Recent work has shown promising performance of frontier large language models (LLMs) and their multimodal counterparts in medical quizzes and diagnostic tasks, highlighting their potential for broad clinical utility given their accessible, general-purpose nature. However, beyond diagnosis, a fundamental aspect of medical image interpretation is the ability to localize pathological findings. Evaluating localization not only has clinical and educational relevance but also provides insight into a model's spatial understanding of anatomy and disease. Here, we systematically assess two general-purpose MLLMs (GPT-4 and GPT-5) and a domain-specific model (MedGemma) in their ability to localize pathologies on chest radiographs, using a prompting pipeline that overlays a spatial grid and elicits coordinate-based predictions. Averaged across nine pathologies in the CheXlocalize dataset, GPT-5 exhibited a localization accuracy of 49.7%, followed by GPT-4 (39.1%) and MedGemma (17.7%), all lower than a task-specific CNN baseline (59.9%) and a radiologist benchmark (80.1%). Despite modest performance, error analysis revealed that GPT-5's predictions were largely in anatomically plausible regions, just not always precisely localized. GPT-4 performed well on pathologies with fixed anatomical locations, but struggled with spatially variable findings and exhibited anatomically implausible predictions more frequently. MedGemma demonstrated the lowest performance on all pathologies, but showed improvements when provided examples through few shot prompting. Our findings highlight both the promise and limitations of current MLLMs in medical imaging and underscore the importance of integrating them with task-specific tools for reliable use.

Foundation models (FMs) promise to generalize medical imaging, but their effectiveness varies. It remains unclear how pre-training domain (medical vs. general), paradigm (e.g., text-guided), and architecture influence embedding quality, hindering the selection of optimal encoders for specific radiology tasks. To address this, we evaluate vision encoders from eight medical and general-domain FMs for chest X-ray analysis. We benchmark classification (pneumothorax, cardiomegaly) and segmentation (pneumothorax, cardiac boundary) using linear probing and fine-tuning. Our results show that domain-specific pre-training provides a significant advantage; medical FMs consistently outperformed general-domain models in linear probing, establishing superior initial feature quality. However, feature utility is highly task-dependent. Pre-trained embeddings were strong for global classification and segmenting salient anatomy (e.g., heart). In contrast, for segmenting complex, subtle pathologies (e.g., pneumothorax), all FMs performed poorly without significant fine-tuning, revealing a critical gap in localizing subtle disease. Subgroup analysis showed FMs use confounding shortcuts (e.g., chest tubes for pneumothorax) for classification, a strategy that fails for precise segmentation. We also found that expensive text-image alignment is not a prerequisite; image-only (RAD-DINO) and label-supervised (Ark+) FMs were among top performers. Notably, a supervised, end-to-end baseline remained highly competitive, matching or exceeding the best FMs on segmentation tasks. These findings show that while medical pre-training is beneficial, architectural choices (e.g., multi-scale) are critical, and pre-trained features are not universally effective, especially for complex localization tasks where supervised models remain a strong alternative.

Chest radiographs (CXRs) are among the most common tests in medicine. Automated image interpretation may reduce radiologists\' workload and expand access to diagnostic expertise. Deep learning multi-task and foundation models have shown strong performance for CXR interpretation but are vulnerable to shortcut learning, where models rely on spurious and off-target correlations rather than clinically relevant features to make decisions. We introduce RoentMod, a counterfactual image editing framework that generates anatomically realistic CXRs with user-specified, synthetic pathology while preserving unrelated anatomical features of the original scan. RoentMod combines an open-source medical image generator (RoentGen) with an image-to-image modification model without requiring retraining. In reader studies with board-certified radiologists and radiology residents, RoentMod-produced images appeared realistic in 93\% of cases, correctly incorporated the specified finding in 89-99\% of cases, and preserved native anatomy comparable to real follow-up CXRs. Using RoentMod, we demonstrate that state-of-the-art multi-task and foundation models frequently exploit off-target pathology as shortcuts, limiting their specificity. Incorporating RoentMod-generated counterfactual images during training mitigated this vulnerability, improving model discrimination across multiple pathologies by 3-19\% AUC in internal validation and by 1-11\% for 5 out of 6 tested pathologies in external testing. These findings establish RoentMod as a broadly applicable tool for probing and correcting shortcut learning in medical AI. By enabling controlled counterfactual interventions, RoentMod enhances the robustness and interpretability of CXR interpretation models and provides a generalizable strategy for improving foundation models in medical imaging.

An essential step in deploying medical imaging models is ensuring alignment with clinical knowledge and interpretability. We focus on mapping clinical concepts into the latent space of generative models to identify Concept Activation Vectors (CAVs). Using a simple reconstruction autoencoder, we link user-defined concepts to image-level features without explicit label training. The extracted concepts are stable across datasets, enabling visual explanations that highlight clinically relevant features. By traversing latent space along concept directions, we produce counterfactuals that exaggerate or reduce specific clinical features. Preliminary results on chest X-rays show promise for large pathologies like cardiomegaly, while smaller pathologies remain challenging due to reconstruction limits. Although not outperforming baselines, this approach offers a path toward interpretable, concept-based explanations aligned with clinical knowledge.

Developing reliable and generalizable deep learning systems for medical imaging faces significant obstacles due to spurious correlations, data imbalances, and limited text annotations in datasets. Addressing these challenges requires architectures robust to the unique complexities posed by medical imaging data. The rapid advancements in vision-language foundation models within the natural image domain prompt the question of how they can be adapted for medical imaging tasks. In this work, we present PRISM, a framework that leverages foundation models to generate high-resolution, language-guided medical image counterfactuals using Stable Diffusion. Our approach demonstrates unprecedented precision in selectively modifying spurious correlations (the medical devices) and disease features, enabling the removal and addition of specific attributes while preserving other image characteristics. Through extensive evaluation, we show how PRISM advances counterfactual generation and enables the development of more robust downstream classifiers for clinically deployable solutions. To facilitate broader adoption and research, we make our code publicly available at https://github.com/Amarkr1/PRISM.

Despite significant advancements in report generation methods, a critical limitation remains: the lack of interpretability in the generated text. This paper introduces an innovative approach to enhance the explainability of text generated by report generation models. Our method employs cyclic text manipulation and visual comparison to identify and elucidate the features in the original content that influence the generated text. By manipulating the generated reports and producing corresponding images, we create a comparative framework that highlights key attributes and their impact on the text generation process. This approach not only identifies the image features aligned to the generated text but also improves transparency but also provides deeper insights into the decision-making mechanisms of the report generation models. Our findings demonstrate the potential of this method to significantly enhance the interpretability and transparency of AI-generated reports.

In this paper, we introduce a novel Multi-Modal Contrastive Pre-training Framework that synergistically combines X-rays, electrocardiograms (ECGs), and radiology/cardiology reports. Our approach leverages transformers to encode these diverse modalities into a unified representation space, aiming to enhance diagnostic accuracy and facilitate comprehensive patient assessments. We utilize LoRA-Peft to significantly reduce trainable parameters in the LLM and incorporate recent linear attention dropping strategy in the Vision Transformer(ViT) for smoother attention. Furthermore, we provide novel multimodal attention explanations and retrieval for our model. To the best of our knowledge, we are the first to propose an integrated model that combines X-ray, ECG, and Radiology/Cardiology Report with this approach. By utilizing contrastive loss, MoRE effectively aligns modality-specific features into a coherent embedding, which supports various downstream tasks such as zero-shot classification and multimodal retrieval. Employing our proposed methodology, we achieve state-of-the-art (SOTA) on the Mimic-IV, CheXpert, Edema Severity, and PtbXl downstream datasets, surpassing existing multimodal approaches. Our proposed framework shows significant improvements in capturing intricate inter-modal relationships and its robustness in medical diagnosis that establishes a framework for future research in multimodal learning in the healthcare sector.

Radiological services are experiencing unprecedented demand, leading to increased interest in automating radiology report generation. Existing Vision-Language Models (VLMs) suffer from hallucinations, lack interpretability, and require expensive fine-tuning. We introduce SAE-Rad, which uses sparse autoencoders (SAEs) to decompose latent representations from a pre-trained vision transformer into human-interpretable features. Our hybrid architecture combines state-of-the-art SAE advancements, achieving accurate latent reconstructions while maintaining sparsity. Using an off-the-shelf language model, we distil ground-truth reports into radiological descriptions for each SAE feature, which we then compile into a full report for each image, eliminating the need for fine-tuning large models for this task. To the best of our knowledge, SAE-Rad represents the first instance of using mechanistic interpretability techniques explicitly for a downstream multi-modal reasoning task. On the MIMIC-CXR dataset, SAE-Rad achieves competitive radiology-specific metrics compared to state-of-the-art models while using significantly fewer computational resources for training. Qualitative analysis reveals that SAE-Rad learns meaningful visual concepts and generates reports aligning closely with expert interpretations. Our results suggest that SAEs can enhance multimodal reasoning in healthcare, providing a more interpretable alternative to existing VLMs.

Large Vision-Language Models (LVLMs) have achieved significant success in recent years, and they have been extended to the medical domain. Although demonstrating satisfactory performance on medical Visual Question Answering (VQA) tasks, Medical LVLMs (MLVLMs) suffer from the hallucination problem, which makes them fail to diagnose complex pathologies. Moreover, they readily fail to learn minority pathologies due to imbalanced training data. We propose two prompting strategies for MLVLMs that reduce hallucination and improve VQA performance. In the first strategy, we provide a detailed explanation of the queried pathology. In the second strategy, we fine-tune a cheap, weak learner to achieve high performance on a specific metric, and textually provide its judgment to the MLVLM. Tested on the MIMIC-CXR-JPG and Chexpert datasets, our methods significantly improve the diagnostic F1 score, with the highest increase being 0.27. We also demonstrate that our prompting strategies can be extended to general LVLM domains. Based on POPE metrics, it effectively suppresses the false negative predictions of existing LVLMs and improves Recall by approximately 0.07.

GPT-4V's purported strong multimodal abilities raise interests in using it to automate radiology report writing, but there lacks thorough evaluations. In this work, we perform a systematic evaluation of GPT-4V in generating radiology reports on two chest X-ray report datasets: MIMIC-CXR and IU X-Ray. We attempt to directly generate reports using GPT-4V through different prompting strategies and find that it fails terribly in both lexical metrics and clinical efficacy metrics. To understand the low performance, we decompose the task into two steps: 1) the medical image reasoning step of predicting medical condition labels from images; and 2) the report synthesis step of generating reports from (groundtruth) conditions. We show that GPT-4V's performance in image reasoning is consistently low across different prompts. In fact, the distributions of model-predicted labels remain constant regardless of which groundtruth conditions are present on the image, suggesting that the model is not interpreting chest X-rays meaningfully. Even when given groundtruth conditions in report synthesis, its generated reports are less correct and less natural-sounding than a finetuned LLaMA-2. Altogether, our findings cast doubt on the viability of using GPT-4V in a radiology workflow.