Abstract
Uveitis is a relatively uncommon and etiologically heterogeneous inflammatory eye disease, and infectious-versus-non-infectious classification is clinically important but challenging for direct vision-language model (VLM) classification. Using real-world fundus images from Tan Tock Seng Hospital, Singapore, we built a fixed 438-image benchmark and evaluated five VLM settings under zero-shot generation, LoRA-style adaptation, and frozen-feature linear probing, with image-only encoders as baselines. Zero-shot hard decisions collapsed to a single class across all VLM settings, yielding 0.5000 balanced accuracy, yet several score streams showed non-random AUC. LoRA adaptation was limited and inconsistent, with one MedGemma-27B setting reaching 0.6441 accuracy and 0.6431 balanced accuracy. Frozen-feature probing better converted encoded representations into classification performance, reaching 0.7797 accuracy and 0.7954 AUC for LLaVA-NeXT and 0.7627 accuracy and 0.8391 AUC for InternVL3.5. In selected rows and metrics, these linear-head results were numerically higher than the strongest image-only baseline, DenseNet121, which achieved 0.7458 accuracy and 0.8230 AUC. A secondary threshold analysis of two preserved score streams further suggested recoverable score-level information. These findings reveal a score/representation–decision mismatch in uncommon, fine-grained ophthalmic classification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_032.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=1ngdAptAn4
BibTex
@InProceedings{WanYun_Benchmarking_MICCAISAT2026,
author = { Wang, Yuning AND Zhong, Zihao AND Diao, Yueqin AND Liu, Huiying AND Fang, Jiansheng AND Li, Kai AND Zhang, Haoran AND Xu, Yanwu AND Agrawal, Rupesh},
title = { { Benchmarking Vision-Language Models for Image-Based Infectious Versus Non-Infectious Uveitis Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}