Abstract
Medical visual question answering (Med-VQA) requires visual evidence tailored to each question and acquired at appropriate spatial scales. Global image context may suffice for modality or anatomy questions, whereas localized findings require regional inspection, and retrieved medical context may provide complementary information. Most methods based on multimodal large language models (MLLMs), however, apply a uniform inference process with limited explicit control over the acquisition, organization, and review of visual evidence. We present agentic Visual Evidence construction and Review Architecture for Medical VQA (VERA-M), an agentic inference framework for question-conditioned visual evidence construction and review. For each image–question pair, VERA-M selects either direct or multi-step inference. On the multi-step route, it selects and executes a question-specific sequence of visual actions drawn from global observation, region-of-interest (ROI) grounding, and regional analysis, and organizes the resulting observations into an ordered, provenance-aware evidence chain. An Evidence Review and Fusion Agent evaluates the sufficiency and cross-scale consistency of this chain and determines whether an additional whole-image cross-check is required before answer fusion. With Qwen3-VL-8B-Instruct as the common backbone, VERA-M achieves the highest values among the compared methods in seven of nine dataset–metric combinations across SLAKE, VQA-RAD, and ImageCLEF VQA-Med 2019. On SLAKE, it improves open-ended accuracy from 52.55% to 68.60% and yields gains across all evaluated metrics with each of three MLLM backbones. These results support the effectiveness of the proposed visual evidence construction and review for Med-VQA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_036.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=RXH2ZJFYig
BibTex
@InProceedings{CheQix_Agentic_MICCAISAT2026,
author = { Chen, Qixian AND Chen, Mingjin AND Zhang, Xin},
title = { { Agentic Visual Evidence Construction and Review Architecture for Medical VQA } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}