Large frontier models such as GPT-5 and Gemini have demonstrated remarkable performance in a wide range of health application benchmarks. However, underneath the seemingly promising results lie salient growth areas, especially in cutting-edge frontiers such as multimodal reasoning. Here we systematically apply and integrate a series of adversarial stress tests to assess the robustness of flagship models and health benchmarks. Our study reveals prevalent brittleness in the presence of simple adversarial transformations: leading systems can guess the correct answer even with key inputs removed yet may get confused by the slightest prompt alterations while fabricating convincing but flawed reasoning traces. Using clinician-guided rubrics, we demonstrate that popular health benchmarks vary widely in what they truly measure. Our study reveals considerable gaps between benchmark performance and the robustness evidence needed to support claims about multimodal medical reasoning in health applications.

Self-supervised learning (SSL) faces a fundamental conflict between semantic understanding and image reconstruction. High-level semantic SSL (e.g., DINO) relies on global tokens that are forced to be location-invariant for augmentation alignment, a process that inherently discards the spatial coordinates required for reconstruction. Conversely, generative models preserves dense feature grids for reconstruction but fails to produce high-level abstractions. We introduce STELLAR, a framework that resolves this tension by factorizing visual features into a low-rank product of semantic concepts and their spatial distributions. This disentanglement allows us to perform DINO-style augmentation alignment on the semantic tokens while maintaining the precise spatial mapping in the localization matrix necessary for pixel-level reconstruction. We demonstrate that as few as 16 sparse tokens under this factorized form are sufficient to simultaneously support high-quality reconstruction (2.60 FID) and match the semantic performance of dense backbones (79.10% ImageNet accuracy). Our results highlight STELLAR as a versatile sparse representation that bridges the gap between discriminative and generative vision by strategically separating semantic identity from spatial geometry.

Vision-language pre-training, i.e., aligning images with paired text, is a powerful paradigm to create encoders that can be directly used for tasks such as classification, retrieval, and segmentation. In the 3D medical image domain, these capabilities allow vision-language encoders (VLEs) to support radiologists by retrieving patients with similar abnormalities, predicting likelihoods of abnormality, or, with downstream adaptation, generating radiological reports. While the methodology holds promise, data availability and domain-specific hurdles limit the capabilities of current 3D VLEs. In this paper, we overcome these challenges by injecting additional supervision via a report generation objective and combining vision-language with vision-only pre-training. This allows us to leverage both image-only and paired image-text 3D datasets, increasing the total amount of data to which our model is exposed. Through these additional objectives, paired with best practices of the 3D medical imaging domain, we develop the Comprehensive Language-Image Pre-training (COLIPRI) encoder family. Our COLIPRI encoders achieve state-of-the-art performance in report generation, semantic segmentation, classification probing, and zero-shot classification. The model is available at https://huggingface.co/microsoft/colipri.

The increasing use of medical imaging in healthcare settings presents a significant challenge due to the additional workload for radiologists, yet it also offers opportunity for enhancing healthcare outcomes if effectively leveraged. Artificial Intelligence (AI)-based 3D medical image retrieval holds the potential to alleviate radiologists' burden by offering evidence-based diagnostics and predictions that can enhance the scale and accuracy of radiologists' work, while simultaneously supporting output verification for safety and regulatory compliance. Despite its promise, the field of 3D medical image retrieval lacks established evaluation benchmarks, comprehensive datasets, and rigorous evaluation studies. This paper aims to address these gaps by introducing the first benchmark for 3D Medical Image Retrieval (3D-MIR) and evaluating various pre-trained models and implementation approaches for retrieval. The benchmark includes four anatomies (Liver, Colon, Pancreas, and Lung) imaged using computed tomography (CT). A range of 3D image search strategies are explored, including those that use aggregated 2D slices/3D volumes (Image-to-Image) and text embeddings from popular foundation models as queries (Text-to-Image). Additionally, novel multi-modal and supervised fine-tuning approaches are investigated to generate multi-modal embeddings for 3D image retrieval. The paper provides quantitative and qualitative assessments of each approach, along with an in-depth discussion offering insights for future research and solutions to support clinical decision-making and healthcare applications. To foster advancement in this field, our benchmark, models, and code are made publicly available.

Automated radiology report generation is advancing rapidly in response to the shortage of radiologists, yet unlike a perception model, existing generation models offer no control over the sensitivity-specificity trade-off of their diagnostic content. Such control is essential because clinical scenarios diverge: emergency triage prioritizes sensitivity to reduce missed findings, whereas confirmatory interpretation emphasizes specificity to limit unnecessary interventions. A single fixed report can neither adapt to these scenarios nor support the ROC-based validation widely expected for regulatory clearance. We introduce RadFusion, a framework that equips report generation with threshold controllability. Our method fuses a multi-label classifier, which provides per-disease confidence scores, with a VQA-based report generator, which describes medical findings in detail; an LLM then rewrites the report so that its stated diagnoses follow the classifier's decisions at the selected threshold while staying grounded in the generator's descriptions. On MIMIC-CXR, the performance of RadFusion conforms to the classifier's ROC curve: sweeping the threshold and mapping the reports back to class labels reproduces the classifier's validated ROC performance. This conformance makes generated reports quantitatively evaluable through ROC analysis, strengthening the case for regulatory clearance, and enables operating-point selection that matches report behavior to clinical context. Moreover, combining the two model types improves diagnostic accuracy over uncontrolled generation: sensitivity increases by 6.9

The limited availability of 3D medical image datasets, due to privacy concerns and high collection or annotation costs, poses significant challenges in the field of medical imaging. There are few solutions for realistic 3D medical image synthesis due to difficulties in backbone design and fewer 3D training samples compared to 2D counterparts. In this paper, we propose GEM-3D, a novel generative approach to the synthesis of 3D medical images and the enhancement of existing datasets using conditional diffusion models. Our method begins with a 2D slice, noted as the informed slice to serve the patient prior, and propagates the generation process using a 3D segmentation mask. By decomposing the 3D medical images into editable masks and patient prior information, GEM-3D offers a flexible yet effective solution for generating versatile 3D images from existing datasets. Moreover, as the informed slice contains patient-wise information, GEM-3D can also facilitate counterfactual image synthesis and dataset-level de-enhancement with desired control. Experiments on brain MRI and abdomen CT images demonstrate that GEM-3D is capable of synthesizing high-quality 3D medical images with volumetric consistency, offering a straightforward solution for dataset enhancement during inference. The code is available at https://github.com/HKU-MedAI/GEM-3D .

Electronic health records (EHRs) have improved data accessibility but have also introduced cognitive burden for physicians, given the sheer volume and complexity of the data involved. Advances in large language models (LLMs) create new opportunities to rethink how clinicians interact with medical data through dynamic, adaptive interfaces. In this position paper, we explore how generative AI can support physicians' information needs by enabling more dynamic interactions with patient data. Through semi-structured interviews with internal physicians at Microsoft, we identify key challenges in data navigation and synthesis, and characterize clinicians' information needs during diagnostic workflows. We further examine how physicians conceptualize AI can help their work process and how these mental models shape expectations for interaction and trust. Based on these insights, we discuss design considerations for generative user interfaces that support clinician-centered workflows.

The National Comprehensive Cancer Network (NCCN) provides evidence-based guidelines for cancer treatment. Translating complex patient presentations into guideline-compliant treatment recommendations is time-intensive, requires specialized expertise, and is prone to error. Advances in large language model (LLM) capabilities promise to reduce the time required to generate treatment recommendations and improve accuracy. We present an LLM agent-based approach to automatically generate guideline-concordant treatment trajectories for patients with non-small cell lung cancer (NSCLC). Our contributions are threefold. First, we construct a novel longitudinal dataset of 121 cases of NSCLC patients that includes clinical encounters, diagnostic results, and medical histories, each expertly annotated with the corresponding NCCN guideline trajectories by board-certified oncologists. Second, we demonstrate that existing LLMs possess domain-specific knowledge that enables high-quality proxy benchmark generation for both model development and evaluation, achieving strong correlation (Spearman coefficient r=0.88, RMSE = 0.08) with expert-annotated benchmarks. Third, we develop a hybrid approach combining expensive human annotations with model consistency information to create both the agent framework that predicts the relevant guidelines for a patient, as well as a meta-classifier that verifies prediction accuracy with calibrated confidence scores for treatment recommendations (AUROC=0.800), a critical capability for communicating the accuracy of outputs, custom-tailoring tradeoffs in performance, and supporting regulatory compliance. This work establishes a framework for clinically viable LLM-based guideline adherence systems that balance accuracy, interpretability, and regulatory requirements while reducing annotation costs, providing a scalable pathway toward automated clinical decision support.

Foundation models, pre-trained on extensive datasets, have significantly advanced machine learning by providing robust and transferable embeddings applicable to various domains, including medical imaging diagnostics. This study evaluates the utility of embeddings derived from both general-purpose and medical domain-specific foundation models for training lightweight adapter models in multi-class radiography classification, focusing specifically on tube placement assessment and related findings, with comparison to the end-to-end training of an established convolutional neural network. A dataset comprising 8842 radiographs classified into seven distinct categories was employed to extract embeddings using seven foundation models: DenseNet121, BiomedCLIP, Med-Flamingo, MedImageInsight, MedSigLIP, Rad-DINO, and CXR-Foundation. Adapter models were subsequently trained using classical machine learning algorithms, including K-nearest neighbors (KNN), logistic regression (LR), support vector machines (SVM), random forest (RF), and multi-layer perceptron (MLP). Among these combinations, MedImageInsight embeddings paired with an SVM or MLP adapter yielded the highest mean area under the curve (mAUC) at 93.1%, followed closely by MedSigLIP with MLP (91.0%), Rad-DINO with SVM (90.7%), and CXR-Foundation with LR (88.6%), achieving a higher mAUC score than a fully finetuned convolutional neural network, DenseNet121 (87.2%). In comparison, BiomedCLIP and DenseNet121 exhibited moderate performance with SVM, obtaining mAUC scores of 82.8% and 81.1%, respectively, whereas Med-Flamingo delivered the lowest performance at 78.5% when combined with RF. Significant differences were found between each embedding model and MedImageInsight using the Wilcoxon signed-rank test at the significance level 0.05 (before Bonferroni correction). Notably, most adapter models demonstrated computational efficiency, achieving training within minutes and inference within seconds on CPU, underscoring their practicality for clinical applications. Furthermore, fairness analysis on adapters trained on MedImageInsight-derived embeddings indicated minimal disparities, with gender differences in performance within 1.8% and standard deviations across age groups not exceeding 1.4%. Further analysis indicated there is no significant difference across gender and age at a significance level of 0.05. These findings confirm that foundation model embeddings-especially those from MedImageInsight-facilitate accurate, computationally efficient, and equitable diagnostic classification using lightweight adapters for radiographic image analysis.

Language-supervised pretraining has proven to be a valuable method for extracting semantically meaningful features from images, serving as a foundational element in multimodal systems within the computer vision and medical imaging domains. However, the computed features are limited by the information contained in the text, which is particularly problematic in medical imaging, in which the findings described by radiologists focus on specific observations. This challenge is compounded by the scarcity of paired imaging–text data due to concerns over the leakage of personal health information. In this work, we fundamentally challenge the prevailing reliance on language supervision for learning general-purpose biomedical imaging encoders. We introduce RAD-DINO, a biomedical image encoder pretrained solely on unimodal biomedical imaging data that obtains similar or greater performance than state-of-the-art biomedical-language-supervised models on a diverse range of benchmarks. Specifically, the quality of learned representations is evaluated on standard imaging tasks (classification and semantic segmentation), and a vision–language alignment task (text report generation from images). To further demonstrate the drawback of language supervision, we show that features from RAD-DINO correlate with other medical records (for example, sex or age) better than language-supervised models, which are generally not mentioned in radiology reports. Finally, we conduct a series of ablations determining the factors in RAD-DINO’s performance. In particular, we observe that RAD-DINO’s downstream performance scales well with the quantity and diversity of training data, demonstrating that image-only supervision is a scalable approach for training a foundational biomedical image encoder. Reliance on text supervision for biomedical image encoders is investigated. The proposed RAD-DINO, pretrained solely on unimodal data, achieves similar or greater performance than state-of-the-art multimodal models on various benchmarks.

Careful selection of skin lesions that require expert evaluation is important for early skin cancer detection. Yet challenges include lack of cost-effective asymptomatic screening, geographical inequality in access to specialty dermatology, and long wait times due to exam inefficiencies and staff shortages. Machine learning models trained on high-quality dermoscopy photos have been shown to aid clinicians in diagnosing individual, hand-selected skin lesions. In contrast, models designed for triage have been less explored due to limited datasets that represent a broader net of skin lesions. 3D total body photography is an emerging technology used in dermatology to document all apparent skin lesions on a patient for skin cancer monitoring. A multi-institutional and global project collected over 900,000 lesion crops off 3D total body photos for an online grand challenge in machine learning. Here we summarize the results of the competition, ‘ISIC 2024 – Skin Cancer Detection with 3D-TBP’, demonstrate superiority of a model that utilized intra-patient context against a prior published approach, and explore clinical plausibility of automated atypical skin lesion triage through an ablation study.

Large language models have demonstrated remarkable performance in a wide range of medical benchmarks. Yet underneath the seemingly promising results lie salient growth areas, especially in cutting-edge frontiers such as multimodal reasoning. In this paper, we introduce a series of adversarial stress tests to systematically assess the robustness of flagship models and medical benchmarks. Our study reveals prevalent brittleness in the presence of simple adversarial transformations: leading systems can guess the right answer even with key inputs removed, yet may get confused by the slightest prompt alterations, while fabricating convincing yet flawed reasoning traces. Using clinician-guided rubrics, we demonstrate that popular medical benchmarks vary widely in what they truly measure. Our study reveals significant competency gaps of frontier AI in attaining real-world readiness for health applications. If we want AI to earn trust in healthcare, we must demand more than leaderboard wins and must hold AI systems accountable to ensure robustness, sound reasoning, and alignment with real medical demands.

AI-assisted report generation offers the opportunity to reduce radiologists' workload stemming from expanded screening guidelines, complex cases and workforce shortages, while maintaining diagnostic accuracy. In addition to describing pathological findings in chest X-ray reports, interpreting lines and tubes (L T) is demanding and repetitive for radiologists, especially with high patient volumes. We introduce MAIRA-X, a clinically evaluated multimodal AI model for longitudinal chest X-ray (CXR) report generation, that encompasses both clinical findings and L T reporting. Developed using a large-scale, multi-site, longitudinal dataset of 3.1 million studies (comprising 6 million images from 806k patients) from Mayo Clinic, MAIRA-X was evaluated on three holdout datasets and the public MIMIC-CXR dataset, where it significantly improved AI-generated reports over the state of the art on lexical quality, clinical correctness, and L T-related elements. A novel L T-specific metrics framework was developed to assess accuracy in reporting attributes such as type, longitudinal change and placement. A first-of-its-kind retrospective user evaluation study was conducted with nine radiologists of varying experience, who blindly reviewed 600 studies from distinct subjects. The user study found comparable rates of critical errors (3.0

This paper presents an overview of the ImageCLEF 2025 lab, which was organized within the Conference and Labs of the Evaluation Forum – CLEF Labs 2025. ImageCLEF is an ongoing evaluation event that started in 2003, promoting the evaluation of technologies for annotation, indexing, and retrieval of multimodal data and aiming to provide access to large collections of data across a veriety of scenarios, domains and contexts. In 2025, the 23rd edition of ImageCLEF consists of four main tasks: (i) the Medical task, comprised of four sub-tasks, approaching a wide array of problems in the medical field, like concept detection, caption prediction, explainability assessment in radiology images, evaluating the veracity of GAN-generated 3D CT scans, providing a segmentation and answers to close-ended questions regarding dermatology images, or visual question answering and synthetic image generation involving gastrointestinal images, (ii) a new Multimodal Reasoning task, involving answering multiple-choice questions in 13 different languages, covering a wide range of subjects and difficulty levels, (iii) the ToPicto task, which focuses on converting either text or speech into a meaningful sequence of pictograms and (iv) the Argument-Image task, which explores the augmentation of arguments using images, by either retrieval or synthetic generation. This edition of the ImageCLEF benchmark attracted 193 teams that registered to the different tasks, of which 56 finished the challenges. This resulted in 493 submitted runs and a total of 45 working note papers. Overall, this year’s edition has been very successful, with the biggest number of teams, submissions and working notes papers since 2019.

Foundation vision encoders such as CLIP and DINOv2, trained on web-scale data, exhibit strong transfer performance across tasks and datasets. However, medical imaging foundation models remain constrained by smaller datasets, limiting our understanding of how data scale and pretraining paradigms affect performance in this setting. In this work, we systematically study continual pretraining of two vision encoders, MedImageInsight (MI2) and RAD-DINO representing the two major encoder paradigms CLIP and DINOv2, on up to 3.5M chest x-rays from a single institution, holding compute and evaluation protocols constant. We evaluate on classification (radiology findings, lines and tubes), segmentation (lines and tubes), and radiology report generation. While prior work has primarily focused on tasks related to radiology findings, we include lines and tubes tasks to counterbalance this bias and evaluate a model's ability to extract features that preserve continuity along elongated structures. Our experiments show that MI2 scales more effectively for finding-related tasks, while RAD-DINO is stronger on tube-related tasks. Surprisingly, continually pretraining MI2 with both reports and structured labels using UniCL improves performance, underscoring the value of structured supervision at scale. We further show that for some tasks, as few as 30k in-domain samples are sufficient to surpass open-weights foundation models. These results highlight the utility of center-specific continual pretraining, enabling medical institutions to derive significant performance gains by utilizing in-domain data.

Molecular Tumor Boards (MTBs) are multidisciplinary forums where oncology specialists collaboratively assess complex patient cases to determine optimal treatment strategies. A central element of this process is the patient summary, typically compiled by a medical oncologist, radiation oncologist, or surgeon, or their trained medical assistant, who distills heterogeneous medical records into a concise narrative to facilitate discussion. This manual approach is often labor-intensive, subjective, and prone to omissions of critical information. To address these limitations, we introduce the Healthcare Agent Orchestrator (HAO), a Large Language Model (LLM)-driven AI agent that coordinates a multi-agent clinical workflow to generate accurate and comprehensive patient summaries for MTBs. Evaluating predicted patient summaries against ground truth presents additional challenges due to stylistic variation, ordering, synonym usage, and phrasing differences, which complicate the measurement of both succinctness and completeness. To overcome these evaluation hurdles, we propose TBFact, a “model-as-a-judge” framework designed to assess the comprehensiveness and succinctness of generated summaries. Using a benchmark dataset derived from de-identified tumor board discussions, we applied TBFact to evaluate our Patient History agent. Results show that the agent captured 94

In this work, we present MedImageInsight, an open-source medical imaging embedding model. MedImageInsight is trained on medical images with associated text and labels across a diverse collection of domains, including X-Ray, CT, MRI, dermoscopy, OCT, fundus photography, ultrasound, histopathology, and mammography. Rigorous evaluations demonstrate MedImageInsight's ability to achieve state-of-the-art (SOTA) or human expert level performance across classification, image-image search, and fine-tuning tasks. Specifically, on public datasets, MedImageInsight achieves SOTA in CT 3D medical image retrieval, as well as SOTA in disease classification and search for chest X-ray, dermatology, and OCT imaging. Furthermore, MedImageInsight achieves human expert performance in bone age estimation (on both public and partner data), as well as AUC above 0.9 in most other domains. When paired with a text decoder, MedImageInsight achieves near SOTA level single image report findings generation with less than 10% the parameters of other models. Compared to fine-tuning GPT-4o with only MIMIC-CXR data for the same task, MedImageInsight outperforms in clinical metrics, but underperforms on lexical metrics where GPT-4o sets a new SOTA. Importantly for regulatory purposes, MedImageInsight can generate ROC curves, adjust sensitivity and specificity based on clinical need, and provide evidence-based decision support through image-image search (which can also enable retrieval augmented generation). In an independent clinical evaluation of image-image search in chest X-ray, MedImageInsight outperformed every other publicly available foundation model evaluated by large margins (over 6 points AUC), and significantly outperformed other models in terms of AI fairness (across age and gender). We hope releasing MedImageInsight will help enhance collective progress in medical imaging AI research and development.

The convergence of text, visual, and audio data is a key step towards human-like artificial intelligence, however the current Vision-Language-Speech landscape is dominated by encoder-only models which lack generative abilities. We propose closing this gap with i-Code V2, the first model capable of generating natural language from any combination of Vision, Language, and Speech data. i-Code V2 is an integrative system that leverages state-of-the-art single-modality encoders, combining their outputs with a new modality-fusing encoder in order to flexibly project combinations of modalities into a shared representational space. Next, language tokens are generated from these representations via an autoregressive decoder. The whole framework is pretrained end-to-end on a large collection of dual- and single-modality datasets using a novel text completion objective that can be generalized across arbitrary combinations of modalities. i-Code V2 matches or outperforms state-of-the-art single- and dual-modality baselines on 7 multimodal tasks, demonstrating the power of generative multimodal pretraining across a diversity of tasks and signals.

Visual Question Answering (VQA) entails answering questions about images. We introduce the first VQA dataset in which all contents originate from an authentic use case. Sourced from online question answering community forums, we call it VQAonline. We characterize this dataset and how it relates to eight mainstream VQA datasets. Observing that answers in our dataset tend to be much longer (i.e., a mean of 173 words) and so incompatible with standard VQA evaluation metrics, we instead utilize popular metrics for longer text evaluation for evaluating six state-of-the-art VQA models on VQAonline and report where they struggle most. Finally, we analyze which evaluation metrics align best with human judgments. We publicly-share the dataset at: https://vqaonline.github.io/.

Advancements in dermatological artificial intelligence research require high-quality and comprehensive datasets that mirror real-world clinical scenarios. We introduce a collection of 18,946 dermoscopic images spanning from 2010 to 2016, collated at the Hospital Clínic in Barcelona, Spain. The BCN20000 dataset aims to address the problem of unconstrained classification of dermoscopic images of skin cancer, including lesions in hard-to-diagnose locations such as those found in nails and mucosa, large lesions which do not fit in the aperture of the dermoscopy device, and hypo-pigmented lesions. Our dataset covers eight key diagnostic categories in dermoscopy, providing a diverse range of lesions for artificial intelligence model training. Furthermore, a ninth out-of-distribution (OOD) class is also present on the test set, comprised of lesions which could not be distinctively classified as any of the others. By providing a comprehensive collection of varied images, BCN20000 helps bridge the gap between the training data for machine learning models and the day-to-day practice of medical practitioners. Additionally, we present a set of baseline classifiers based on state-of-the-art neural networks, which can be extended by other researchers for further experimentation.