CAGU-Net: Category Attention Guidance U-Net for Retinal Blood Vessel Segmentation

Abstract

The characteristics of retinal blood vessels are the basis for physicians to diagnose cardiovascular diseases such as diabetes and hypertension. The accurate segmentation of retinal blood vessels has crucial clinical medical significance. In this paper, we propose a novel retinal vessel segmentation network, which is a Category Attention Guidance U-Net (CAGU-Net) to alleviate the problems of unbalanced sample categories and low contrast in retinal fundus images. Firstly, a Category Attention Guidance (CAG) module is proposed to build a category attention mechanism that provides global guidance for pixel classification. Secondly, a deep supervision strategy is introduced to supervise the encoder to learn more detailed semantic information. Finally, experimental results show that the proposed method can achieve advanced performance in retinal fundus images and outperform the state-of-the-art methods.