The development of therapeutics builds on testing their efficiency in vitro. To optimize gene therapies, for example, fluorescent reporters expressed by treated cells are typically utilized as readouts. Traditionally, their global fluorescence signal has been used as an estimate of transduction efficiency. However, analysis in individual cells within a living 3D tissue remains a challenge. Readout on a single-cell level can be realized via fluo-rescence-based flow cytometry at the cost of tissue dissociation and loss of spatial information. Complementary, spatial information is accessible via immunofluorescence of fixed samples. Both approaches impede time-dependent studies on the delivery of the vector to the cells. Here, quantitative 3D characterization of viral transduction efficiencies in living retinal organoids is introduced. The approach combines quantified gene delivery efficiency in space and time, leveraging human retinal organ-oids, engineered adeno-associated virus (AAV) vectors, confocal live imaging, and deep learning-based image segmentation. The integration of these tools in an organoid imaging and analysis pipeline allows quantitative testing of future treatments and other gene delivery methods. It has the potential to guide the development of therapies in biomedical applications.

Purpose: Intensity-modulated proton therapy (IMPT) offers precise tumor coverage while sparing organs at risk (OARs) in head and neck (H&N) cancer. However, its sensitivity to anatomical changes requires frequent adaptation through online adaptive radiation therapy (oART), which depends on fast, accurate dose calculation via Monte Carlo (MC) simulations. Reducing particle count accelerates MC but degrades accuracy. To address this, denoising low-statistics MC dose maps is proposed to enable fast, high-quality dose generation. Methods: We developed a diffusion transformer-based denoising framework. IMPT plans and 3D CT images from 80 H&N patients were used to generate noisy and high-statistics dose maps using MCsquare (1 min and 10 min per plan, respectively). Data were standardized into uniform chunks with zero-padding, normalized, and transformed into quasi-Gaussian distributions. Testing was done on 10 H&N, 10 lung, 10 breast, and 10 prostate cancer cases, preprocessed identically. The model was trained with noisy dose maps and CT images as input and high-statistics dose maps as ground truth, using a combined loss of mean square error (MSE), residual loss, and regional MAE (focusing on top/bottom 10% dose voxels). Performance was assessed via MAE, 3D Gamma passing rate, and DVH indices. Results: The model achieved MAEs of 0.195 (H&N), 0.120 (lung), 0.172 (breast), and 0.376 Gy[RBE] (prostate). 3D Gamma passing rates exceeded 92% (3%/2mm) across all sites. DVH indices for clinical target volumes (CTVs) and OARs closely matched the ground truth. Conclusion: A diffusion transformer-based denoising framework was developed and, though trained only on H&N data, generalizes well across multiple disease sites.

We present ReXVQA, the largest and most comprehensive benchmark for visual question answering (VQA) in chest radiology, comprising approximately 696,000 questions paired with 160,000 chest X-rays studies across training, validation, and test sets. Unlike prior efforts that rely heavily on template based queries, ReXVQA introduces a diverse and clinically authentic task suite reflecting five core radiological reasoning skills: presence assessment, location analysis, negation detection, differential diagnosis, and geometric reasoning. We evaluate eight state-of-the-art multimodal large language models, including MedGemma-4B-it, Qwen2.5-VL, Janus-Pro-7B, and Eagle2-9B. The best-performing model (MedGemma) achieves 83.24% overall accuracy. To bridge the gap between AI performance and clinical expertise, we conducted a comprehensive human reader study involving 3 radiology residents on 200 randomly sampled cases. Our evaluation demonstrates that MedGemma achieved superior performance (83.84% accuracy) compared to human readers (best radiology resident: 77.27%), representing a significant milestone where AI performance exceeds expert human evaluation on chest X-ray interpretation. The reader study reveals distinct performance patterns between AI models and human experts, with strong inter-reader agreement among radiologists while showing more variable agreement patterns between human readers and AI models. ReXVQA establishes a new standard for evaluating generalist radiological AI systems, offering public leaderboards, fine-grained evaluation splits, structured explanations, and category-level breakdowns. This benchmark lays the foundation for next-generation AI systems capable of mimicking expert-level clinical reasoning beyond narrow pathology classification. Our dataset will be open-sourced at https://huggingface.co/datasets/rajpurkarlab/ReXVQA

BackgroundCoeliac disease, an autoimmune disorder affecting approximately 1% of the global population, is typically diagnosed on a duodenal biopsy. However, inter-pathologist agreement on coeliac disease diagnosis is only around 80%. Existing machine learning solutions designed to improve coeliac disease diagnosis often lack interpretability, which is essential for building trust and enabling widespread clinical adoption. ObjectiveTo develop an interpretable AI model capable of segmenting key histological structures in duodenal biopsies, generating explainable segmentation masks, estimating intraepithelial lymphocyte (IEL)-to-enterocyte and villus-to-crypt ratios, and diagnosing coeliac disease. DesignSemantic segmentation models were trained to identify villi, crypts, IELs, and enterocytes using 49 annotated 2048x2048 patches at 40x magnification. IEL-to-enterocyte and villus-to-crypt ratios were calculated from segmentation masks, and a logistic regression model was trained on 172 images to diagnose coeliac disease based on these ratios. Evaluation was performed on an independent test set of 613 duodenal biopsy scans from a separate NHS Trust. ResultsThe villus-crypt segmentation model achieved a mean PR AUC of 80.5%, while the IEL-enterocyte model reached a PR AUC of 82%. The diagnostic model classified WSIs with 96% accuracy, 86% positive predictive value, and 98% negative predictive value on the independent test set. ConclusionsOur interpretable AI models accurately segmented key histological structures and diagnosed coeliac disease in unseen WSIs, demonstrating strong generalization performance. These models provide pathologists with reliable IEL-to-enterocyte and villus-to-crypt ratio estimates, enhancing diagnostic accuracy. Interpretable AI solutions like ours are essential for fostering trust among healthcare professionals and patients, complementing existing black-box methodologies. What is already known on this topicPathologist concordance in diagnosing coeliac disease from duodenal biopsies is consistently reported to be below 80%, highlighting diagnostic variability and the need for improved methods. Several recent studies have leveraged artificial intelligence (AI) to enhance coeliac disease diagnosis. However, most of these models operate as "black boxes," offering limited interpretability and transparency. The lack of explainability in AI-driven diagnostic tools prevents widespread adoption by healthcare professionals and reduces patient trust. What this study addsThis study presents an interpretable semantic segmentation algorithm capable of detecting the four key histological structures essential for diagnosing coeliac disease: crypts, villi, intraepithelial lymphocytes (IELs), and enterocytes. The model accurately estimates the IEL-to-enterocyte ratio and the villus-to-crypt ratio, the latter being an indicator of villous atrophy and crypt hyperplasia, thereby providing objective, reproducible metrics for diagnosis. The segmentation outputs allow for transparent, explainable decision-making, supporting pathologists in coeliac disease diagnosis with improved accuracy and confidence. This study presents an AI model that automates the estimation of the IEL-to-enterocyte ratio--a labour-intensive task currently performed manually by pathologists in limited biopsy regions. By minimising diagnostic variability and alleviating time constraints for pathologists, the model provides an efficient and practical solution to streamline the diagnostic workflow. Tested on an independent dataset from a previously unseen source, the model demonstrates explainability and generalizability, enhancing trust and encouraging adoption in routine clinical practice. Furthermore, this approach could set a new standard for AI-assisted duodenal biopsy evaluation, paving the way for the development of interpretable AI tools in pathology to address the critical challenges of limited pathologist availability and diagnostic inconsistencies.

While Deep Learning (DL) models trained on Magnetic Resonance Imaging (MRI) have shown promise for prostate cancer detection, their lack of direct biological validation often undermines radiologists trust and hinders clinical adoption. Radiologic-histopathologic (rad-path) correlation has the potential to validate MRI-based lesion detection using digital histopathology. This study uses automated and manually annotated digital histopathology slides as a standard of reference to evaluate the spatial extent of lesion annotations derived from both radiologist interpretations and DL models previously trained on prostate bi-parametric MRI (bp-MRI). 117 histopathology slides were used as reference. Prospective patients with clinically significant prostate cancer performed a bp-MRI examination before undergoing a robotic radical prostatectomy, and each prostate specimen was sliced using a 3D-printed patient-specific mold to ensure a direct comparison between pre-operative imaging and histopathology slides. The histopathology slides and their corresponding T2-weighted MRI images were co-registered. We trained DL models for cancer detection on large retrospective datasets of T2-w MRI only, bp-MRI and histopathology images and did inference in a prospective patient cohort. We evaluated the spatial extent between detected lesions and between detected lesions and the histopathological and radiological ground-truth, using the Dice similarity coefficient (DSC). The DL models trained on digital histopathology tiles and MRI images demonstrated promising capabilities in lesion detection. A low overlap was observed between the lesion detection masks generated by the histopathology and bp-MRI models, with a DSC = 0.10. However, the overlap was equivalent (DSC = 0.08) between radiologist annotations and histopathology ground truth. A rad-path correlation pipeline was established in a prospective patient cohort with prostate cancer undergoing surgery. The correlation between rad-path DL models was low but comparable to the overlap between annotations. While DL models show promise in prostate cancer detection, challenges remain in integrating MRI-based predictions with histopathological findings.

IntroductionExplainable Artificial Intelligence (XAI) methods enhance the diagnostic efficiency of clinical decision support systems by making the predictions of a convolutional neural networks (CNN) on brain imaging more transparent and trustworthy. However, their clinical adoption is limited due to limited validation of the explanation quality. Our study introduces a framework that evaluates XAI methods by integrating neuroanatomical morphological features with CNN-generated relevance maps for disease classification. MethodsWe trained a CNN using brain MRI scans from six cohorts: ADNI, AIBL, DELCODE, DESCRIBE, EDSD, and NIFD (N=3253), including participants that were cognitively normal, with amnestic mild cognitive impairment, dementia due to Alzheimers disease and frontotemporal dementia. Clustering analysis benchmarked different explanation space configurations by using morphological features as proxy-ground truth. We implemented three post-hoc explanations methods: i) by simplifying model decisions, ii) explanation-by-example, and iii) textual explanations. A qualitative evaluation by clinicians (N=6) was performed to assess their clinical validity. ResultsClustering performance improved in morphology enriched explanation spaces, improving both homogeneity and completeness of the clusters. Post hoc explanations by model simplification largely delineated converters and stable participants, while explanation-by-example presented possible cognition trajectories. Textual explanations gave rule-based summarization of pathological findings. Clinicians qualitative evaluation highlighted challenges and opportunities of XAI for different clinical applications. ConclusionOur study refines XAI explanation spaces and applies various approaches for generating explanations. Within the context of AI-based decision support system in dementia research we found the explanations methods to be promising towards enhancing diagnostic efficiency, backed up by the clinical assessments.

Diffusion models, while trained for image generation, have emerged as powerful foundational feature extractors for downstream tasks. We find that off-the-shelf diffusion models, trained exclusively to generate natural RGB images, can identify semantically meaningful correspondences in medical images. Building on this observation, we propose to leverage diffusion model features as a similarity measure to guide deformable image registration networks. We show that common intensity-based similarity losses often fail in challenging scenarios, such as when certain anatomies are visible in one image but absent in another, leading to anatomically inaccurate alignments. In contrast, our method identifies true semantic correspondences, aligning meaningful structures while disregarding those not present across images. We demonstrate superior performance of our approach on two tasks: multimodal 2D registration (DXA to X-Ray) and monomodal 3D registration (brain-extracted to non-brain-extracted MRI). Code: https://github.com/uncbiag/dgir

Compound figures, which are multi-panel composites containing diverse subfigures, are ubiquitous in biomedical literature, yet large-scale subfigure extraction remains largely unaddressed. Prior work on subfigure extraction has been limited in both dataset size and generalizability, leaving a critical open question: How does high-fidelity image-text alignment via large-scale subfigure extraction impact representation learning in vision-language models? We address this gap by introducing a scalable subfigure extraction pipeline based on transformer-based object detection, trained on a synthetic corpus of 500,000 compound figures, and achieving state-of-the-art performance on both ImageCLEF 2016 and synthetic benchmarks. Using this pipeline, we release OPEN-PMC-18M, a large-scale high quality biomedical vision-language dataset comprising 18 million clinically relevant subfigure-caption pairs spanning radiology, microscopy, and visible light photography. We train and evaluate vision-language models on our curated datasets and show improved performance across retrieval, zero-shot classification, and robustness benchmarks, outperforming existing baselines. We release our dataset, models, and code to support reproducible benchmarks and further study into biomedical vision-language modeling and representation learning.

Although existing semi-supervised image segmentation methods have achieved good performance, they cannot effectively utilize multiple sources of voxel-level uncertainty for targeted learning. Therefore, we propose two main improvements. First, we introduce a novel pignistic co-evidential fusion strategy using generalized evidential deep learning, extended by traditional D-S evidence theory, to obtain a more precise uncertainty measure for each voxel in medical samples. This assists the model in learning mixed labeled information and establishing semantic associations between labeled and unlabeled data. Second, we introduce the concept of information volume of mass function (IVUM) to evaluate the constructed evidence, implementing two evidential learning schemes. One optimizes evidential deep learning by combining the information volume of the mass function with original uncertainty measures. The other integrates the learning pattern based on the co-evidential fusion strategy, using IVUM to design a new optimization objective. Experiments on four datasets demonstrate the competitive performance of our method.

Multi-modal brain magnetic resonance imaging (MRI) plays a crucial role in clinical diagnostics by providing complementary information across different imaging modalities. However, a common challenge in clinical practice is missing MRI modalities. In this paper, we apply SwinUNETR to the synthesize of missing modalities in brain MRI. SwinUNETR is a novel neural network architecture designed for medical image analysis, integrating the strengths of Swin Transformer and convolutional neural networks (CNNs). The Swin Transformer, a variant of the Vision Transformer (ViT), incorporates hierarchical feature extraction and window-based self-attention mechanisms, enabling it to capture both local and global contextual information effectively. By combining the Swin Transformer with CNNs, SwinUNETR merges global context awareness with detailed spatial resolution. This hybrid approach addresses the challenges posed by the varying modality characteristics and complex brain structures, facilitating the generation of accurate and realistic synthetic images. We evaluate the performance of SwinUNETR on brain MRI datasets and demonstrate its superior capability in generating clinically valuable images. Our results show significant improvements in image quality, anatomical consistency, and diagnostic value.