Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias and hallucination. Validated against board-certified clinicians, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 state-of-the-art LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability—the ‘benchmarking gap’. Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70%, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive bias priming altered recommendations in 81% of fairness tests and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants and broader clinical workflows. Using a suite of adversarial agents designed to test large language models in interactive conversations in a range of safety-related metrics, this study shows strong gaps between static performance and reliability in multi-turn conversations and high failure rates with respect to privacy, fairness and accuracy.

Tool-augmented large language model (LLM) agents can orchestrate specialist classifiers, segmentation models, and visual question-answering modules to interpret chest X-rays. However, these agents still solve each case in isolation: they fail to accumulate experience across cases, correct recurrent reasoning mistakes, or adapt their tool-use behavior without expensive reinforcement learning. While a radiologist naturally improves with every case, current agents remain static. In this work, we propose Evo-MedAgent, a self-evolving memory module that equips a medical agent with the capacity for inter-case learning at test time. Our memory comprises three complementary stores: (1) Retrospective Clinical Episodes that retrieve problem-solving experiences from similar past cases, (2) an Adaptive Procedural Heuristics bank curating priority-tagged diagnostic rules that evolves via reflection, much like a physician refining their internal criteria, and (3) a Tool Reliability Controller that tracks per-tool trustworthiness. On ChestAgentBench, Evo-MedAgent raises multiple-choice question (MCQ) accuracy from 0.68 to 0.79 on GPT-5-mini, and from 0.76 to 0.87 on Gemini-3 Flash. With a strong base model, evolving memory improves performance more effectively than orchestrating external tools on qualitative diagnostic tasks. Because Evo-MedAgent requires no training, its per-case overhead is bounded by one additional retrieval pass and a single reflection call, making it deployable on top of any frozen model.

Purpose: To quantify how evaluation annotations influence measured pulmonary embolism (PE) segmentation performance relative to model training changes, and to establish a human-referenced framework. Materials and Methods: This retrospective study screened 166 voxel-annotated CT pulmonary angiography cases from CADPE (n=91), FUMPE (n=35), and READ (n=40); 149 were included. A primary rater annotated PE by protocol, and a senior thoracic radiologist reviewed and revised all segmentations. Three additional raters at three centers annotated a 15-case subset. The label effect was measured by evaluating two pretrained nnU-Net models (nnU-Net-A, nnU-Net-B) against original and refined annotations. The model effect was measured by comparing the same architecture trained on different dataset combinations with annotations fixed. The benchmark model (nnPE) was trained with leave-one-dataset-out and pooled five-fold cross-validation. Four metric categories were analyzed with case-paired Wilcoxon signed-rank tests, Benjamini-Hochberg correction, and bootstrap 95 Results: Changing only the annotation increased mean DSC by 0.143 (0.122-0.166) for nnU-Net-A and 0.188 (0.163-0.213) for nnU-Net-B (both P < .001), whereas changing training-dataset composition changed DSC by 0.028. The label effect exceeded the model effect on CADPE and FUMPE and was 0.045 on READ. Within-mask attenuation SD fell in all three datasets after re-annotation (all P < .001). nnPE reached DSC 0.72 +/- 0.22 on pooled cross-validation but scored below all four annotators across 52 paired comparisons (all corrected P < .05). Conclusion: Evaluation annotations affected measured PE segmentation performance at least as much as model training choices. A human-referenced evaluation framework is publicly available for future study.

Clinical abnormality grounding for rare diseases is often hindered by data scarcity, rendering supervised fine-tuning infeasible and single-pass inference highly unstable. Thus, we propose Dynamic Decision Learning (DDL), a framework that enables frozen LVLMs to refine their decisions across language and visual spaces by optimizing instructions and consolidating predictions under visual perturbations, thereby improving localization quality and producing a consensus‑based reliability score that quantifies the model’s confidence. Results on brain‑imaging benchmarks, including a rare‑disease dataset with 281 pathology types across 3B-72B models, show that DDL improves mAP@75 by up to 105\% on rare‑disease cases and surpasses adaptation baselines and supervised fine‑tuning. Moreover, we show that DDL yields stronger calibration between consensus‑based reliability scores and localization accuracy under severe distribution shifts and increasing task difficulty. The code will be open-sourced.

While contrast-enhanced CT (CECT) is standard for assessing abdominal aortic aneurysms (AAA) , the required iodinated contrast agents pose significant risks, including patient allergies and environmental harm. To reduce contrast agent use, methods have focused on generating synthetic CECT from non-contrast CT (NCCT) scans. However, most adopt a multi-stage pipeline that first generates images and then performs segmentation, which leads to error accumulation and fails to leverage shared information. To address this, we propose a unified framework that generates synthetic CECT images from NCCT scans while simultaneously segmenting the aortic lumen and thrombus. Our approach integrates conditional diffusion models (CDM) with multi-task learning, enabling end-to-end joint optimization of image synthesis and anatomical segmentation. Unlike previous multitask diffusion models, our approach requires no initial predictions (e.g., a coarse segmentation mask), shares both encoder and decoder parameters across tasks, and employs a semi-supervised training strategy to learn from scans with missing segmentation labels, a common constraint in clinical data. Evaluated on a cohort of 264 patients, our method consistently outperformed state-of-the-art single-task and multi-stage models. For image synthesis, it achieved a PSNR of 25.61 dB, compared to 23.80 dB from a single-task CDM. For segmentation, it improved the lumen Dice score to 0.89 from 0.87 and the challenging thrombus Dice score to 0.53 from 0.48 (nnU-Net). These segmentation enhancements led to more accurate clinical measurements, reducing the lumen diameter MAE to 4.19 mm from 5.78 mm and the thrombus area error to 33.85% from 41.45%. Code is at https://github.com/yuxuanou623/AortaDiff.

Intensive care units (ICU) generate long, dense and evolving streams of clinical information, where physicians must repeatedly reassess patient states under time pressure, underscoring a clear need for reliable AI decision support. Existing ICU benchmarks typically treat historical clinician actions as ground truth. However, these actions are made under incomplete information and limited temporal context of the underlying patient state, and may therefore be suboptimal, making it difficult to assess the true reasoning capabilities of AI systems. We introduce RealICU, a hindsight-annotated benchmark for evaluating large language models (LLMs) under realistic ICU conditions, where labels are created after senior physicians review the full patient trajectory. We formulate four physician-motivated tasks: assess Patient Status, Acute Problems, Recommended Actions, and Red Flag actions that risk unsafe outcomes. We partition each trajectory with 30-min windows and release two datasets: RealICU-Gold with 930-window annotations from 94 MIMIC-IV patients, and RealICU-Scale with 11,862 windows extended by Oracle, a physician-validated LLM hindsight labeler. Existing LLMs including memory-augmented ones performed poorly on RealICU, exposing two failure modes: a recall-safety tradeoff for clinical recommendations, and an anchoring bias to early interpretations of the patient. We further introduce ICU-Evo to study structured-memory agents that improves long-horizon reasoning but does not fully eliminate safety failures. Together, RealICU provides a clinically grounded testbed for measuring and improving AI sequential decision-support in high-stakes care. Project page: https://chengzhi-leo.github.io/RealICU-Bench/

LLM agents acting in structured environments fail in operational rather than conversational ways, and reliability depends on procedural knowledge of the environment. Prior self-improvement methods accumulate natural-language guidance without checking that each new item preserves previously correct behavior, so a note that fixes one trajectory can silently regress another. We introduce GRASP (Gated Regression-Aware Skill Proposer), which treats agent improvement as a sequence of edits to a bounded skill library, admitting each candidate only if it produces a net improvement on a balanced held-out probe under a hard regression budget. We evaluate GRASP across five base models (gpt-oss-120b, DeepSeek V4 Flash, Gemini 3.1 Flash Lite, GPT-4.1, GPT-5.4) on two FHIR-based clinical benchmarks. On MedAgentBench, GRASP lifts gpt-oss-120b from 40.6

State-of-the-art large language models (LLMs) show high performance in general visual question answering. However, a fundamental limitation remains: current architectures lack the native 3D spatial reasoning required for direct analysis of volumetric medical imaging, such as CT or MRI. Emerging agentic AI offers a new solution, eliminating the need for intrinsic 3D processing by enabling LLMs to orchestrate and leverage specialized external tools. Yet, the feasibility of such agentic frameworks in complex, multi-step radiological workflows remains underexplored. In this work, we present a training-free agentic pipeline for automated brain MRI analysis. Validating our methodology on several LLMs (GPT-5.1, Gemini 3 Pro, Claude Sonnet 4.5) with off-the-shelf domain-specific tools, our system autonomously executes complex end-to-end workflows, including preprocessing (skull stripping, registration), pathology segmentation (glioma, meningioma, metastases), and volumetric analysis. We evaluate our framework across increasingly complex radiological tasks, from single-scan segmentation and volumetric reporting to longitudinal response assessment requiring multi-timepoint comparisons. We analyze the impact of architectural design by comparing single-agent models against multi-agent "domain-expert" collaborations. Finally, to support rigorous evaluation of future agentic systems, we introduce and release a benchmark dataset of image-prompt-answer tuples derived from public BraTS data. Our results demonstrate that agentic AI can solve highly neuro-radiological image analysis tasks through tool use without the need for training or fine-tuning.

A key challenge for multi-modal brain MRI registration is that most training data consists of only T1-weighted scans, making it difficult to generalize to other less frequently acquired modalities. We show that purely randomized intensity augmentations, without any annotations or modality synthesis, transform T1-trained networks from complete failure to state-of-the-art performance in T1—T2-weighted registration. Further, enforcing transformation composition consistency across image triplets yields substantial accuracy gains. Together, these simple yet powerful strategies redefine how robust multi-modal registration can be achieved under minimal supervision and data assumptions.

The most plausible near-term role of medical LLMs is to assist rather than replace physicians, yet current evaluations often test isolated capabilities: clinical knowledge, EHR system interaction, or patient communication. Physician assistance instead requires coordinating these capabilities within the same interaction, where physicians issue underspecified requests, patients describe symptoms ambiguously, and EHR systems demand precise tool use. We introduce PhysAssistBench, a benchmark for interactive doctor-patient-EHR assistance. Built from real MIMIC-IV cases, PhysAssistBench uses a scalable pipeline to construct agentic patients: interactive, record-grounded agents that turn static EHR records into multi-turn clinical scenarios while preserving clinical factuality. PhysAssistBench provides a curated bilingual evaluation set of 1,296 manually reviewed and physician-validated turns. Experiments with leading LLMs show that current models remain unreliable in this setting, which exposes a key bottleneck for clinical LLMs: reliable assistance requires coordination across knowledge, communication, and systems, not isolated gains in any of them.

MLLMs are beginning to appear in clinical workflows, but their ability to perform complex medical reasoning remains unclear. We present Med-CMR, a fine-grained Medical Complex Multimodal Reasoning benchmark. Med-CMR distinguishes from existing counterparts by three core features: 1) Systematic capability decomposition, splitting medical multimodal reasoning into fine-grained visual understanding and multi-step reasoning to enable targeted evaluation; 2) Challenging task design, with visual understanding across three key dimensions (small-object detection, fine-detail discrimination, spatial understanding) and reasoning covering four clinically relevant scenarios (temporal prediction, causal reasoning, long-tail generalization, multi-source integration); 3) Broad, high-quality data coverage, comprising 20,653 Visual Question Answering (VQA) pairs spanning 11 organ systems and 12 imaging modalities, validated via a rigorous two-stage (human expert + model-assisted) review to ensure clinical authenticity. We evaluate 18 state-of-the-art MLLMs with Med-CMR, revealing GPT-5 as the top-performing commercial model: 57.81 accuracy on multiple-choice questions (MCQs) and a 48.70 open-ended score, outperforming Gemini 2.5 Pro (49.87 MCQ accuracy, 45.98 open-ended score) and leading open-source model Qwen3-VL-235B-A22B (49.34 MCQ accuracy, 42.62 open-ended score). However, specialized medical MLLMs do not reliably outperform strong general models, and long-tail generalization emerges as the dominant failure mode. Med-CMR thus provides a stress test for visual–reasoning integration and rare-case robustness in medical MLLMs, and a rigorous yardstick for future clinical systems.

Conventional clinical CMR pipelines rely on a sequential "reconstruct-then-analyze" paradigm, forcing an ill-posed intermediate step that introduces avoidable artifacts and information bottlenecks. This creates a fundamental mathematical paradox: it attempts to recover high-dimensional pixel arrays (i.e., images) from undersampled k-space, rather than directly extracting the low-dimensional physiological labels actually required for diagnosis. To unlock the direct diagnostic potential of k-space, we propose k-MTR (k-space Multi-Task Representation), a k-space representation learning framework that aligns undersampled k-space data and fully-sampled images into a shared semantic manifold. Leveraging a large-scale controlled simulation of 42,000 subjects, k-MTR forces the k-space encoder to restore anatomical information lost to undersampling directly within the latent space, bypassing the explicit inverse problem for downstream analysis. We demonstrate that this latent alignment enables the dense latent space embedded with high-level physiological semantics directly from undersampled frequencies. Across continuous phenotype regression, disease classification, and anatomical segmentation, k-MTR achieves highly competitive performance against state-of-the-art image-domain baselines. By showcasing that precise spatial geometries and multi-task features can be successfully recovered directly from the k-space representations, k-MTR provides a robust architectural blueprint for task-aware cardiac MRI workflows.

Missing input sequences are common in medical imaging data, posing a challenge for deep learning models reliant on complete input data. In this work, inspired by MultiMAE [2], we develop a masked autoencoder (MAE) paradigm for multi-modal, multi-task learning in 3D medical imaging with brain MRIs. Our method treats each MRI sequence as a separate input modality, leveraging a late-fusion-style transformer encoder to integrate multi-sequence information (“multi-modal”) and individual decoder streams for each modality for “multi-task” reconstruction. This pretraining strategy guides the model to learn rich representations per modality while also equipping it to handle missing inputs through cross-sequence reasoning. The result is a flexible and generalizable encoder for brain MRIs that infers missing sequences from available inputs and can be adapted to various downstream applications. We demonstrate the performance and robustness of our method against an MAE-ViT baseline in downstream segmentation and classification tasks, showing absolute improvement of 10.1 overall Dice score and 0.46 MCC over the baselines with missing input sequences. Our experiments demonstrate the strength of this pretraining strategy. The implementation is made available (The repository for our work: https://github.com/chris-beischl/multimae-for-brain-mri ).

Medical diagnosis is not a single prediction from a fully specified vignette. It is a sequential workup: clinicians decide what evidence to obtain, revise a differential diagnosis, and stop when the diagnosis is sufficiently supported. Most medical AI benchmarks instead reveal the relevant context upfront and score only the final answer, making unsupported correct guesses, premature closure, inefficient workups, and poor uncertainty updating invisible. We introduce DDX-TRACE, a physician-adjudicated benchmark for multimodal neuroradiology that evaluates diagnostic trajectories under hidden evidence over 211 challenging cases. Each case begins with limited clinical history; models request imaging studies in free form, receive matched image bundles when available, update a probabilistic differential diagnosis after each turn, and stop with a localized final diagnosis. Evaluating state-of-the-art VLMs, we find that final diagnosis scores can substantially misrepresent workup quality: models may guess plausible diagnoses without essential evidence, request useful studies but misinterpret raw images, or acquire evidence inefficiently while updating uncertainty poorly. Controlled evidence variants isolate bottlenecks in planning, visual evidence extraction, and downstream differential reasoning. DDX-TRACE shifts medical AI evaluation from final answers to evidence-supported diagnostic trajectories.

Medical imaging benchmarks often evaluate VLMs on pre-selected 2D images, slices, crops, or patches, making evaluation closer to visual recognition. Real clinical workflows impose a different burden: readers must search through complete studies, operate imaging software, navigate across slices and magnifications, and document visual evidence that can be audited. We argue that this evidence-producing workflow is a critical missing evaluation axis for medical imaging agents. To study it, we introduce MedFlowBench, a full-study benchmark for VLM agents, together with MedOpenClaw, a controlled and replayable runtime in which agents operate medical imaging viewers such as 3D Slicer and QuPath. In each episode, an agent inspects a complete radiology study or whole-slide pathology image, returns a task answer, and submits structured evidence, including key slices, coordinates, regions of interest, or lesion-state fields. This evidence is automatically checked against withheld masks, annotations, and labels. Across evaluated models, final answer-only scoring gives an overly optimistic picture: when answers must also be supported by correct evidence, performance drops substantially on complex workflows. We further find that adding image-analysis tools does not by itself solve the problem. Tools help when they make a complex procedure simple and reliable, but agents still struggle when they must choose inputs, manage viewer state, and verify intermediate outputs over multiple steps. MedFlowBench exposes whether medical imaging agents can produce auditable evidence from complete studies, rather than plausible answers from selected images.

Longitudinal brain analysis is essential for understanding healthy aging and identifying pathological deviations. Longitudinal registration of sequential brain MRI underpins such analyses. However, existing methods are limited by reliance on densely sampled time series, a trade-off between accuracy and temporal smoothness, and an inability to prospectively forecast future brain states. To overcome these challenges, we introduce TimeFlow, a learning-based framework for longitudinal brain MRI registration. TimeFlow uses a U-Net backbone with temporal conditioning to model neuroanatomy as a continuous function of age. Given only two scans from an individual, TimeFlow estimates accurate and temporally coherent deformation fields, enabling non-linear extrapolation to predict future brain states. This is achieved by our proposed inter-/extrapolation consistency constraints applied to both the deformation fields and deformed images. Remarkably, these constraints preserve temporal consistency and continuity without requiring explicit smoothness regularizers or densely sampled sequential data. Extensive experiments demonstrate that TimeFlow outperforms state-of-the-art methods in terms of both future timepoint forecasting and registration accuracy. Moreover, TimeFlow supports novel biological brain aging analyses by differentiating neurode-generative trajectories from normal aging without requiring segmentation, thereby eliminating the need for labor-intensive annotations and mitigating segmentation inconsistency. TimeFlow offers an accurate, data-efficient, and annotation-free framework for longitudinal analysis of brain aging and chronic diseases, capable of forecasting brain changes beyond the observed study period.

Cardiovascular diseases are the leading cause of death. Cardiac phenotypes (CPs), e.g., ejection fraction, are the gold standard for assessing cardiac health, but they are derived from cine cardiac magnetic resonance imaging (CMR), which is costly and requires high spatio-temporal resolution. Every magnetic resonance (MR) examination begins with rapid and coarse localizers for scan planning, which are discarded thereafter. Despite non-diagnostic image quality and lack of temporal information, localizers can provide valuable structural information rapidly. In addition to imaging, patient-level information, including demographics and lifestyle, influence the cardiac health assessment. Electrocardiograms (ECGs) are inexpensive, routinely ordered in clinical practice, and capture the temporal activity of the heart. Here, we introduce C-TRIP (Cardiac Tri-modal Representations for Imaging Phenotypes), a multi-modal framework that aligns localizer MRI, ECG signals, and tabular metadata to learn a robust latent space and predict CPs using localizer images as an opportunistic alternative to CMR. By combining these three modalities, we leverage cheap spatial and temporal information from localizers, and ECG, respectively while benefiting from patient-specific context provided by tabular data. Our pipeline consists of three stages. First, encoders are trained independently to learn uni-modal representations. The second stage fuses the pre-trained encoders to unify the latent space. The final stage uses the enriched representation space for CP prediction, with inference performed exclusively on localizer MRI. Proposed C-TRIP yields accurate functional CPs, and high correlations for structural CPs. Since localizers are inherently rapid and low-cost, our C-TRIP framework could enable better accessibility for CP estimation.

Cardiac Magnetic Resonance (CMR) imaging is a critical tool for diagnosing and managing cardiovascular disease, yet its utility is often limited by the sparse acquisition of 2D short-axis slices, resulting in incomplete volumetric information. Accurate 3D reconstruction from these sparse slices is essential for comprehensive cardiac assessment, but existing methods face challenges, including reliance on predefined interpolation schemes (e.g., linear or spherical), computational inefficiency, and dependence on additional semantic inputs such as segmentation labels or motion data. To address these limitations, we propose a novel Cardiac Latent Interpolation Diffusion (CaLID) framework that introduces three key innovations. First, we present a data-driven interpolation scheme based on diffusion models, which can capture complex, non-linear relationships between sparse slices and improves reconstruction accuracy. Second, we design a computationally efficient method that operates in the latent space and speeds up 3D whole-heart upsampling time by a factor of 24, reducing computational overhead compared to previous methods. Third, with only sparse 2D CMR images as input, our method achieves SOTA performance against baseline methods, eliminating the need for auxiliary input such as morphological guidance, thus simplifying workflows. We further extend our method to 2D+T data, enabling the effective modeling of spatiotemporal dynamics and ensuring temporal coherence. Extensive volumetric evaluations and downstream segmentation tasks demonstrate that CaLID achieves superior reconstruction quality and efficiency. By addressing the fundamental limitations of existing approaches, our framework advances the state of the art for spatio and spatiotemporal whole-heart reconstruction, offering a robust and clinically practical solution for cardiovascular imaging.

Motivation: CMR is the golden standard for cardiac diagnosis, and medical data annotation is time-consuming. Thus, screening techniques from unlabeled data can help streamline the cardiac diagnosis process. Goal: This work aims to enable cardiac function assessment from unlabeled cardiac MR images using an unsupervised approach with masked image modeling. Approach: Our model creates a robust latent space by reconstructing sparse 2D+T planes (SAX, 2CH, 3CH, and 4CH views) with 70% masking, which can be further disentangled into distinct cardiac temporal states. Results: t-SNE visualization and kNN clustering analysis confirm the association between latent space and cardiac phenotypes, highlighting strong temporal feature extraction. Impact: This method offers a scalable approach for cardiac screening by creating a latent space as well as distinct time-segment embeddings, enabling diverse preliminary analysis of cardiac function and potentially advancing research in cardiovascular disease applications.