Generalized Analysis of Vessels in Eye Edition 2

Retinal artery–vein (AV) identification plays a critical role in the diagnosis and monitoring of numerous ocular and systemic diseases, including diabetic retinopathy, retinal vascular occlusions, hypertension, and glaucoma. Quantitative retinal vascular biomarkers—such as vessel caliber, density, topological complexity, and the arteriovenous ratio—depend on accurate segmentation and classification of retinal arteries and veins from imaging data. Consequently, AV segmentation has become a foundational task in computational ophthalmology. However, existing datasets and algorithmic research almost exclusively focus on color fundus photographs (CFP), where arteries and veins often exhibit similar color intensities, illumination variations, and ambiguous crossings. These factors substantially limit the accuracy and robustness of CFP-only AV segmentation methods.

Fundus fluorescein angiography (FFA), by contrast, captures dynamic vascular perfusion and provides unique functional cues essential for distinguishing arteries from veins. Early-phase FFA highlights arterial filling, whereas late-phase FFA shows venous and mixed vessel filling patterns. These complementary temporal-functional characteristics offer strong supervisory signals that can significantly improve AV classification on CFP images. Building upon these insights, this year’s challenge introduces a novel cross-modal setting: leveraging paired FFA images to guide and enhance artery–vein segmentation on CFP. Participants are encouraged to explore multimodal fusion, cross-modal consistency learning, temporal-phase guidance, and representation alignment between structural (CFP) and functional (FFA) modalities.

In addition to AV segmentation, the challenge emphasizes clinically meaningful retinal vascular biomarker quantification. Based on the segmentation results, participants are tasked with automatically measuring a comprehensive set of vascular biomarkers, including central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arteriolar-to-venular ratio (AVR), artery and vein densities, as well as artery and vein fractal dimensions. Together, these metrics capture complementary aspects of vessel caliber, spatial distribution, and vascular topology, providing a richer characterization of retinal vascular health.

This challenge provides a curated dataset of 150 CFP–FFA pairs with expert annotations for AV segmentation. In the preliminary stage, 50 labeled cases will be released for model training, followed by 50 additional labeled cases for validation. An online evaluation platform will allow participants to iteratively refine their models based on leaderboard rankings. In the final stage, the remaining 50 cases will be used for blind evaluation. From a technical standpoint, this challenge advances research in multimodal medical image segmentation and cross-domain vascular analysis. From a biomedical perspective, it aims to improve the precision of retinal vascular assessment and facilitate downstream computation of clinically meaningful biomarkers. By integrating CFP with early- and late-phase FFA signals, this challenge promotes the development of next-generation AV segmentation and vascular quantification methods that are more reliable, interpretable, and clinically applicable.