T3SSLNet: Triple-Method Self-Supervised Learning for Enhanced Brain Tumor Classification in MRI

Abstract

Classification of brain tumors from MRI images is crucial for early diagnosis and effective treatment planning. However, there are still obstacles to overcome, including low image quality, sparsely labeled data, and variability in tumor characteristics. In this study, we explored the use of self-supervised learning techniques to improve the classification performance for brain tumors. Specifically, we tested three SSL approaches SimCLR, MoCo, and BYOL, with ResNet-50 as the backbone architecture on a newly constructed dataset created by combining five public datasets. We further extended our work by integrating EfficientNet to evaluate its computational efficiency, demonstrating its feasibility for low-processor systems. We introduce T3SSLNet, a novel framework consisting of four key components: the imaging spectrum enhancement block for data augmentation, the Frozen Feature Extractor block for hierarchical feature extraction, Neural Representation Projection Learning block for contrastive-positive pair learning, and Unfrozen Classification block for tumor classification. Our experimental results paired with ResNet-50 indicate that, without fine-tuning, MoCo achieved the highest accuracy at 95.76%, followed by SimCLR at 92.25% and BYOL at 81.80%. Following fine-tuning, BYOL showed a significant improvement, reaching 96.42%, while MoCo and SimCLR reached 96.87% and 97.02%, respectively.