Abstract
Medical Visual Question Answering (Med-VQA) is a cornerstone task for the advancement of AI-assisted medical image analysis.
Current state-of-the-art methods are primarily based on Large Vision Language Models (LVLMs).
However, these models face significant limitations in tasks requiring fine-grained visual grounding and spatial reasoning, such as precise measurement and counting.
Alternatively, the Visual Programming paradigm operates at the intersection of multimodal reasoning and agentic AI.
By leveraging code generation to answer visual questions, it explicitly disentangles reasoning and execution.
We argue that its improved auditability and complementary capabilities compared to end-to-end models are of great interest for medical applications.
To this end, we introduce MeVisQA, an off-the-shelf Med-VQA framework based on Visual Programming.
Additionally, inspired by radiologist workflows, we present SliceVQA, a novel benchmark designed to assess capabilities empowered by the Visual Programming approach,
including quantitative measurements, multi-step reasoning, and multi-image analysis.
Our experiments demonstrate that MeVisQA achieves an almost two-fold accuracy improvement over recent medical LVLMs on this benchmark.
SliceVQA is available at https://github.com/NdkArthur/SliceVQA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SAFER_007.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=gkmkZCrUq6
BibTex
@InProceedings{HonFan_MeVisQA_MICCAISAT2026,
author = { Hong, Fan AND Anselmet, Marie AND Ndoko, Arthur},
title = { { MeVisQA: Medical Visual Question Answering via Visual Programming } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}