Abstract
Multi label retinal disease classification from colour fundus photography is challenging because retinal findings can co occur, appear at different spatial locations, and follow a highly imbalanced distribution. This paper presents ReGraFT, a Retinal Graph and Region Fusion Transformer for broad multi label fundus disease recognition. ReGraFT combines a ConvNeXt Tiny visual encoder with multi scale feature fusion, learned region token pooling, label to region decoding, direct spatial attention, graph based label refinement, and independent label prediction heads. We also propose MuRFiD, a harmonised Multi Label Retinal Fundus Image Dataset constructed from BRSET and MuReD. MuRFiD contains 18,474 fundus images and 21,064 positive annotations, with shared and source specific disease labels, strong class imbalance, and label co occurrence patterns. On MuRFiD, ReGraFT achieved the best overall performance among convolutional, transformer, self supervised, vision language, and retinal foundation model baselines, with an F1 score of 0.7475, precision of 0.7245, recall of 0.7719, mAP of 0.7914, and AUC of 0.9797. ReGraFT also required only 31.51M parameters and 11.41 GFLOPs, substantially less than RETFound. Class wise results and qualitative explanation maps further indicate that ReGraFT can model both common and rare retinal findings under a broad and imbalanced multi disease setting. (MuRFiD and the implementation code will be released upon paper acceptance.)
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_024.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=EPn15JRZII
BibTex
@InProceedings{BilHaz_ReGraFT_MICCAISAT2026,
author = { Bilal, Hazrat AND Keles, Ayse AND Kafieh, Raheleh AND Bendechache, Malika},
title = { { ReGraFT: Broad Multi Label Retinal Disease Classification on MuRFiD } },
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}
}