Automatic Detection of Peripheral Retinal Lesions From Ultrawide-Field Fundus Images Using Deep Learning

Abstract

Purpose: To establish a multilabel-based deep learning (DL) algorithm for automatic detection and categorization of clinically significant peripheral retinal lesions using ultrawide-field fundus images. Methods: A total of 5958 ultrawide-field fundus images from 3740 patients were randomly split into a training set, validation set, and test set. A multilabel classifier was developed to detect rhegmatogenous retinal detachment, cystic retinal tuft, lattice degeneration, and retinal breaks. Referral decision was automatically generated based on the results of each disease class. t-distributed stochastic neighbor embedding heatmaps were used to visualize the features extracted by the neural networks. Gradient-weighted class activation mapping and guided backpropagation heatmaps were generated to investigate the image locations for decision-making by the DL models. The performance of the classifier(s) was evaluated by sensitivity, specificity, accuracy, F1 score, area under receiver operating characteristic curve (AUROC) with 95% CI, and area under the precision-recall curve. Results: In the test set, all categories achieved a sensitivity of 0.836–0.918, a specificity of 0.858–0.989, an accuracy of 0.854–0.977, an F1 score of 0.400–0.931, an AUROC of 0.9205–0.9882, and an area under the precision-recall curve of 0.6723–0.9745. The referral decisions achieved an AUROC of 0.9758 (95% CI= 0.9648–0.9869). The multilabel classifier had significantly better performance in cystic retinal tuft detection than the binary classifier (AUROC= 0.9781 vs 0.6112, P < 0.001). The model showed comparable performance with human experts. Conclusions: This new DL model of a multilabel classifier is capable of automatic, accurate, and early detection of clinically significant peripheral retinal lesions with various sample sizes. It can be applied in peripheral retinal screening in clinics.