Abstract
Medical vision-language models hold considerable promise for AI-assisted diagnosis and treatment, yet their progress remains constrained by the quality of available datasets. Existing resources suffer from coarse answer granularity, oversimplified evaluation, or image-text alignment noise in large-scale pretraining corpora. To address these limitations, we build PSV2026, a 61K medical VQA dataset for model training and fine-grained factuality evaluation. The training split is built by refining existing VQA data and simulating how clinicians learn from errors to introduce additional supervisory signals. Each test sample is equipped with five question-specific scoring criteria, enabling fine-grained measurement of clinical factuality. Clinician evaluation confirms that both the dataset and the LLM-as-a-Judge strategy achieve high agreement with expert judgments. Building upon PSV2026, we propose MOFT, a lightweight medical factuality-oriented fine-tuning method that conducts multi-objective optimization through medical entity weighting and parameter-space fusion. Experiments reveal that current medical and general VLMs still leave substantial headroom on complex tasks such as diagnostic inference, and that both the PSV2026 and MOFT transfer effectively to the external OmniAbnormalCT dataset.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{LiYuc_PSV2026_MICCAISAT2026,
author = { Li, Yuchong AND Zeng, Xiaojun AND You, Caizhen AND Wu, Pengbo AND Guo, Zixian AND Yang, Jian AND Jia, Fucang AND Zhang, Lei},
title = { { PSV2026: A Fine-Grained Clinical Factuality Dataset for Medical Vision-Language Models } },
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}
}