HEAL-MedVQA. We introduce HEAL-MedVQA (Hallucination Evaluation via Localization in Medical VQA), a benchmark designed to evaluate hallucination robustness and localization ability of medical LMMs using 67K doctor-annotated VQA pairs and anatomical segmentation masks.
Evaluation Protocols. HEAL-MedVQA includes two protocols—Textual Perturbation Test (TPT) and Visual Perturbation Test (VPT)—to diagnose shortcut reliance on language or irrelevant image areas.
LobA Framework. We propose Localize-before-Answer (LobA), a method that improves visual grounding by training models to localize pathological regions and self-prompt for more accurate answers.
Performance. LobA significantly outperforms state-of-the-art medical LMMs on HEAL-MedVQA, advancing both robustness and reliability in medical VQA.
Contradicting Source Evidence: LMMs frequently generate responses that are false or contradict the visual information presented in images, often due to poor localization reasoning.
Reliance on Shortcuts Instead of Analysis: Models often bypass analyzing relevant image regions. Instead, they rely on:
Textual Perturbation Test (TPT): Tests language pattern reliance by swapping anatomical terms or diseases in correctly answered questions. Measures if model's answer changes appropriately, verifying true visual understanding.
Visual Perturbation Test (VPT): Evaluates visual grounding by overlaying disease-free anatomical regions onto test images. Verifies if model's decisions are based on correct image areas.
HEAL-MedVQA contains 67K QA pairs with doctor-annotated segmentation masks. Our pipeline:
Anatomy Segmentation: Generate precise masks for core anatomical structures
Disease Extraction: Obtain disease labels and bounding boxes
Anatomy-Disease Mapping: Link anatomies to diseases using IoU scores
QA Generation: Create QA pairs from validated associations
Region Localization: When a medical image question is received, our model first localizes the region of interest as a mask.
Patch Interpolation: The mask is interpolated into patches, enhancing focus on the relevant region.
Answer Generation: The highlighted input is processed again to generate the final, region-focused answer.
Performance on the proposed HEAL-MedVQA benchmark: LobA significantly outperforms existing medical VQA models, achieving a 15% increase in accuracy across diverse medical image types and question categories.
Textual and Visual Perturbation Tests: Our framework demonstrates superior resilience against both textual perturbations (Left) and visual perturbations (Right), maintaining consistent performance even under challenging test conditions.
Ablation studies of different components of LobA: LobA achieves precise anatomical localization with an average IoU of 0.85, significantly reducing hallucination instances and improving answer reliability.
