Myopic-Net: Deep Learning-Based Direct Identification of Myopia Onset and Progression.

Purpose: Identifying and monitoring the onset and progression of myopia (myopia onset and progression [MOP]) based on the changes in anatomical structures in fundus retinal images has significant clinical application prospects. For this purpose, we tested the performance of deep neural networks. Methods: We established a deep neural network, called Myopic-Net, to detect anatomical changes owing to the MOP from a pair of retinal images collected during different fundoscopies. Myopic-Net was developed using 3964 fundus image pairs without MOP and 2380 fundus image pairs with MOP. Five indicators-accuracy, precision, recall, specificity, and F1-score-were evaluated on the internal test set and the independent external test set. In addition, we use a deep network visualization method to explore the factors driving Myopic-Net to predict. Results: On the internal test set, Myopic-Net achieved an accuracy of 87.3%; the precision, recall, and specificity were 86.2%, 80.1%, and 91.9% respectively, while the identification accuracy of two ophthalmologists is only 66.1% and 73.5%, respectively. Even on the external test set, Myopic-Net still achieved an accuracy of 84.1%. In addition, we found that the factors driving Myopic-Net to predict are mainly anatomical changes in the optic disc and surrounding areas. Conclusions: Myopic-Net has been shown to be able to identify the MOP from fundus image pairs using anatomical changes in optic disc and surrounding areas. And Myopic-Net has good accuracy, reliability, and generalization ability. These factors show that deep neural networks have strong potential in monitoring and final diagnosing the MOP based on fundus image analysis. Translational Relevance: With the development of fundus imaging technology based on intelligent mobile terminals, embedding the program based on Myopic-Net has great potential to achieve convenient and fast personalized monitoring of myopia.