SimCLR-based Self-Supervised Learning Approach for Limited Brain MRI and Unlabeled Images

Abstract

In this study, a SimCLR-based model is proposed for the classification of unlabeled brain tumor images in medical imaging using a self-supervised learning (SSL) technique. Additionally, the performances of different SSL techniques (Barlow Twins, NnCLR, and SimCLR) are analyzed to evaluate the performance of the proposed model. Three different datasets, consisting of pituitary, meningioma, and glioma brain tumors as well as non-tumor images, were used as the dataset. Out of a total of 7,671 images, 6,128 were used as unlabeled data, and the model was trained with both labeled and unlabeled data. The proposed model achieved high performance with unlabeled data, reducing the need for manual labeling. As a result, the model demonstrated superior performance compared to other models, with high performance values such as 99.35% c_acc and 96.31% p_acc.