基于超参数优化的TransCNN眼底图像分类算法  

Eye fundus image classification using hyperparameter optimization based TransCNN

作　　者：王小芳 余柯欣 王张怡 王剑华 王静 穆楠 WANG Xiaofang;YU Kexin;WANG Zhangyi;WANG Jianhua;WANG Jing;MU Nan(School of Intelligence Technology,Geely University of China,Chengdu 641423,China;School of Computing,Chengdu College of University of Electronic Science and Technology of China,Chengdu 611731,China;Chengdu Medical Plus Medical Optics Co.,Ltd,Chengdu 610031,China;College of Computer Science,Sichuan Norman University,Chengdu 610066,China)

机构地区：[1]吉利学院智能科技学院,四川成都641423 [2]电子科技大学成都学院计算机学院,四川成都611731 [3]成都医加医光学有限责任公司,四川成都610031 [4]四川师范大学计算机科学学院,四川成都610066

出　　处：《中国医学物理学杂志》2023年第6期672-682,共11页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(62006165);成都市科技局项目(2018-YFYF-00191-SN);四川省民办教育协会项目(MBXH21YB119);吉利学院科研项目(2022xzky004)。

摘　　要：针对深度学习模型参数多、随机、训练时间长,眼底图像病变处交织重叠、数据集样本不均等问题,提出基于超参数优化的TransCNN(Deho-TransCNN)眼底图像分类算法。该算法以TransCNN网络模型为基础,利用差分进化算法分别对模型网络权重进行初始化寻优和对模型进行超参数组合实现参数自适应优化;最后利用MEB-KSVM对眼底病变图像进行多分类。实验结果表明,改进算法的准确率、敏感性、特异性以及AUC值最优,分别为0.947、0.926、0.937、0.945,与文中9种传统算法比较,分别平均提升5.6%、6.4%、5.1%、7.9%;改进算法检测时间最低,与最佳算法改进CNN相比,平均检测时间降低158.3%。改进算法在一定程度上提升图像多分类效果,降低图像检测时间,对图像多分类处理有一定泛化能力。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.

关 键 词：眼底图像 差分进化算法 TransCNN 超参数优化 MEB-KSVM 

分 类 号：R318[医药卫生—生物医学工程] R744.1[医药卫生—基础医学]