Medical VLMs

Large-scale construction of medical vision-language model (VLM) benchmarks is increasingly feasible with richly annotated imaging datasets and large language models (LLMs), yet existing automation largely focuses on generating evaluation items within predefined benchmark specifications. We study the broader problem of automatically deriving the specification itself: what to evaluate, which annotations support each task, and how to translate this evidence into reliable evaluation items. We formulate benchmark construction as constrained compilation, in which the benchmark specification is progressively derived from evaluation requirements, heterogeneous annotations, and medical knowledge. Based on this formulation, we introduce MedBenchAgent, a multi-agent framework with a Benchmark Intermediate Representation (BIR) that encodes task definitions, evidence mappings, evaluation protocols, and item specifications across construction stages. MedBenchAgent separates planning, which derives and verifies the specification, from instantiation, which constructs and audits items under the locked specification. MedBenchAgent achieves a Task-Space F1 of 90.9%, outperforming direct task induction (79.2-80.0%) and prior-guided induction (85.1%); 994 of 1,000 sampled items from correctly identified tasks pass human audit. We further demonstrate portability to a specialized medical domain and evaluate twelve VLMs, revealing task- and setting-specific variation obscured by aggregate scores. These results establish constrained compilation as a scalable and auditable framework for medical VLM benchmark construction beyond question generation.

Medical vision-language models (VLMs) have shown increasing potential for radiological image interpretation. Medical VLMs encode radiological images into visual representations that capture both anatomical and phenotypic information for diagnosis. Existing approaches improve pathological phenotype representations through semantic-guided representation alignment. However, pathological phenotypes arise as lesion-specific visual changes superimposed on underlying normal anatomy. Such semantic alignment approaches fail to model the phenotype-specific increment relative to the corresponding normal anatomical representation. To address this gap, we propose \textbf{$Δ$Representation}, a visual phenotype representation learning framework based on counterfactual reasoning for medical VLMs. It comprises \textbf{BaseAnatomy}, a geometry-supervised representation learning module, and \textbf{$Δ$Phenotype}, a counterfactual incremental representation learning module. BaseAnatomy provides fine-grained geometric supervision through spatial relationships across and within anatomical structures. $Δ$Phenotype computes the representation increment between lesion representations and their corresponding normal anatomical representations, and supervises increments associated with the same phenotype to cluster in the representation space. Experiments on \textit{ReXGroundingCT} and \textit{LIDC-IDRI} demonstrate that $Δ$Representation effectively structures pathological phenotype representations and improves lesion grounding and phenotype characterization accuracy in medical VLMs. Code is available at https://anonymous.4open.science/r/deltarep-CF6D.

Autism spectrum disorder (ASD) is diagnosed through specialist observation of a child's social behavior, and access to that expertise is the bottleneck for early identification. Vision-language models (VLMs) describe a child's behavior from video well; the verdict drawn from the description is unstable: at temperature~0, across eight pipeline configurations on one backbone, 16--37% of clips change their predicted label between repeated runs, and the cause lies in the serving stack. We keep the VLM frozen and move the decision out of the model. Under our grounded perception constraints the VLM writes an event table of timestamped, glossary-labeled events that names the eliciting press and logs counter-evidence; a text-only stage age-calibrates the confidence of each row; a deterministic weight-of-evidence scorer sums it into an evidence total and stratifies it into a risk category, so every decision decomposes into named per-feature contributions and can be re-scored from the saved table. On 43 caregiver-recorded, protocol-free home free-play clips of preschool children, the pipeline reaches AUC $0.851 \pm 0.012$, 86.0% accuracy, and F$_1$ 71.8 over three runs. It labels 74.4% of clips correctly in every run (60.5% for the zero-shot baseline) and flags no typically developing clip in every run (9 of 31 at zero-shot). An ablation on the same backbone attributes the gain to the grounded perception constraints read through the deterministic scorer.

Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve. This is especially concerning in medicine, where new clinical evidence, updated guidelines, and new therapies can change established practice. Fine-tuning can update the model, but it requires access to model weights and additional training. Parameter-free methods avoid training, but they may overfit a fixed validation set, lack reliable domain knowledge, or lose visual details by saving experience only as text. To address these limitations, we present a model-agnostic framework that allows frozen LLMs and VLMs to learn from deployment experience through three forms of external expertise: a Skill that guides reasoning and tool use, a Knowledge Memory that stores reliable facts supported by earlier cases or trusted external evidence, and a Multimodal Knowledge Base that keeps visual examples and guides the model to relate each retrieved case to the current image. Instead of relying on a fixed validation set, a validation strategy keeps an update only if it helps on new cases without degrading performance on earlier ones. Across six benchmarks covering clinical diagnosis, clinical workflows, medical reasoning, and medical and non-medical visual reasoning, and with four open-weight and closed-source base models, our framework improves performance during online deployment by up to 34.2% over the base model on medical tasks, generalizes to unseen cases, transfers to other models without further optimization, and works in non-medical domains.

Clinical diagnosis is inherently a structured reasoning process, yet existing deep learning models often bypass this structure by mapping image features directly to disease labels without explicitly interrogating the morphological and textural criteria that clinicians systematically evaluate. This limits diagnostic transparency and may compromise safe clinical deployment. We present MedCORE (Medical Criteria-Oriented Reasoning and Evidence), a structured diagnostic framework that operationalizes clinical reasoning within a vision-language architecture. For each input image, MedCORE decomposes the diagnostic process into clinically defined criteria, spatially localizes each criterion to diagnostically relevant image regions, encodes evidence through multi-scale representations that capture macro-structural and micro-textural pathological characteristics, and refines criterion representations using a Graph Attention Network that explicitly models inter-criteria dependencies. Criterion representations are further aligned with clinical text descriptors, reinforced through class-wise visual prototypes, and aggregated using uncertainty-calibrated weighting that proportionally discounts low-confidence diagnostic evidence. MedCORE is validated across three clinically heterogeneous imaging modalities, including dermoscopic lesion classification on ISIC 2018, breast ultrasound lesion characterization on BUSI, and diabetic retinopathy grading on IDRiD. Quantitatively, MedCORE achieves 89.2% accuracy, 85.7% macro-F1, and 96.4% AUC on ISIC 2018; 96.1% accuracy, 95.2% macro-F1, and 98.4% AUC on BUSI; and 84.3% accuracy, 80.2% macro-F1, and 92.8% AUC on IDRiD. These results demonstrate consistent improvements over strong CNN, transformer, biomedical vision-language, concept-based, and prototype-based baselines.

Tactical Combat Casualty Care (TC3) requires responders to connect visual observations of injuries and interventions with established clinical guidance. Developing vision-language models to support this process requires supervision that links visible evidence to traceable doctrine. We present TC3-VQA, a dataset constructed from public instructional and field TC3 videos and authoritative TC3 documents. It contains 581 items spanning 11 concepts, with 1,860 questions covering intervention recognition, doctrine, clinical reasoning, procedural guidance, and refusal when visual information is insufficient. Doctrine-based answers preserve verbatim source passages and character offsets. Construction combines visual annotation, passage retrieval, entailment checks, and verification across model families. Equipment boxes, anatomical labels, temporal segments, and source metadata accompany the question-answer pairs. Automated audits and ratings by two physicians and two medical students characterize annotation quality, with human ratings available for 88 retained items. The dataset provides a resource for adapting vision-language models to TC3, studying the connection between visual evidence and clinical knowledge, and evaluating recognition, doctrine recall, and abstention.

Multimodal large language models (MLLMs) are rapidly advancing clinical diagnosis, yet their adaptation pipelines remain anchored to accuracy-based objectives. Clinical data are heavily class-imbalanced: a constant-majority predictor can score above 90% accuracy while being clinically useless. We therefore evaluate and optimize for AUROC, a threshold-free score that ranks positives above negatives and is invariant to class balance. We focus on prompt optimization in MLLMs. Reflective methods such as GEPA use a binary scores matrix with one row per evaluation instance and one column per candidate prompt; cells record per-instance correctness, so the column average is accuracy and drives candidate selection. We introduce pair-level Pareto prompt evolution (Ranking-PE), which replaces each correctness row with a pairwise-ordering row over (positive, negative) instance pairs: the cell is 1 if the candidate scores the positive higher than the paired negative. The column average then equals empirical AUROC (by the Wilcoxon-Mann-Whitney identity). We apply this swap at all three layers the prompt evolution search reads from - the scores matrix that decides Pareto dominance, the per-example feedback to the reflection LM, and final candidate selection - at no extra model calls and with no surrogate loss. Across three diseases on MIMIC, accuracy-based prompt evolution can degrade ranking; Ranking-PE reverses this, beating the accuracy-based recipe by +5.8 AUROC pp on fine-tuned Qwen3-VL-8B and +16.2 pp on MedGemma-4B. Ablations examine each design component and show that a medical-grade visual backbone - via vision-encoder-tuned SFT or medical pretraining - is a prerequisite that prompt search cannot replace - our recipe extends reflective prompt evolution from text-only data to multimodal clinical decision-making.

Automated report generation can ease the burden radiolo gists face when interpreting multi-sequence MRI studies. Unlike CT, MRI examinations comprise multiple sequences and imaging planes, each con tributing complementary diagnostic information. Existing methods en code a study as a single volume and combine multiple acquisitions by fixed rules. Findings visible in only one plane are thus diluted and of ten missed, lowering recall on clinical efficacy metrics, where a missed abnormality is most costly. We propose GateSPINE, a vision-language framework that fuses sagittal T1 and T2 volumes with a training-free operator, encodes the fused sagittal and axial volumes with two parallel 3D encoders, and decodes their combined representation into a report. Its core mechanism is a gated cross view fusion module that predicts, per feature channel and token, how much of each view to admit, so the more informative view dominates at each spatial location. We evaluate GateSPINE on three lumbar MRI datasets, comprising two public bench marks and a private cohort collected from Phenikaa University Hospital, using both natural language generation (NLG) and clinical efficacy (CE) metrics. GateSPINE achieves the highest CE F1 through improved re call on all three datasets; on SPIDER, which lacks an axial sequence, this reflects the sagittal fusion component rather than the gated cross-view mechanism, which is validated on the two cohorts with both imaging planes. GateSPINE also remains competitive on standard NLG metrics.

Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark built on UK Biobank, comprising 50,401 clinical questions derived from real-world ICD-10-coded diagnoses of 4,739 participants. Each question links to a patient-specific environment containing clinical context and multi-sequence MRI from baseline and follow-up visits, where agents autonomously select which visits, organs, modalities, slices, and specialist tools to inspect and compare. Supervised fine-tuning transfers evidence-seeking workflows from 14,734 frontier-model interaction trajectories, followed by agentic reinforcement learning on the learner's own environment interactions. Privileged on-policy self-distillation and rubric-based LLM feedback refine evidence-to-conclusion reasoning without prescribing tool sequences. Experiments show that CASE moves beyond question-answer imitation toward transferable investigation policies, strengthening evidence-grounded longitudinal reasoning. Under matched evaluation conditions, our Qwen3-VL-8B based agent achieves over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8. Code will be available at https://github.com/VinyehShaw/CASE.

Medical models are judged not only on correctness, but on whether reported confidence matches actual accuracy. Generalist multimodal medical models have expanded what a single model can perceive, yet they still express bounded decisions such as diagnoses, findings or cell counts as generated text, so the reported probability reflects the next token rather than the decision itself. Motivated by decision-native interfaces such as Jev, we introduce OmniMed-Jev, which represents each medical decision as a Choice, Noul or Score decision over a runtime-supplied candidate set and returns a full distribution over that set: mutually exclusive classes, binary presence of a finding, or a bounded ordered value. The design is omni in three respects: it accepts diverse imaging modalities, covers different prediction tasks, and expresses them through one candidate-conditioned probability model, so heterogeneous outputs become comparable probabilities rather than task-specific strings. In an interface-controlled comparison against a generative baseline trained on the same backbone, data and schedule, OmniMed-Jev's reported probabilities track observed correctness far more closely, reducing calibration error by up to an order of magnitude and reliability error by up to two, while point-prediction performance remains comparable; counting is the one family where the generative baseline stays ahead. Making the decision distribution the model's output is not a format change but what turns reported numbers into probabilities that mean what they say. These results support explicit decision modeling as a way to make reported confidence meaningful within the evaluated tasks, and they are not evidence of clinical readiness: the comparison cannot separate the interface from associated training differences, which we state alongside the results. Code is available at github.com/lytang63/OmniMed-Jev.

As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.

Medical multimodal large language models (MLLMs) are increasingly expected not only to answer clinical questions, but also to localize the visual evidence behind their predictions. A common strategy connects a vision--language model (VLM) with SAM-style segmentation through a special <SEG> token, yet full-parameter training of this unified architecture is difficult because image-level reasoning and pixel-level segmentation impose different requirements on the shared representation space. To address this issue, we propose a parameter-decoupled training framework for unified medical reasoning and segmentation. The framework treats the <SEG> hidden state as a semantic-to-spatial prompt for the mask decoder and encourages it to become separable from generic language states, reducing ambiguous segmentation prompts and potential disruption to reasoning representations. It first performs medical shallow alignment to adapt visual features to clinical language without disturbing the LLM; then controlled instruction tuning shapes separable <SEG> prompt states, monitored by the Davies--Bouldin Index (DBI), while scaling segmentation gradients entering the language backbone; finally, the SAM branch is specialized with the VLM frozen to improve mask precision without altering reasoning parameters. Experiments on medical referring segmentation, grounding, visual QA, and textual QA benchmarks show that our framework achieves strong language-conditioned segmentation while preserving competitive reasoning ability. Ablations show that two-phase instruction tuning, gradient scaling, and segmentation specialization all contribute to the model.

Medical Vision-Language Models (VLMs) show significant promise for clinical image understanding, offering accurate diagnosis with interpretable reasoning. However, a critical performance gap exists between their strong vision encoders and the full multimodal model: in dermatology, the MedSigLIP encoder outperforms MedGemma by an average of 10.26 percentage points even when both use zero target-task labels; few-shot linear probing provides further evidence of strong visual representations. This gap motivates an investigation of how visual information is used in end-to-end diagnosis and why plausible-sounding predictions can lack grounding in image evidence. Using dermatology as our primary testbed, we systematically investigate three hypotheses for this phenomenon. We further provide a mechanistic analysis of the model's internal attention patterns, showing that a simple describe-then-decide prompting strategy increases vision attention by 30-40% during generation. Task-specific fine-tuning improves dermatology classification but reduces cross-domain medical question-answering performance in our evaluation. To address these challenges, we combine label-free prompting with low-label encoder-assisted reranking while keeping the VLM frozen. We validate the interventions across five VLM backbones in dermatology and provide supporting representation and attention analyses across additional medical modalities.

Recent medical multimodal models have benefited from larger corpora, broader modality coverage, and stronger reasoning-oriented training, yet effective data design across continued pretraining (CPT) and post-training remains challenging. Medical sources vary substantially in structure, granularity, and information density, and their utility shifts as training progresses from broad knowledge acquisition to late-stage consolidation. Meanwhile, post-training is often dominated by short-form visual question answering, providing limited supervision for informative and answer-consistent explanations. We introduce InfiMed2, a family of 4B and 27B generalist medical multimodal foundation models built around stage-aware data design. We curate a 55.68B-token corpus that combines broad clinical knowledge with context-rich biomedical visual evidence through source-specific processing. Our CPT pipeline first adapts the vision encoder, then builds broad medical knowledge, and finally transitions to an evidence-focused data mixture during learning-rate decay. For supervised fine-tuning (SFT), we regenerate visual question-answering responses using answer stability, answer-masked reconstruction, and correctness-constrained selection to produce more informative and answer-consistent supervision. The 4B model is further optimized with reinforcement learning with verifiable rewards (RLVR). Across five medical multimodal benchmarks, InfiMed2-4B achieves 66.73% mean accuracy after RLVR, surpassing the larger Qwen3.5-9B, while InfiMed2-27B reaches 73.72%, the highest among the evaluated open-weight models.

Acne assessment is crucial for clinical decision-making, yet traditional grading and counting are subjective and fail to account for lesion size. While area-based assessment has emerged as a promising alternative, acne segmentation has continued to rely on general-purpose architectures. To address this gap, we propose VL-AcneSeg, a multimodal framework for acne lesion segmentation that leverages CLIP and region-level text prompts to incorporate spatial priors, enabling lesions to be localized across the whole face. Because region-level prompts indicate which facial areas contain lesions, we report a single global prompt, which requires no such information, as our primary setting. On our internal clinical dataset, VL-AcneSeg achieves a Dice score of 0.5082 and an IoU of 0.3407 under this protocol, the highest among all compared methods, including recent vision-language segmentation methods that are themselves given region-level prompts; region-level prompting raises these to 0.5296 and 0.3602. Moreover, lesion area measurements derived from our segmentation correlate with IGA scores at a level comparable to expert annotations (Pearson r = 0.719 versus 0.658). Notably, our framework maintains consistent performance across external validation datasets, performing reliably even on uncontrolled smartphone images without requiring additional training or fine-tuning. By pairing a protocol that requires no lesion-location information with area-based severity estimation, this work provides a foundation for objective acne assessment outside the clinic. Our implementation is publicly available at: https://github.com/sukjuoh/VL-AcneSeg

Medical vision-language pre-training largely follows a visit-level image-report matching paradigm, aligning paired images and reports at individual visits. While effective for static cross-modal correspondence, this paradigm provides limited supervision for longitudinal clinical change, such as whether abnormalities improve, remain stable, or worsen over time. Learning such change is challenging because temporal semantics are implicit in free-text reports, and different abnormalities within the same patient may evolve asynchronously or even in opposite directions. We propose MedCTA, which reframes medical vision-language pre-training from visit-level cross-modal matching to learning clinical change. Rather than compressing a patient history into a single temporal representation, MedCTA models clinical change at two complementary scopes. At the abnormality scope, clinically grounded queries construct abnormality-conditioned visual and textual trajectories to capture heterogeneous abnormality evolution. At the patient-course scope, global image and report sequences are modeled to capture overall clinical progression beyond any individual abnormality. Structured trend supervision is extracted from longitudinal reports by an offline LLM parser, removing the need for manual temporal annotations. Combined with static image-report alignment, MedCTA learns representations that preserve visit-level cross-modal correspondence while encoding longitudinal change semantics. Experiments on temporal image classification, image-text retrieval, and zero-shot classification show consistent gains over strong medical vision-language baselines.

Pathological assessment relies on recognizing fine-grained visual details in histological images. Vision-language models (VLMs) increasingly support pathology interpretation, yet their ability to perceive these details remains inadequate. This weakness leads to inaccurate cellular observations that can persist even when final answers are correct. In this paper, we propose ASPECT to improve visually grounded reasoning through explicit supervision of cellular appearance and abundance. ASPECT trains intermediate visual tokens through pathology feature reconstruction, cell feature alignment, and count supervision. Three-stage supervised fine-tuning teaches the model to perceive, generate visual tokens, and reason, followed by reinforcement learning that rewards answer correctness and consistency with reported measurements. We also introduce PathoVernier, a benchmark of 759 expert-reviewed questions from five pathology datasets covering four cellular composition tasks. It evaluates both final answers and intermediate measurements to expose errors hidden by answer accuracy. On PathoVernier, ASPECT achieves relative accuracy gains of approximately 19.2% over the strongest baseline, Gemini-3.1-Pro, and 99.3% over its Qwen3-VL-8B backbone, while reducing RAWR, which measures counting errors within correct responses, by 28.1% and 42.7%, respectively. ASPECT also improves over its backbone on three external pathology benchmarks covering classification and question answering beyond cellular composition tasks.

Pretrained biomedical vision-language models achieve strong zero-shot performance in biomedical image classification. However, downstream biomedical classification often depends on subtle visual differences between classes that may not be fully captured by pretrained representations. Few-shot adaptation addresses this mismatch by optimizing a task-specific predictor on a small labeled support set. Because the selected examples capture only part of the visual variation within the target classes, the adapted predictions can depend strongly on their composition. We propose Prompt-Anchored Residual Adaptation (PARA), which retains the frozen prompt prediction as a support-invariant semantic anchor and incorporates a visual prediction learned from the support set through an anchor-relative residual. The residual step is computed in a closed form from frozen support embeddings using anchor discrepancy and support agreement. Support-set dependence also limits evaluation: comparisons are fair within a shared draw but remain conditional on its composition. To obtain more reliable comparisons, we introduce a repeated-support protocol that separates support-selection variation from optimization randomness and reports both average and worst-20% performance. PARA achieves state-of-the-art performance in both few-shot classification and base-to-novel generalization.

Pathology vision-language foundation models vary in performance across diseases and tasks, with no single model consistently performing best. The high cost of expert pathology annotation can also limit the labeled data available for task-specific adaptation. Combining complementary pretrained representations is a potential approach to these limitations, yet learning an effective fusion from few labeled examples remains challenging. We introduce Few-shot Fusion Foundation Models of Computational Pathology (FFM-CP), which is a framework that combines multiple pathology vision-language models in the few-shot learning setting. The framework first aligns heterogeneous representations using a closed-form Orthogonal Procrustes transformation estimated from corresponding support images. This alignment preserves within-model feature geometry without training an additional alignment network. Within the aligned space, a unified graph enables information exchange across backbones by jointly refining support-image features and visual and textual class prototypes. These refined representations support complementary text-prototype and case-retrieval branches that capture semantic class knowledge and within-class visual variation, respectively. Each branch learns to combine predictions from all ordered backbone pairs, allowing queries encoded by one model to draw on evidence represented by another. We evaluate three backbone combinations on six histopathology datasets at 4, 8, and 16 shots per class. FFM-CP achieves higher mean macro-F1 than the strongest individually adapted member of each fused set in 50 of 54 comparisons. These findings suggest that combining complementary pretrained representations can improve histopathological classification when annotations are limited.

Artificial intelligence (AI) has advanced at a rapid pace in recent years. Initially, breakthroughs in large language models caught widespread attention. However, recent generations of frontier AI models have adopted multimodal capabilities as a first class citizen, with vision capabilities being central to that. In this paper, we evaluate three recently released models on the task of diagnosing dermatological conditions from patient-submitted images. The models chosen are at the low to mid tier in terms of pricing and thus represent a floor on current AI capabilities, not a ceiling. We evaluate AI performance relative to a panel of three certified dermatologists, who grade each image, and we present four interesting findings. Firstly, depending on the metric, the tested AI models are either on par or slightly trail humans in terms of inter-clinician agreement. Secondly, we find that asking AI models for a confidence rating produces poorly calibrated answers, meaning use of confidence thresholds should not be relied upon in a clinical setting. Thirdly, the effect of providing additional patient metadata is strongly model-specific, with one of the three models degrading on every metric considered. Finally, model cost is not predictive of performance. The best-performing model we tested costs on average $0.0045 per case.

We developed BrainVLM to classify all 12 World Health Organization (WHO) 2021 brain tumor types. BrainVLM integrates an uncertainty quantification strategy to indicate prediction reliability and a module for generating radiology reports to elucidate the clinical rationale. BrainVLM was trained on multi-modal data (MRI scans, demographics, and radiology reports) from 40,043 individuals. It was validated on 5,211 patients with pathologically confirmed brain tumors, including 3,877 held-out patients from the primary hospital and 1,334 patients from 11 independent hospitals. We further conducted two proof-of-concept studies to validate its clinical utility in AI-clinician workflows: 1) a blinded multireader study where 12 neuroradiologists across varying experience levels interpreted 248 retrospective cases with or without AI assistance, and 2) a real-world prospective study in which 1,009 patients were independently and blindly assessed by BrainVLM and radiologists before surgery. Additionally, we demonstrated BrainVLM's utility in preoperative molecular subgroup prediction for adult-type diffuse gliomas, using a multi-center cohort of 632 patients. In primary evaluation, BrainVLM achieved an area under the curve (macro-AUC) of 0.85 (95% CI: 0.84-0.86), and an F1 score of 0.82 (95% CI: 0.81-0.83), surpassing neuroradiologists (F1 = 0.80 (95% CI: 0.79-0.81)). In external validation across 11 centers, BrainVLM achieved an AUC = 0.80 (95% CI: 0.79-0.82) and F1 = 0.75 (95% CI: 0.73-0.78), compared with F1 = 0.71 (95% CI: 0.69-0.73) for neuroradiologists. In prospective real-world evaluation, BrainVLM maintained performance comparable to neuroradiologists.

Accurate PSMA PET/CT interpretation is central to prostate cancer management, yet existing PET/CT AI models typically address isolated tasks. We propose a unified PSMA PET/CT vision-language model for report generation, visual question answering, and lesion segmentation. The framework adopts an LLaVA-style architecture, comprising a PET/CT vision encoder, an MLP-Mixer projection module, a LoRA-tuned large language model, and a 3D segmentation branch. Training followed a four-stage strategy: vision encoder pretraining, projection-layer alignment, VLM fine-tuning, and final multitask tuning. Language tasks used 5,747 PSMA PET/CT datasets with paired reports, while segmentation used the PSMA subset of AutoPET. The model outperformed PET2REP and a CT-based baseline across standard report-generation metrics, improved performance across VQA question types, and achieved higher Dice and lesion-level overlap F1 than SegAnyPET and nnUNet. These results support the feasibility of a unified framework for structured, interactive, interpretable PSMA PET/CT analysis with voxel-level grounding within a single multitask model architecture.

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer \emph{semantics} rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0% to 57.5% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number $k$ of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5%) at both $k=0$ and the adapter's training-time $k=1$, 188/200 (94.0%) at the packaged runtime's default $k=3$ --- and the submitted confidence-gated override adds no net accuracy on top of $k=3$, selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~\cite{zhao2024ratescore} decrease as $k$ grows, but paired sign tests on token-F1 differences are nonsignificant ($p \ge 0.29$); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0% MCQ accuracy on the development holdout and 93.20% on the organizer's official pre-evaluation, versus 29.43% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

Vision-language models are increasingly explored for clinical prediction from electronic health records and medical images, where identifying unreliable predictions is important for safe deployment. Uncertainty estimation (UE) enables detecting such predictions, but its evaluation depends on a correctness criterion that determines whether each model output is correct. If this criterion disagrees with human judgement or distorts downstream UE performance, conclusions about model reliability can be misleading. We introduce a two-axis framework that evaluates correctness criteria by their agreement with human judgements and fidelity to human-referenced UE performance. We assess eight criteria across three clinical prediction tasks and three models using 450 predictions annotated by two reviewers. Across the audited tasks, canonical exact matching (EM) achieved the highest observed human agreement and lowest UE distortion, while the BERT-based matching (BEM) and LLM-judge also showed strong human agreement. Across four UE methods and 23,254 clinical predictions, criterion choice changed error-detection AUROC by up to 0.146 and reversed the relative ranking of UE methods. The LLM-judge also selectively accepted invalid or uncertain outputs, accepting 16 of 30 such human-identified errors. These results demonstrate that correctness assessment is an integral component of clinical UE evaluation and should be validated before UE methods are compared.

Vision-Language Models such as CLIP enable effective few-shot medical anomaly detection (AD) via strong image-text semantic alignment. However, their globally contrastive pretraining lacks explicit spatial supervision, limiting precise lesion localization. In contrast, Vision Foundation Models (VFMs) such as DINO learn spatially coherent patch representations via self-distillation and local-to-global consistency, better capturing fine-grained anatomical structures. Leveraging this complementarity, we propose Spatial-FAD, a spatial-aware few-shot medical AD framework that improves lesion localization by combining VFM spatial priors with CLIP semantics. Specifically, we introduce a VFM-enhanced adapter that injects a structural affinity prior derived from DINO into CLIP features. This structure-guided refinement encourages visual embeddings to better adhere to lesion boundaries while maintaining semantic alignment. To address the loss of spatial detail from patchification and the limited input resolution of CLIP, we adopt a sliding-window aggregation strategy. This generates high-resolution, spatially dense embeddings to further enhance localization granularity. Moreover, we introduce a prototype-enhanced support memory scheme to efficiently exploit the few-shot support set. This module stores compact prototypes for normal and abnormal patterns, reducing memory costs while boosting performance by fusing patch-to-prototype and image-text similarities. Extensive experiments on three benchmark datasets, including Liver CT, Retinal OCT, and Brain MRI, demonstrate that Spatial-FAD significantly outperforms state-of-the-art methods, especially in lesion segmentation. Notably, in the 4-shot scenario, our method achieves an average improvement of over 11.4% in Dice score and 1.8% in AUC. Code is available at: https://github.com/JuzhengMiao/Spatial-FAD.

Uncertainty estimation for medical vision--language models (VLMs) using conformal prediction has gained increasing attention due to its distribution-free coverage guarantees. However, standard conformal prediction relies on exchangeability between calibration and test data and typically requires a sufficiently large calibration set to obtain reliable coverage. These assumptions are difficult to satisfy in few-shot transfer settings, where only a small labeled support set is available to adapt a pretrained VLM to a new medical task, while an unlabeled query set is used for evaluation. Supervised fine-tuning on the support set changes the model parameters and consequently shifts the nonconformity score distribution, breaking exchangeability between calibration and query samples and leading to unreliable coverage under distribution shift. Existing transductive conformal adaptation methods often preserve validity by avoiding supervised updates. While this helps maintain conformal assumptions, it underutilizes the scarce labeled support data and limits task adaptation, which is the primary objective in few-shot learning. In this setting, conformal prediction should serve as an uncertainty estimation layer that supports the adapted model, rather than preventing adaptation itself. To this end, we propose AlignCP, a framework that reconciles supervised few-shot adaptation with conformal uncertainty estimation under non-exchangeability. AlignCP learns a reweighted calibration distribution that reduces the score-level discrepancy between the labeled support set and the unlabeled query set. By aligning the one-dimensional nonconformity score distributions, AlignCP aims to close the coverage gap induced by adaptation without requiring query labels.

Out-of-distribution (OOD) detection is crucial for safe deployment of medical AI systems, where domain shifts arise across institutions, acquisition protocols, and patient populations. VLMs enable zero-shot OOD detection by embedding images into a language-aligned latent space, where cross-modal similarity serves as a non-parametric confidence signal for identifying in-distribution samples. Yet existing methods rely almost exclusively on final-layer embeddings, implicitly assuming that the deepest representations are universally optimal. We first show that this assumption does not hold in medical imaging: intermediate layers provide complementary OOD signals, and the optimal representational depth depends on the respective image modality. While prior work selects layer combinations via entropy minimization of normalized histograms, we demonstrate that single-resolution entropy estimation is highly sensitive to binning choices, leading to performance variations of up to 19.3% AUROC. To address this instability, we propose a multi-resolution entropy estimation strategy that aggregates histogram statistics across multiple discretization scales, enabling robust and stable intermediate-layer selection. Across two medical OOD benchmarks, namely MIDOG and OASIS, covering distinct imaging modalities, diverse shift types, and different VLM backbones, our method consistently outperforms state-of-the-art approaches, offering a lightweight and stable solution for zero-shot OOD detection.

Large language models (LLMs) have demonstrated strong capabilities across diverse domains, showing considerable potential in medicine. However, their application in medical settings remains limited by the scarcity of visual question answering (VQA) datasets that capture clinical reasoning and explicit image-text alignment. Here, we leverage de-identified medical images and expert commentaries shared through clinician-oriented online resources. By combining an advanced LLM with clinician-in-the-loop verification, we established a rigorous pipeline to construct ThoughtMed-1M, a long-form medical VQA dataset containing over one million VQA pairs and designed to capture structured clinical reasoning and medical image-text alignment. To demonstrate its utility, we developed a FOundational LLM Trained on ThoughtMed-1M (FOLTMed). FOLTMed achieved state-of-the-art performance across 42 medical VQA benchmark datasets, with a macro accuracy of 85.4 percent. It also generated more clinically coherent responses on the ThoughtMed-1M test set, outperforming state-of-the-art models by 3 to 5 percent across factuality and similarity metrics, highlighting a scalable paradigm for advancing research on clinically grounded multimodal LLMs.

The joint interpretation of metabolic function and anatomical structure is essential for clinical diagnosis in whole-body PET/CT. Although recent advances in 3D medical vision-language models have demonstrated remarkable progress, current efforts are limited to regional CT imaging, leaving a critical void in comprehensive whole-body PET/CT analysis. In this work, we introduce MetaStructAtlas, a large-scale dataset for grounded whole-body PET/CT interpretation that synthesizes multimodal imaging with integrated anatomical, metabolic, and semantic annotations. MetaStructAtlas provides 490 co-registered 3D PET and CT volumes with 50,470 organ-level segmentation masks and grounded radiology reports. To facilitate interactive reasoning, we further developed MetaStructVQA, a standardized 3D grounded visual question-answering benchmark containing 100,565 QA pairs. This framework explicitly links diagnostic queries to visual evidence across modalities, encompassing anatomical, morphological, and metabolic characteristics. Finally, we evaluate state-of-the-art 3D medical VLMs on MetaStructVQA, establishing a robust foundation for multimodal representation learning and integrated whole-body reasoning in nuclear medicine.

Federated learning (FL) lets institutions train a shared model without exchanging data, and Low-Rank Adaptation (LoRA) makes this practical at scale by communicating only compact low-rank updates. Biomedical imaging is a compelling setting for this combination: patient data are archived behind privacy regulations, and institutions differ widely in scanners, protocols, and compute. Such heterogeneity raises the question of how federated LoRA updates should be aggregated, increasingly pressing as multimodal vision-language models become central to medical image analysis. We benchmark federated Parameter-efficient fine-tuning (PEFT) of BiomedCLIP for chest radiograph classification across four public cohorts on three continents (USA, Vietnam, Spain). Federated LoRA adaptation improves shared-class AUC on all four cohorts over the unadapted BiomedCLIP backbone (mean 0.687 to 0.802), showing that the gains come from federated adaptation rather than from the pretrained model's zero-shot ability. Relative to isolated single-cohort training, federation improves the weaker cohorts while largely preserving the strongest and approaches a centralized reference (0.812) that pools all data. The singular value decomposition (SVD)-based product-space aggregation introduced by FlexLoRA is essential to this gain (naive factor averaging drops mean AUC by 0.097), whereas a drift-correcting optimizer (FedProx) shows no benefit over FedAvg in our single-seed runs, consistent with LoRA's low-rank updates already limiting client drift. Biomedical vision-language models can thus be adapted collaboratively across heterogeneous, geographically distributed institutions without centralizing data.