Although new vision foundation models such as Segment Anything Model 2 (SAM2) have significantly enhanced zero-shot image segmentation capabilities, reliance on human-provided prompts poses significant challenges in adapting SAM2 to medical image segmentation tasks. Moreover, SAM2's performance in medical image segmentation was limited by the domain shift issue, since it was originally trained on natural images and videos. To address these challenges, we proposed SAM2 with support-set guided prompting (SAM2-SGP), a framework that eliminated the need for manual prompts. The proposed model leveraged the memory mechanism of SAM2 to generate pseudo-masks using image-mask pairs from a support set via a Pseudo-mask Generation (PMG) module. We further introduced a novel Pseudo-mask Attention (PMA) module, which used these pseudo-masks to automatically generate bounding boxes and enhance localized feature extraction by guiding attention to relevant areas. Furthermore, a low-rank adaptation (LoRA) strategy was adopted to mitigate the domain shift issue. The proposed framework was evaluated on both 2D and 3D datasets across multiple medical imaging modalities, including fundus photography, X-ray, computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), and ultrasound. The results demonstrated a significant performance improvement over state-of-the-art models, such as nnUNet and SwinUNet, as well as foundation models, such as SAM2 and MedSAM2, underscoring the effectiveness of the proposed approach. Our code is publicly available at https://github.com/astlian9/SAM_Support.

The success of deep learning in medical imaging applications has led several companies to deploy proprietary models in diagnostic workflows, offering monetized services. Even though model weights are hidden to protect the intellectual property of the service provider, these models are exposed to model stealing (MS) attacks, where adversaries can clone the model's functionality by querying it with a proxy dataset and training a thief model on the acquired predictions. While extensively studied on general vision tasks, the susceptibility of medical imaging models to MS attacks remains inadequately explored. This paper investigates the vulnerability of black-box medical imaging models to MS attacks under realistic conditions where the adversary lacks access to the victim model's training data and operates with limited query budgets. We demonstrate that adversaries can effectively execute MS attacks by using publicly available datasets. To further enhance MS capabilities with limited query budgets, we propose a two-step model stealing approach termed QueryWise. This method capitalizes on unlabeled data obtained from a proxy distribution to train the thief model without incurring additional queries. Evaluation on two medical imaging models for Gallbladder Cancer and COVID-19 classification substantiates the effectiveness of the proposed attack. The source code is available at https://github.com/rajankita/QueryWise.

Foundation models for volumetric medical image segmentation have emerged as powerful tools in clinical workflows, enabling radiologists to delineate regions of interest through intuitive clicks. While these models demonstrate promising capabilities in segmenting previously unseen anatomical structures, their performance is strongly influenced by prompt quality. In clinical settings, radiologists often provide suboptimal prompts, which affects segmentation reliability and accuracy. To address this limitation, we present SafeClick, an error-tolerant interactive segmentation approach for medical volumes based on hierarchical expert consensus. SafeClick operates as a plug-and-play module compatible with foundation models including SAM 2 and MedSAM 2. The framework consists of two key components: a collaborative expert layer (CEL) that generates diverse feature representations through specialized transformer modules, and a consensus reasoning layer (CRL) that performs cross-referencing and adaptive integration of these features. This architecture transforms the segmentation process from a prompt-dependent operation to a robust framework capable of producing accurate results despite imperfect user inputs. Extensive experiments across 15 public datasets demonstrate that our plug-and-play approach consistently improves the performance of base foundation models, with particularly significant gains when working with imperfect prompts. The source code is available at https://github.com/yifangao112/SafeClick.

Existing foundation models for neuroimaging are often prohibitively large and data-intensive. We introduce BrainSymphony, a lightweight, parameter-efficient foundation model that achieves state-of-the-art performance while being pre-trained on significantly smaller public datasets. BrainSymphony's strong multimodal architecture processes functional MRI data through parallel spatial and temporal transformer streams, which are then efficiently distilled into a unified representation by a Perceiver module. Concurrently, it models structural connectivity from diffusion MRI using a novel signed graph transformer to encode the brain's anatomical structure. These powerful, modality-specific representations are then integrated via an adaptive fusion gate. Despite its compact design, our model consistently outperforms larger models on a diverse range of downstream benchmarks, including classification, prediction, and unsupervised network identification tasks. Furthermore, our model revealed novel insights into brain dynamics using attention maps on a unique external psilocybin neuroimaging dataset (pre- and post-administration). BrainSymphony establishes that architecturally-aware, multimodal models can surpass their larger counterparts, paving the way for more accessible and powerful research in computational neuroscience.

Endoscopic image classification plays a pivotal role in medical diagnostics by identifying anatomical landmarks and pathological findings. However, conventional closed-set classification frameworks are inherently limited in open-world clinical settings, where previously unseen conditions can arise andcompromise model reliability. To address this, we explore the application of Open Set Recognition (OSR) techniques on the Kvasir dataset, a publicly available and diverse endoscopic image collection. In this study, we evaluate and compare the OSR capabilities of several representative deep learning architectures, including ResNet-50, Swin Transformer, and a hybrid ResNet-Transformer model, under both closed-set and open-set conditions. OpenMax is adopted as a baseline OSR method to assess the ability of these models to distinguish known classes from previously unseen categories. This work represents one of the first efforts to apply open set recognition to the Kvasir dataset and provides a foundational benchmark for evaluating OSR performance in medical image analysis. Our results offer practical insights into model behavior in clinically realistic settings and highlight the importance of OSR techniques for the safe deployment of AI systems in endoscopy.

Positron emission tomography (PET)/computed tomography (CT) myocardial perfusion imaging (MPI) is a vital diagnostic tool, especially in patients with cardiometabolic syndrome. Low-dose CT scans are routinely performed with PET for attenuation correction and potentially contain valuable data about body tissue composition. Deep learning and image processing were combined to automatically quantify skeletal muscle (SM), bone and adipose tissue from these scans and then evaluate their associations with death or myocardial infarction (MI). In PET MPI from three sites, deep learning quantified SM, bone, epicardial adipose tissue (EAT), subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and intermuscular adipose tissue (IMAT). Sex-specific thresholds for abnormal values were established. Associations with death or MI were evaluated using unadjusted and multivariable models adjusted for clinical and imaging factors. This study included 10 085 patients, with median age 68 (interquartile range 59-76) and 5767 (57%) male. Body tissue segmentations were completed in 102 ± 4 s. Higher VAT density was associated with an increased risk of death or MI in both unadjusted [hazard ratio (HR) 1.40, 95% confidence interval (CI) 1.37-1.43] and adjusted (HR 1.24, 95% CI 1.19-1.28) analyses, with similar findings for IMAT, SAT, and EAT. Patients with elevated VAT density and reduced myocardial flow reserve had a significantly increased risk of death or MI (adjusted HR 2.49, 95% CI 2.23-2.77). Volumetric body tissue composition can be obtained rapidly and automatically from standard cardiac PET/CT. This new information provides a detailed, quantitative assessment of sarcopenia and cardiometabolic health for physicians.

Extracting clinical entities from unstructured medical documents is critical for improving clinical decision support and documentation workflows. This study examines the performance of various encoder and decoder models trained for Named Entity Recognition (NER) of clinical parameters in pathology and radiology reports, highlighting the applicability of Large Language Models (LLMs) for this task. Three NER methods were evaluated: (1) flat NER using transformer-based models, (2) nested NER with a multi-task learning setup, and (3) instruction-based NER utilizing LLMs. A dataset of 2013 pathology reports and 413 radiology reports, annotated by medical students, was used for training and testing. The performance of encoder-based NER models (flat and nested) was superior to that of LLM-based approaches. The best-performing flat NER models achieved F1-scores of 0.87-0.88 on pathology reports and up to 0.78 on radiology reports, while nested NER models performed slightly lower. In contrast, multiple LLMs, despite achieving high precision, yielded significantly lower F1-scores (ranging from 0.18 to 0.30) due to poor recall. A contributing factor appears to be that these LLMs produce fewer but more accurate entities, suggesting they become overly conservative when generating outputs. LLMs in their current form are unsuitable for comprehensive entity extraction tasks in clinical domains, particularly when faced with a high number of entity types per document, though instructing them to return more entities in subsequent refinements may improve recall. Additionally, their computational overhead does not provide proportional performance gains. Encoder-based NER models, particularly those pre-trained on biomedical data, remain the preferred choice for extracting information from unstructured medical documents.

Major Adverse Cardiovascular Events (MACE) remain the leading cause of mortality globally, as reported in the Global Disease Burden Study 2021. Opportunistic screening leverages data collected from routine health check-ups and multimodal data can play a key role to identify at-risk individuals. Chest X-rays (CXR) provide insights into chronic conditions contributing to major adverse cardiovascular events (MACE), while 12-lead electrocardiogram (ECG) directly assesses cardiac electrical activity and structural abnormalities. Integrating CXR and ECG could offer a more comprehensive risk assessment than conventional models, which rely on clinical scores, computed tomography (CT) measurements, or biomarkers, which may be limited by sampling bias and single modality constraints. We propose a novel predictive modeling framework - MOSCARD, multimodal causal reasoning with co-attention to align two distinct modalities and simultaneously mitigate bias and confounders in opportunistic risk estimation. Primary technical contributions are - (i) multimodal alignment of CXR with ECG guidance; (ii) integration of causal reasoning; (iii) dual back-propagation graph for de-confounding. Evaluated on internal, shift data from emergency department (ED) and external MIMIC datasets, our model outperformed single modality and state-of-the-art foundational models - AUC: 0.75, 0.83, 0.71 respectively. Proposed cost-effective opportunistic screening enables early intervention, improving patient outcomes and reducing disparities.

BackgroundGenerative Pre-trained Transformer 4 (GPT-4) has demonstrated strong performance in standardized medical examinations but has limitations in real-world clinical settings. The newly released multimodal GPT-4o model, which integrates text and image inputs to enhance diagnostic capabilities, and the multimodal o1 model, which incorporates advanced reasoning, may address these limitations. ObjectiveThis study aimed to compare the performance of GPT-4o and o1 against clinicians in real-world clinical case challenges. MethodsThis retrospective, cross-sectional study used Medscape case challenge questions from May 2011 to June 2024 (n = 1,426). Each case included text and images of patient history, physical examination findings, diagnostic test results, and imaging studies. Clinicians were required to choose one answer from among multiple options, with the most frequent response defined as the clinicians decision. Data-based decisions were made using GPT models (3.5 Turbo, 4 Turbo, 4 Omni, and o1) to interpret the text and images, followed by a process to provide a formatted answer. We compared the performances of the clinicians and GPT models using Mixed-effects logistic regression analysis. ResultsOf the 1,426 questions, clinicians achieved an overall accuracy of 85.0%, whereas GPT-4o and o1 demonstrated higher accuracies of 88.4% and 94.3% (mean difference 3.4%; P = .005 and mean difference 9.3%; P < .001), respectively. In the multimodal performance analysis, which included cases involving images (n = 917), GPT-4o achieved an accuracy of 88.3%, and o1 achieved 93.9%, both significantly outperforming clinicians (mean difference 4.2%; P = .005 and mean difference 9.8%; P < .001). o1 showed the highest accuracy across all question categories, achieving 92.6% in diagnosis (mean difference 14.5%; P < .001), 97.0% in disease characteristics (mean difference 7.2%; P < .001), 92.6% in examination (mean difference 7.3%; P = .002), and 94.8% in treatment (mean difference 4.3%; P = .005), consistently outperforming clinicians. In terms of medical specialty, o1 achieved 93.6% accuracy in internal medicine (mean difference 10.3%; P < .001), 96.6% in major surgery (mean difference 9.2%; P = .030), 97.3% in psychiatry (mean difference 10.6%; P = .030), and 95.4% in minor specialties (mean difference 10.0%; P < .001), significantly surpassing clinicians. Across five trials, GPT-4o and o1 provided the correct answer 5/5 times in 86.2% and 90.7% of the cases, respectively. ConclusionsThe GPT-4o and o1 models achieved higher accuracy than clinicians in clinical case challenge questions, particularly in disease diagnosis. The GPT-4o and o1 could serve as valuable tools to assist healthcare professionals in clinical settings.

PurposeTo assess and compare the accuracy and legitimacy of multimodal large language models (LLMs) on the Japan Diagnostic Radiology Board Examination (JDRBE). Materials and methodsThe dataset comprised questions from JDRBE 2021, 2023, and 2024, with ground-truth answers established through consensus among multiple board-certified diagnostic radiologists. Questions without associated images and those lacking unanimous agreement on answers were excluded. Eight LLMs were evaluated: GPT-4 Turbo, GPT-4o, GPT-4.5, GPT-4.1, o3, o4-mini, Claude 3.7 Sonnet, and Gemini 2.5 Pro. Each model was evaluated under two conditions: with inputting images (vision) and without (text-only). Performance differences between the conditions were assessed using McNemars exact test. Two diagnostic radiologists (with 2 and 18 years of experience) independently rated the legitimacy of responses from four models (GPT-4 Turbo, Claude 3.7 Sonnet, o3, and Gemini 2.5 Pro) using a five-point Likert scale, blinded to model identity. Legitimacy scores were analyzed using Friedmans test, followed by pairwise Wilcoxon signed-rank tests with Holm correction. ResultsThe dataset included 233 questions. Under the vision condition, o3 achieved the highest accuracy at 72%, followed by o4-mini (70%) and Gemini 2.5 Pro (70%). Under the text-only condition, o3 topped the list with an accuracy of 67%. Addition of image input significantly improved the accuracy of two models (Gemini 2.5 Pro and GPT-4.5), but not the others. Both o3 and Gemini 2.5 Pro received significantly higher legitimacy scores than GPT-4 Turbo and Claude 3.7 Sonnet from both raters. ConclusionRecent multimodal LLMs, particularly o3 and Gemini 2.5 Pro, have demonstrated remarkable progress on JDRBE questions, reflecting their rapid evolution in diagnostic radiology. Secondary abstract Eight multimodal large language models were evaluated on the Japan Diagnostic Radiology Board Examination. OpenAIs o3 and Google DeepMinds Gemini 2.5 Pro achieved high accuracy rates (72% and 70%) and received good legitimacy scores from human raters, demonstrating steady progress.