Edge-aware U-net with gated convolution for retinal vessel segmentation

Abstract

Morphological changes of retinal vessels, especially tiny vessels play an essential role in the diagnosis and clinical prognosis of specific cardiovascular and ophthalmic diseases. However, the precise segmentation of retinal vessels, particular the vessel terminals of the retinal vascular branches, tends to be poor. In this proposed approach, we introduce new edge-aware flows into U-Net encoder-decoder architecture to guide the retinal vessel segmentation, which makes segmentation more sensitive to the fine edges of the capillaries. The edge-gated flow with gated convolution only focuses on edge presentation and learns to emphasize the vessel edges using features extracted from the encoder path, and exports the edge prediction results. The edge-downsamling flow then extracts the edge features from the edge prediction results and feeds them back into the decoder path to refine the segmentation results. The proposed method achieves state-of-the-art performance in term of accuracy of 0.9701, 0.9691, and 0.9811 and gains an increment of 0.0056, 0.0026, and 0.0047 compared with U-Net baseline on three publicly available datasets: DRIVE, STARE, and CHASEDB1, respectively. The experimental results show that the proposed Edge-Aware U-Net is an effective architecture that provides more accurate segmentation around the vessel edges and significantly boosts the performance on tiny vessels.