A diagnostic method of high myopia based on transfer learning. Issue 1 (July 2021)

• Record Type: Journal Article
• Authors: Chen, Baibing; Wan, Cheng
• Is Part Of: Journal of physics. Volume 1976:Issue 1(2021)
• Publication Date: 2021-07
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1976/1/012031
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: In order to promote the development of computer-aided diagnosis, improve the efficiency of high myopia diagnosis, and help healthcare workers assist in diagnosis, this paper uses 6571 fundus color images of high myopia and 6212 normal color images from the Geriatric Hospital of Nanjing Medical University as the dataset. The initialization parameters pre-training on ImageNet are loaded into the VGG-16, VGG-19, ResNet-50, ResNet-101, Inception-V3, and EfficientNet-B0 models, Learning rate changes dynamically according to the number of training epochs, and saving the model with the highest area under curve (AUC) value during the training process. Taking the best classification model from different networks, this paper achieves classification accuracy of 93.86%, sensitivity of 0.9126, specificity of 0.9677 and AUC value of 0.9866 on the test set. Each image takes 0.039 seconds, which is within the acceptable range meeting real-time of medically assisted diagnosis. Results show the fine-tuned ResNet-101 network meets the needs of medical assist diagnosis and can be used for screening high myopia, with high accuracy and real-time performance.