Self-Supervised Learning for Early-Stage Glaucoma Detection Using Fundus Images

Abstract

Deep learning techniques have proven to be rather effective as far as the field of medical image analysis is concerned. However, a major limitation to the clinical usage of such solutions is that they require large amounts of annotated data, a limitation that is especially burdensome in the domain of ophthalmology, in which manual annotation is both time-consuming and expensive. We propose in this manuscript a self-supervised deep learning model that detects glaucomatous retina fundus during retina fundus photography to reduce the use of a labeled data set, without negatively impacting or marginally, but significantly, improving diagnostic performance. The suggested methodology is a two-step training procedure. First, the ResNet-18 encoder is trained through contrastive learning with a SimCLR architecture on a set of unlabeled fundus images, which in turn, makes it possible to extract semantically rich image representations. Afterward, an encoder is frozen, and a lightweight classifier is trained using a small number of glaucoma cased. Analysis of the ACRIMA, Drishti_GS and Rim-One-R2 datasets shows that our model is doing an excellent job of weak supervision. Comparative experiments attest that self-supervised pretraining is more effective than training frozen, with an area under the receiver operating characteristic curve (AUC) of 0.947 (94.7%). These results show that self-supervised learning is an opportunity that has the potential to build clinically deployable models in a situation where annotated medical imagery is limited.