A Classification Method for Diabetic Retinopathy Based on Self-supervised Learning

Diabetic retinopathy, a prevalent eye disease in diabetic patients, poses a high risk of blindness. Current computer-aided diagnostic methods require extensive labeled datasets, which are labor-intensive and time-consuming. Given the scarcity of large labeled datasets in medical imaging, self-supervised learning presents a promising alternative, capable of leveraging unlabeled data to improve diagnostic accuracy. Addressing this challenge, our paper introduces a classification method for diabetic retinopathy using self-supervised learning. The method encompasses three stages: data preprocessing, self-supervised learning, and classification. In preprocessing stage, fundus images are resized to 255 x 255 pixels, divided into nine sub-images each, and labeled with serial numbers. Additionally, masks are applied to sub-image edges to minimize model reliance on edge features. The self-supervised learning stage employs a VGG16-based network with nine branches to learn intrinsic features of fundus images, thus decreasing the requirement for labeled samples. The network inputs are these sub-images, shuffled in order, with output being their sequence numbers. This stage produces a pre-trained network. For the classification stage, this pre-trained network is further fine-tuned using a small labeled dataset (300 images), modifying the final fully connected layer for either binary or five-category classification. Our approach has demonstrated impressive results: 92.5% accuracy for binary and 66.7% for five-category classification on APTOS dataset, and 92.7% for binary and 62.4% for five-category classification on the Kaggle EyePACS dataset. These outcomes underscore the substantial potential of self-supervised learning in diabetic retinopathy diagnosis, offering a significant reduction in the dependency on extensive labeled datasets and thereby enhancing the diagnostic process.