Abstract

Info

Title:: Eye fundus image classification using hyperparameter optimization based TransCNN

Author(s):: WANG Xiaofang1; YU Kexin2; WANG Zhangyi2; WANG Jianhua2; WANG Jing3; MU Nan4; 1. School of Intelligence Technology, Geely University of China, Chengdu 641423, China 2. School of Computing, Chengdu College of University of Electronic Science and Technology of China, Chengdu 611731, China 3. Chengdu Medical Plus Medical Optics Co., Ltd, Chengdu 610031, China 4. College of Computer Science, Sichuan Norman University, Chengdu 610066, China

Keywords:: Keywords: fundus image differential evolution algorithm TransCNN hyperparameter optimization MEB-KSVM

Abstract:: Abstract: A novel algorithm (hyperparameter optimization based TransCNN, Deho-TransCNN) for fundus image classification is proposed to solve the problems of the multiple parameters, randomness and long training time in deep learning model, overlapping lesions in fundus images, and uneven samples in data sets. Based on TransCNN model, the algorithm uses differential evolution algorithm to initialize the weight of model network, combines the hyperparameters to achieve adaptive optimization, and complete the fundus image multi-classification using MEB-KSVM. The experimental results show that the proposed algorithm has an accuracy, sensitivity, specificity and AUC value of 0.947, 0.926, 0.937 and 0.945 for image multi-classification, with average improvements of 5.6%, 6.4%, 5.1% and 7.9% as compared with the other 9 traditional algorithms. The average detection time of the proposed algorithm is 158.3% lower than that of the optimal CNN. The improved algorithm can enhance the performance of image multi-classification, reduce the time required for image detection, and has certain generalization ability for image multi-classification.