Enhancing Contrastive Learning for Retinal Imaging via Adjusted Augmentation Scales

Abstract

Contrastive learning, a typical self-supervised learning strategy, operates on bringing similar data together while pushing dissimilar data apart in latent space. This approach extracts robust and discriminative representations, thus being widely used in natural computer vision tasks, such as object classification. However, unlike natural images, medical images (e.g., retinal images) tend to share substantial similarities in imaging area and anatomical tissues, leading to a denser distribution in latent space. As a result, the default use of strong augmentations in contrastive learning potentially exacerbates this intensive distribution in retinal images, making it difficult to distinguish between genuinely similar and dissimilar data, and therefore hindering model pre-training convergence. In this paper, we hypothesise that weaker augmentations are better suited to contrastive learning for medical image applications, and we investigate model performance under various augmentation strategies. Our study includes six publicly available retinal datasets covering multiple clinically relevant tasks. We assess the models' performance and generalizability via extensive experiments. The model pre-trained with weak augmentation outperforms the one pre-trained with strong augmentation, achieving approximately a 6\% increase in AUPR ($P$<$0.001) and a 12.5\% increase in sensitivity ($P$<$0.001) on MESSIDOR-2. Similar improvements are observed across other datasets. Our findings suggest that optimizing the scale of augmentation is critical for enhancing the efficacy of contrastive learning in medical imaging. The model weights and relevant code are available at: https://github.com/ziijiecheng/Enhance-contrastive-SSL-for-Retinal-Imaging.