Abstract
Most retinal vision-language foundation models are adapted through global image-text alignment, which can miss the subtle, localized findings decisive for detecting early retinal disease. Naively fine-tuning such a model (FLAIR) on a narrow, single-disease dataset risks degrading its general representations, so we propose Pixel-FLAIR, which adds a region-level text supervisory signal that approximates pixel-level text supervision alongside the global one. An off-the-shelf segmentation model locates anatomical landmarks (optic disc and cup) that anchor the automatic cropping of clinically relevant sub-regions; these are paired with expert text descriptors, yielding region-text supervision with no manual region annotation. A modified contrastive loss injects a within-sample, bidirectional region-alignment term into the diagonal of the similarity matrix, preserving the pretrained contrastive geometry. Across REFUGE, MESSIDOR, and FIVES, region supervision with weight interpolation consistently improves performance, most clearly in the few-shot linear-probe setting.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_007.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{AlaSas_PixelFLAIR_MICCAISAT2026,
author = { Alavala, Sasidhar AND Kamalesh, Akash AND Seelamantula, Chandra Sekhar},
title = { { Pixel-FLAIR: Leveraging Anatomical Segmentation for Region-Specific Supervision in Retinal Foundational 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}
}