Self-Supervised Learning Achieves 99.64% Accuracy in Brain Tumor MRI Classification
A recent research paper on arXiv (2605.01999) presents TumorXAI, an innovative self-supervised learning (SSL) framework aimed at explainable classification of brain tumors from MRI images. Researchers employed a ResNet-50 backbone to assess SimCLR, BYOL, DINO, and MoCo v3 using a public dataset consisting of 4,448 MRIs representing 17 different tumor types. Notably, SimCLR reached an impressive accuracy, precision, recall, and F1-score of 99.64%. The methodology encompasses preprocessing, fine-tuning, linear evaluation, and SSL pretraining with various augmentations. When labels were scarce, SSL-pretrained models surpassed traditional supervised baselines. Additionally, the framework offers visual insights into the decisions made by the model.
Key facts
- arXiv paper 2605.01999 introduces TumorXAI framework
- Uses self-supervised learning (SSL) for brain tumor classification
- ResNet-50 backbone evaluated with SimCLR, BYOL, DINO, MoCo v3
- Dataset: 4,448 MRIs with 17 distinct tumor types
- SimCLR achieved 99.64% accuracy, precision, recall, F1-score
- Workflow includes preprocessing, fine-tuning, linear evaluation, SSL pretraining
- SSL outperforms supervised baselines with limited labels
- Provides visual explanations for model decisions
Entities
Institutions
- arXiv