Abstract
Medical VQA combines discrete recognition with free-form evidence, but a shared adaptation objective can blur their requirements, while low-resolution preprocessing may remove diagnostic detail. We propose a parameter-efficient framework integrating data-calibrated native-detail preservation, concise image-grounded targets, and task-routed low-rank adapters. A general adapter is trained on all tasks while freezing the vision encoder and multimodal projector; an open-answer specialist is further trained on balanced free-form examples. At inference, question type deterministically selects the adapter, while schema-first generation and failure-triggered recovery isolate serialization errors. We evaluate Qwen3-VL on a 512-case holdout from 17,722 Med-CMR-derived cases across 12 modalities, 11 organ systems, and seven reasoning categories. Multiple-choice accuracy improves from 94.71% to 96.63% at 8B; 32B reaches 97.12% (404/416; exact paired p = 0.0129 versus baseline). On 96 open-ended cases, answer/evidence-trace token F1 rises from 0.231/0.240 to 0.253/0.250, although bootstrap intervals include zero. The 32B setup trains 536.9M parameters (∼ 1.68%), avoiding full-model fine-tuning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedReason_013.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=8MWHO7jusA
BibTex
@InProceedings{LuMen_TaskRouted_MICCAISAT2026,
author = { Lu, Mengkang AND Ma, Rongze AND Zeng, Qingjie AND Lu, Zilin AND Xia, Yong},
title = { { Task-Routed Low-Rank Adaptation with Grounded Supervision for Medical Visual Question Answering } },
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}
}