How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images

• URL: http://arxiv.org/abs/2603.14323v1
• Date: Sun, 15 Mar 2026 10:46:27 GMT
• Title: How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images
• Abstract summary: Generalist multimodal large language models (MLLMs) have achieved impressive performance across a wide range of vision-language tasks.<n>However, their performance on medical tasks, particularly in zero-shot settings where generalization is critical, remains suboptimal.<n>We present a pioneering systematic investigation into the visual grounding capabilities of state-of-the-art medical MLLMs.
• Score: 16.362951636873248
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Generalist multimodal large language models (MLLMs) have achieved impressive performance across a wide range of vision-language tasks. However, their performance on medical tasks, particularly in zero-shot settings where generalization is critical, remains suboptimal. A key research gap is the limited understanding of why medical MLLMs underperform in medical image interpretation. In this work, we present a pioneering systematic investigation into the visual grounding capabilities of state-of-the-art medical MLLMs. To disentangle visual grounding from semantic grounding, we design VGMED, a novel evaluation dataset developed with expert clinical guidance, explicitly assessing the visual grounding capability of medical MLLMs. We introduce new quantitative metrics and conduct detailed qualitative analyses. Our study across eight state-of-the-art (SOTA) medical MLLMs validates that they often fail to ground their predictions in clinically relevant image regions. We note that this finding is specific to medical image analysis; in contrast, prior work has shown that MLLMs are capable of grounding their predictions in the correct image regions when applied to natural scene images. Motivated by these findings, we propose VGRefine, a simple yet effective inference-time method that refines attention distribution to improve visual grounding in medical settings. Our approach achieves SOTA performance across 6 diverse Med-VQA benchmarks (over 110K VQA samples from 8 imaging modalities) without requiring additional training or external expert models. Overall, our work, for the first time, systematically validates inadequate visual grounding as one of the key contributing factors for medical MLLMs' under-performance. Additional experiments are included in the Supp.

Related papers

• MMBU: A Massive Multi-modal Biomedical Understanding Benchmark to Probe the Perception Capabilities of Vision-Language Models [83.50100003741628]
  Vision and language models (VLMs) hold immense promise to transform biomedical imaging.<n>We introduce the Massive Multimodal Biomedical Understanding (MMBU) benchmark.<n>It is the largest vision and language benchmark to date, covering 35 submodalities with rich structured metadata.
  arXiv  Detail & Related papers  (2026-06-04T20:24:47Z)
• Lost in the Hype: Revealing and Dissecting the Performance Degradation of Medical Multimodal Large Language Models in Image Classification [14.247959730104085]
  multimodal large language models (MLLMs) have sparked an unprecedented wave of applications in the field of medical imaging analysis.<n>However, medical image classification reveals a sobering reality: state-of-the-art medical MLLMs consistently underperform compared to traditional deep learning models.<n>In this paper, we conduct extensive experiments on 14 open-source medical MLLMs across three representative image classification datasets.
  arXiv  Detail & Related papers  (2026-04-09T15:07:26Z)
• Does medical specialization of VLMs enhance discriminative power?: A comprehensive investigation through feature distribution analysis [2.243145970857166]
  This study investigates the feature representations produced by publicly available open source medical vision-language models (VLMs)<n>Our experiments showed that medical VLMs can extract discriminative features that are effective for medical classification tasks.<n>Our results imply that enhancing text encoder is more crucial than training intensively on medical images when developing medical VLMs.
  arXiv  Detail & Related papers  (2026-01-21T08:53:40Z)
• Evaluating the Diagnostic Classification Ability of Multimodal Large Language Models: Insights from the Osteoarthritis Initiative [14.002322217782364]
  Multimodal large language models (MLLMs) show promising performance on medical visual question answering (VQA) and report generation.<n>We evaluated MLLM architectures on knee osteoarthritis (OA) radiograph classification.
  arXiv  Detail & Related papers  (2026-01-05T13:31:44Z)
• TemMed-Bench: Evaluating Temporal Medical Image Reasoning in Vision-Language Models [54.48710348910535]
  Existing medical reasoning benchmarks primarily focus on analyzing a patient's condition based on an image from a single visit.<n>We introduce TemMed-Bench, the first benchmark designed for analyzing changes in patients' conditions between different clinical visits.
  arXiv  Detail & Related papers  (2025-09-29T17:51:26Z)
• Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning [57.873833577058]
  We build a multimodal dataset enriched with extensive medical knowledge.<n>We then introduce our medical-specialized MLLM: Lingshu.<n>Lingshu undergoes multi-stage training to embed medical expertise and enhance its task-solving capabilities.
  arXiv  Detail & Related papers  (2025-06-08T08:47:30Z)
• Medical Large Vision Language Models with Multi-Image Visual Ability [46.889345205047675]
  We present the Med-MIM instruction dataset, comprising 83.2K medical multi-image QA pairs.<n>We fine-tune Mantis and LLaVA-Med, resulting in two specialized medical VLMs: MIM-LLaVA-Med and Med-Mantis.<n>We also develop the Med-MIM benchmark to evaluate the medical multi-image understanding capabilities of LVLMs.
  arXiv  Detail & Related papers  (2025-05-25T08:31:22Z)
• LLaVA-RadZ: Can Multimodal Large Language Models Effectively Tackle Zero-shot Radiology Recognition? [59.81732629438753]
  We propose LLaVA-RadZ, a simple yet effective framework for zero-shot medical disease recognition via utilizing the existing MLLM features.<n>Specifically, we design an end-to-end training strategy, termed Decoding-Side Feature Alignment Training (DFAT) to take advantage of the characteristics of the MLLM decoder architecture.<n>We also introduce a Domain Knowledge Anchoring Module (DKAM) to exploit the intrinsic medical knowledge of large models.
  arXiv  Detail & Related papers  (2025-03-10T16:05:40Z)
• Improving Medical Large Vision-Language Models with Abnormal-Aware Feedback [57.98393950821579]
  We propose a novel UMed-LVLM designed to unveil medical abnormalities.<n>We propose a prompt method utilizing the GPT-4V to generate diagnoses based on identified abnormal areas in medical images.<n>Our UMed-LVLM significantly outperforms existing Med-LVLMs in identifying and understanding medical abnormalities.
  arXiv  Detail & Related papers  (2025-01-02T17:37:20Z)
• Parameter-Efficient Fine-Tuning Medical Multimodal Large Language Models for Medical Visual Grounding [9.144030136201476]
  Multimodal large language models (MLLMs) inherit the superior text understanding capabilities of LLMs and extend these capabilities to multimodal scenarios. These models achieve excellent results in the general domain of multimodal tasks. However, in the medical domain, the substantial training costs and the requirement for extensive medical data pose challenges to the development of medical MLLMs.
  arXiv  Detail & Related papers  (2024-10-31T11:07:26Z)
• MedVH: Towards Systematic Evaluation of Hallucination for Large Vision Language Models in the Medical Context [21.562034852024272]
  Large Vision Language Models (LVLMs) have recently achieved superior performance in various tasks on natural image and text data. Despite their advancements, there has been scant research on the robustness of these models against hallucination when fine-tuned on smaller datasets. We introduce a new benchmark dataset, the Medical Visual Hallucination Test (MedVH), to evaluate the hallucination of domain-specific LVLMs.
  arXiv  Detail & Related papers  (2024-07-03T00:59:03Z)
• OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM [48.16696073640864]
  We introduce OmniMedVQA, a novel comprehensive medical Visual Question Answering (VQA) benchmark. All images in this benchmark are sourced from authentic medical scenarios. We have found that existing LVLMs struggle to address these medical VQA problems effectively.
  arXiv  Detail & Related papers  (2024-02-14T13:51:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.