A Data-Efficient Approach for Automated Classification of OCT Images Using Generative Adversarial Network

Abstract

Deep learning algorithms can offer a reliable automated interpretation of retinal optical coherence tomography (OCT) images to assist clinicians in disease diagnosis and management. However, retinal image processing presents pertinent obstacles such as the struggle of large-scale data acquisition and high cost of annotation. To address this, we have developed a data-efficient semisupervised generative adversarial network based classifier for automated diagnosis with limited labeled data. The framework consists of a generator and a discriminator. The adversarial learning between them assists in building a generalizable classifier to predict progressive retinal diseases like age-related macular degeneration and diabetic macular edema. Experimental results on clinical-grade OCT images show an overall improvement of more than 10% in accuracy, compared to the state-of-the-art methods.