Preprint: Shape-Aware Deep Segmentation in Retinal Vessel Imaging from Sparse and Weak Labels

Abstract

Retinal vessel segmentation in optical coherence tomography (OCT) is challenging because dense annotations require expert knowledge and are time consuming. We investigated shape-aware deep segmentation from sparse vessel center annotations covering only a subset of visible vessels. Using UNet++ as baseline, we evaluated three shape priors: SPM, SPGNet, and an HBS-inspired loss. All methods were compared on a generated proxy dataset with dense masks. SPM achieved the highest validation scores and was subsequently evaluated on real OCT using blob weak labels derived from sparse annotations. Although Dice, Gen eralized Dice, and mean Intersection over Union favored the conservative UNet++ baseline against blob labels, visual and quantitative foreground analyses showed that SPM predicted more vessel-like foreground and connected components and achieved higher weak-label recall. These results indicate that SPM can produce plausible vessel-like predictions beyond weak-label coverage, while overlap metrics alone cannot establish anatomical correctness from sparse annotations