Detecting Retinopathy From Optical Coherence Tomography Images Using a Novel Augmentation-Based Semi-Supervised Learning Approach

Optical coherence tomography (OCT) facilitates the diagnosis of retinal diseases with a noninvasive and rapid imaging solution. When automatically detecting retinopathy in OCT images using machine learning (ML), semi-supervised learning (SSL) technologies can reduce the burden on data annotation which is expensive and tedious. The conventional pseudo-label-based SSL often suffers from an accuracy decline due to confirmation bias. To address this, we propose an augmentation-based SSL method for the automatic detection of retinopathy in OCT images. Specifically, we employ both weak and strong augmentations on unlabeled images and learn a pretrained deep network. Pseudo-labels are derived from predictions on weakly augmented OCT images, which diminishes confirmation bias and enhances the model's robustness. We validate our method on two public OCT image datasets with 2% labels and 0.1% labels, respectively, precisely detecting four types of retinal diseases. The experimental results show that we surpass existing state-of-the-art models in the retinopathy classification.