Abstract: Objective To promote the development of computer-aided diagnosis, and to improve the diagnosing efficiency of high myopia. Methods ResNeXt-50 network was used to diagnose high myopia with few parameters and fast training speed. It was used to distinguish normal fundus and high myopia fundus in this study. Results This study used 6571 high myopia color photos and 6212 normal color photos from Jiangsu Province Geriatric Hospital as the data set. In the end, the diagnostic method got an accuracy of 94.10%, a sensitivity of 92.33%, a specificity of 95.94% and an AUC of 0.9861 for high myopia. The average time of each image took 0.035 s, which was within the acceptable range of real-time diagnosis, which met the real-time performance of medical auxiliary diagnosis. Conclusions ResNeXt-50 has a good classification performance. It can diagnose high myopia efficiently and accurately.

Key words: high myopia, computer-aided diagnosis, convolutional neural network, image classification