Attention residual convolution neural network based on U-net (AttentionResU-Net) for retina vessel segmentation
Shun Zhao, Tao Liu, Bowen Liu, Kun Ruan
Abstract
Abstract Extraction of retinal vessels from fundus images is one of the basic steps for diagnosis of diabetic retinopathy. Although some scholars have proposed several segmentation methods, this segmentation is still challenging due to differences in retinal vascular network and image quality. At present, the main challenges of retinal vascular segmentation are noise (due to uneven light) and capillaries. Due to the high complexity of retinal vascular characteristic information, the existing algorithms exists some problems such as microvascular segmentation and optic disc segmentation, We propose a network segmentation model based on fusion residual block, Attention mechanism and U-net, Firstly, the image was enhanced by limiting contrast histogram equalization. Secondly, Gamma correction is used to improve image brightness information and reduce artifact interference. Finally, AttentionResU-et segmentation model is used for segmentation. This algorithm was tested on DRIVE dataset, and its AUC reached 97.93%.