Abstract

Info

Title:: Automatic diagnosis of diabetic retinopathy based on interpretable features fusion

Author(s):: JIANG Jiewei1; LEI Shutao2; GENG Miaomiao1; GONG Jiamin1; 2; ZHU Zehao2; ZHANG Yunsheng1; LIU Fang2; WU Yijie2; WANG Yuwen3; LI Zhongwen3; 1. School of Electronic Engineering, Xian University of Posts & Telecommunications, Xian 710121, China 2. School of Communication and Information Engineering, Xian University of Posts & Telecommunications, Xian 710121, China 3. Ningbo Eye Hospital, Wenzhou Medical University, Ningbo 315000, China

Keywords:: Keywords: diabetic retinopathy deep learning ensemble learning interpretable model support vector machine

Abstract:: Abstract: Diabetic retinopathy (DR) has become one of the 4 major common causes of blindness, and the early diagnosis can effectively reduce the risk of visual impairment. By the fusion of the interpretable features of deep learning, an automatic diagnosis method of DR is proposed in the study. After different labeled lesion images are generated by two interpretable techniques, namely guided gradient-weighted class activation mapping and saliency map, the feature vectors of the original image and the two generated images are extracted via convolutional neural networks. Finally, the fusion of 3 kinds of feature vectors is carried out, and the results are input into support vector machine for realizing the automatic diagnosis of DR. For the data set with 1 443 color fundus images, compared with those of basic ResNet50 model, the accuracy, specificity, sensitivity, precision and Kappa coefficient of the proposed method are improved by 3.6%, 2.4%, 5.8%, 4.6% and 7.9%, respectively. The experimental results show that the proposed method can effectively reduce the risk of DR misdiagnosis.