Abstract

Info

Title:: Multi-layer feature attention enhanced network for diabetic retinopathy staging

Author(s):: LIANG Bingxue1; WANG Wenjing1; WANG Haoqi2; GUAN Quan1; QIN Yuhua1; 1. School of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China 2. School of Business, East China University of Science and Technology, Shanghai 200237, China

Keywords:: Keywords: diabetes retinopathy image classification feature fusion computer-aided diagnosis

Abstract:: Abstract: A multi-layer feature attention enhanced network is proposed to further improve the diagnostic accuracy of the severity of diabetic retinopathy. To address the inconsistent expression of global and local features when processing diabetic retinopathy images, a dual-branch parallel model combining ResNet-50 and DeiT-S is employed as the backbone architecture, and a feature fusion module is designed at the end of the network. Concurrently, a multi-scale location awareness enhancement module is developed to extract multi-scale information through dilated convolution with positional attention mechanism for enhancing the feature representation of lesions in fundus images, and a local feature enhancement module is constructed to strengthen the models capability in extracting local information, thus improving models capability to identify small lesions and minor changes. The experimental results show that the proposed multi-layer feature attention enhanced network achieves an accuracy of 87.61%, exhibiting excellent classification performance. This advancement provides a strong support for further development of diabetic retinopathy detection technology.