Robust semi-automatic vessel tracing in the human retinal image by an instance segmentation neural network

Vasculature morphology and hierarchy are essential for blood perfusion. Human retinal 20 circulation is an intricate vascular system emerging and remerging at the optic nerve head 21 (ONH). Tracing retinal vascular branching from ONH can allow detailed morphological 22 quantification, and yet remains a challenging task. We presented a robust semi-automatic 23 vessel tracing algorithm on human fundus images by an instance segmentation neural 24 network (InSegNN). InSegNN separates and labels individual vascular trees and enables 25 tracing each tree throughout its branching. We have three strategies to improve robustness 26 and accuracy: pseudo-temporal learning, spatial multi-sampling, and dynamic probability 27 map. We achieved 83% specificity, 50% improvement in Symmetric Best Dice (SBD) 28 compared to literature, and outperformed baseline U-net, and achieved 91% precision with 29 71% sensitivity. We have demonstrated tracing individual vessel trees from fundus 30 images, and simultaneously retain vessel hierarchy information. InSegNN paves a way for 31 subsequent analysis of vascular morphology in relation to retinal diseases.