Self-Supervised Learning in Medical Diagnostics: An Examination of SimCLR and BYOL in Image Classification

Supervised learning algorithms have accomplished significant advancements in the field of medical image classification. However, these methods highly rely on large labelled datasets and it is costly to obtain image datasets with proper labels. To solve this issue, contrastive learning aims to learn representations by contrasting positive and negative examples by maximising the similarity between positive pairs and minimising the similarity between negative pairs. In this paper, the performance of two prominent contrastive self-supervised learning (SSL) frameworks, SimCLR and BYOL, will be compared in the specific context of medical image classification using the CheXpert dataset. ResNet-18, is used as the model backbone for comparison purposes for its low compute requirements and common usage for image classification. The result shows that SimCLR provides the best performance compared to BYOL with hardware constraints. However, neither of which outperform current supervised models due to limited hyperparameter configurations. The objectives of this paper is to overcome the challenges of limited availability and high costs of annotated data in medical image classification by using self-supervised learning techniques, to enhance the performance of image classification models by effectively utilising unlabelled medical images, and to assess the performance and drawbacks associated with different self-supervised learning frameworks.