Abstract
Background: Pathologic myopia (PM) associated with myopic maculopathy (MM) and "Plus" lesions is a major cause of irreversible visual impairment worldwide. Therefore, we aimed to develop a series of deep learning algorithms and artificial intelligence (AI)-models for automatic PM identification, MM classification, and "Plus" lesion detection based on retinal fundus images. Materials and Methods: Consecutive 37,659 retinal fundus images from 32,419 patients were collected. After excluding 5,649 ungradable images, a total dataset of 32,010 color retinal fundus images was manually graded for training and cross-validation according to the META-PM classification. We also retrospectively recruited 1,000 images from 732 patients from the three other hospitals in Zhejiang Province, serving as the external validation dataset. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and quadratic-weighted kappa score were calculated to evaluate the classification algorithms. The precision, recall, and F1-score were calculated to evaluate the object detection algorithms. The performance of all the algorithms was compared with the experts' performance. To better understand the algorithms and clarify the direction of optimization, misclassification and visualization heatmap analyses were performed. Results: In five-fold cross-validation, algorithm I achieved robust performance, with accuracy = 97.36% (95% CI: 0.9697, 0.9775), AUC = 0.995 (95% CI: 0.9933, 0.9967), sensitivity = 93.92% (95% CI: 0.9333, 0.9451), and specificity = 98.19% (95% CI: 0.9787, 0.9852). The macro-AUC, accuracy, and quadratic-weighted kappa were 0.979, 96.74% (95% CI: 0.963, 0.9718), and 0.988 (95% CI: 0.986, 0.990) for algorithm II. Algorithm III achieved an accuracy of 0.9703 to 0.9941 for classifying the "Plus" lesions and an F1-score of 0.6855 to 0.8890 for detecting and localizing lesions. The performance metrics in external validation dataset were comparable to those of the experts and were slightly inferior to those of cross-validation. Conclusion: Our algorithms and AI-models were confirmed to achieve robust performance in real-world conditions. The application of our algorithms and AI-models has promise for facilitating clinical diagnosis and healthcare screening for PM on a large scale.