Abstract
Automatic extraction of clinical retinal biomarkers relies on artery-vein (A/V) segmentation methods, typically trained and evaluated using overlap metrics, that lack clinical interpretability and do not reflect biomarker fidelity. Additionally, most A/V segmentation models are benchmarked on single-center datasets that are not representative of acquisition and population variety across clinical settings. Here, we present OCULAR (Open-source Collection of Unified Large-scale Artery-vein Retinae), a curated and carefully harmonized aggregation of 1,200 publicly available A/V annotations spanning diverse imaging conditions. OCULAR is an order of magnitude larger than previously proposed retinal A/V datasets, and contains an in-distribution split together with several out-of-distribution test sets. Leveraging OCULAR, we develop OCULAR-Net, a model trained by focusing on clinically relevant vascular regions and the preservation of biomarkers, rather than optimizing standard overlap losses. Across all OCULAR’s OoD test datasets, OCULAR-Net substantially improves agreement between ground-truth and segmentation-derived biomarkers, including topological complexity, junction geometry and large-vessel calibers. Remarkably, OCULAR-Net does not reach the best performance in terms of standard pixel-wise overlap metrics (e.g., Dice), but it clearly outperforms an array of state-of-the-art models in clinical biomarker fidelity, a metric much closer to real-life usefulness. Our results highlight the need for introducing domain knowledge and clinically-oriented evaluation in retinal A/V segmentation in order to ensure potential for clinical translation. OCULAR-Net and the OCULAR dataset, together with an expanded set of 1,800 refined A/V annotations derived from existing public binary vessel segmentation datasets, will be released at https://github.com/GonzaloPlaaza/OCULAR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_030.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=UmQjRhJdXT
BibTex
@InProceedings{PlaGon_Beyond_MICCAISAT2026,
author = { Plaza, Gonzalo AND Sanchez, Maria AND Serra-Castanera, Alicia AND Valls-Esteve, Arnau AND Mata Miquel, Christian AND Camara, Oscar AND Galdran, Adrian},
title = { { Beyond Dice: Clinically Meaningful Large-Scale Retinal Artery/Vein Segmentation } },
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}
}