Abstract
Diagnostic errors occur in an estimated 3–5% of medical imaging interpretations, amounting to approximately 40 million errors worldwide each year and potentially contributing to missed or delayed diagnoses and unintentional patient harm. Medical visual question-answering (VQA) aims to support image interpretation, but reliable performance under domain shift requires both i) precise closed-ended answers and ii) visually grounded open-ended explanations. However, these two VQA formats impose distinct optimization requirements because exact option selection differs substantially from free-form, image-grounded generation. To this end, we introduce GraDe (Grouped Adapter Decoding) to explicitly address the MICCAI 2026 MedReason challenge, by routing the two VQA formats through task-specific low-rank adapters attached to a shared Qwen2.5-VL-7B-Instruct backbone, allowing the 2 VQA formats to be optimized independently. Closed-ended predictions are decoded directly from the logits associated with the option letters, whereas open-ended reasoning is supervised only using evidence that a grounding judge confirms as visible in the image. GraDe is trained exclusively on MedReason’s 17,722 released cases, and evaluated on MedReason’s official hidden pre-evaluation set, achieving i) 87.38% multiple-choice accuracy, ii) open-ended ground-truth score of 1.398/4, and iii) visual-answer score of 2.256/4. Code and trained weights are available on GitHub and Zenodo.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedReason_009.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=QTzpYQdUeZ
BibTex
@InProceedings{KacSan_GrADe_MICCAISAT2026,
author = { Kachole, Sanket AND Bakas, Spyridon},
title = { { GrADe: Grouped Adapter Decoding for Medical Visual Question-Answering under Domain Shift } },
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}
}
