Retinal Blood Vessel Segmentation via Attention Gate Network

Automatic retinal vessel segmentation is a challenging problem in the clinical diagnosis of eye diseases. Accurate segmentation of retinal vessel can efficiently assist the physicians to make a more precise symptom detection. However, there are various shapes and sizes, complex backgrounds and noise in the retinal vessel images. To address these problems, in this paper, we design an attention gate network to model long-range dependencies and capture rich contextual information. Specifically, we adopt an attention gate module, which includes a spatial attention module to model spatial long-range contextual information. Moreover, to improve the contrast of original retinal fundus images, we employ green channel extraction and contrast limited adaptive histogram equalization as pre-processing steps. Experiments on the DRIVE and STARE show the proposed AGNET achieves the outstanding performance with 0.8247/0.8361 sensitivity, 0.9871/0.9899 specificity, 0.9764/0.9791 accuracy, and 0.9881/0.9928 AUC respectively.