate the effects of data collection strategies on performance. Uncertainty calibration was examined. The model's robustness and interpretability were examined using gradient-weighted class activation mapping (Grad-CAM) to visualize influential image regions and t-distributed stochastic neighbor embedding (t-SNE) to analyze learned feature embeddings. Lastly, a graphical user interface was developed for a prospective evaluation in the laboratory. RESULTS: The model achieved approximately 97% overall accuracy on the test set. Experiments revealed sensitivity to data set variability, suggesting that performance may improve by integrating additional training examples. Confidence scores aligned with accuracy, and interpretability analyses showed that the model focused on relevant image regions and learned embeddings demonstrated clear class separation. In the prospective evaluation, top 1 and top 3 accuracies decreased to approximately 78% and 92%, respectively. CONCLUSIONS: Our results indicate that a lightweight deep learning model can achieve high performance in classifying urine particles. Analysis of discrepancies between retrospective and prospective evaluations provides important insights toward reliable clinical application.
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 →J Allergy Clin Immunol
Explainable artificial intelligence predicts inflammatory and spatial heterogeneity from nasal polyp histology
BACKGROUND: Chronic rhinosinusitis with nasal polyps is a heterogeneous disorder characterized by diverse inflammatory signatures and endotypes. OBJECTIVE: We sought to develop a histology-based deep learning network for predicting inflammatory gene signatures and spatial patterns in chronic rhinosinusitis with nasal polyps. METHODS: We developed HE2Signature, a deep learning model, using 70 hematoxylin and eosin-stained whole-slide images of nasal polyps paired with corresponding endotypic signature gene expression profiles derived from transcriptomic data. The model was validated in an internal cohort (n = 30) and tested in an independent external cohort (n = 224) from 4 medical centers. The performance was evaluated using correlation and confusion matrix analyses and receiver-operating characteristic curves. Spatial predictions were validated by immunohistochemistry. RESULTS: The HE2Signature demonstrated strong positive correlations between the predicted and actual expression levels of 28 of the 33 signature genes. The internal validation cohort was stratified into distinct endotypes on the basis of predicted signature gene expression, achieving area under the curve values of 0.833, 0.903, and 0.935 for the T1, T2, and T3 endotypes, respectively. In the external cohort, the combination of HE2Signature-predicted Fc epsilon receptor II and CST1 effectively identified the T2 endotype, with a receiver-operating characteristic value of 0.716. Histopathologic examination of high-prediction patches revealed characteristic features associated with signature gene expression. Heatmaps and immunohistochemistry confirmed the accuracy of the model in mapping spatial gene expression patterns. HE2Signature-predicted spatial expression of signature genes correlated with blood and tissue eosinophil counts, Lund-Kennedy score, and Sino-Nasal Outcome Test-22 score, particularly within subepithelial regions. CONCLUSIONS: Our study introduces the first histology-based, explainable deep learning model capable of predicting inflammatory gene signatures and spatial molecular heterogeneity. This proof-of-concept highlights that artificial intelligence-powered histopathologic analysis can generate digital biomarkers by linking tissue patterns with molecular profiles, providing a clinically applicable framework for endotype-guided precision medicine in chronic rhinosinusitis with nasal polyps.
1 Dec 2025
Read appraisal →