How to design a deep neural network for retinal vessel segmentation: an empirical study

•We perform an empirical study of retinal vessel segmentation from four aspects.•From empirical study, we derive a well-designed neural network framework.•We achieve state-of-the-art performance in three retinal vessel segmentation benchmarks. Retinal vessel segmentation is a critical step towards the accurate visualization, diagnosis, early treatment and surgery planning of ocular diseases. However, due to limited densely annotated data, inhomogeneous lighting and poor target contrast, the accurate segmentation of the retinal vessels remains challenging. This paper presents an empirical study of retinal vessel segmentation in data processing, architecture design, attention mechanism and regularization strategy, which reveals several best practices for producing state-of-the-art retinal vessel segmentation performance. Some of them are first investigated in retinal vessels segmentation, and then thoroughly analyzed with various parameters. Furthermore, based on our empirical study, a well-designed deep neural network is proposed for retinal vessel segmentation on three digital retinal images benchmarks: DRIVE, STARE and CHASE_DB1. In addition, to verify the generalization capability of the proposed framework on different publicly available datasets, a cross-dataset evaluation between DRIVE and STARE is performed. Extensive experimental results suggest that our proposed framework not only achieves state-of-the-art performance among existing methods with AUCs of 98.279%, 98.862%, 98.724% on DRIVE, STARE, and CHASE_DB1, respectively, but also exhibits stronger generalization than previous models in retinal vessel segmentation.