Comparison of self-supervised learning methods for optical coherence tomography image classification

Abstract

Self-supervised learning enables deep learning algorithms to achieve high performance even with limited labeled data. Numerous self-supervised learning methods have been proposed, and several medical artificial intelligence studies have confirmed the effectiveness of this technique in enhancing model performance. Nonetheless, only a few studies have investigated which self-supervised learning methods are appropriate for medical images. In this study, a comparative analysis to determine the most suitable self-supervised learning methods for developing medical artificial intelligence models using optical coherence tomography images. Four self-supervised learning methods were tested, and models were developed and validated using two publicly available optical coherence tomography datasets. The reliability of the model output was assessed through Gradient-weighted Class Activation Mapping analysis. As a result, using contrastive learning-based self-supervised learning methods, better performance was observed compared to other models, resulting in 99.90% and 99.05% accuracy for the two public optical coherence tomography datasets, which represents a 0.8% and 2.86% improvement over conventional supervised learning training. In addition, we confirmed that the contrastive learning-based model can more reliably localize disease lesions in the Gradient-weighted Class Activation Mapping analysis. © 2024 Copyright held by the owner/author(s)