Residual Module and Multi-scale Feature Attention Module for Exudate Segmentation

Abstract

Exudates are primary signals of diabetic retinopathy (DR) which has become the leading cause of blindness in diabetic patients. Therefore, exudate segmentation is of great significance for disease screening and clinical diagnosis. The paper proposes an end-to-end method for the detection and segmentation of exudates in fundus images utilizing deep convolutional neural network (DCNN). Firstly, residual module is designed to extract more abundant exudate features. Besides, we consider fusing multi-scale feature through an improved feature pyramid attention module called Multi-scale Feature Attention module. The proposed method has been evaluated on publicly available DIARETDB1 v2, HEI-MED and Messidor datasets. It achieves better results than many existing methods in terms of AUC metrics, which make it befitting for practical clinical applications.