An Explainable SSL-Based Model for Robust Multi-Class Brain Tumor Classification from MRI Images

Abstract Accurate and interpretable brain tumor classification from magnetiSSc resonance imaging (MRI) is important for timely detection and effective treatment planning. Deep supervised learning methods, though strong, are limited by their reliance on vast labeled datasets and their lack of explainability in clinical decision-making. In this work, we introduce a self-supervised learning (SSL) approach based on SimCLR with EfficientNetB3 backbone for four-class brain tumor segmentation: glioma, meningioma, pituitary tumor, and no-tumor. The method employs SSL-based model pre-training on large amounts of unlabeled data to learn salient feature representations prior to performing supervised fine-tuning with an optimal classifier head. The technique effectively enhances generalization with minimal dependence on large-scale human annotation. The envisioned framework has a test accuracy of 98.32%, per-class precision, recall, and F1-measures over 96%, and best classification performance in no-tumor and pituitary classes. For improving interpretability and clinical confidence, Gradient-weighted Class Activation Mapping (Grad-CAM) was used with discriminative tumor region visualization and validation that model attention is in agreement with radiological features. To the best knowledge of the authors, it is the first work that combines an optimized SimCLR-based SSL with brain tumor classification using MRI and explainability. The results show that SSL-driven and interpretable models can have the capability of producing highly accurate, reliable, and clinically relevant decision support for neuro-oncology.