Robust MRI-Based Brain Tumor Detection Using Hybrid Feature Learning and Self-Supervised Pretraining

Abstract

Brain tumors represent one of the most critical and challenging health conditions, requiring early and accurate diagnosis for effective treatment. Traditional machine learning and standalone deep learning models often struggle to capture the complex features in MRI brain images, leading to suboptimal classification performance. To address these limitations, this work proposes a novel method titled Robust MRI-Based Brain Tumor Detection Using Hybrid Feature Learning and Self-Supervised Pretraining. The approach integrates a hybrid model that combines Convolutional Neural Networks (CNN) for local feature extraction with Transformer encoders for global feature representation, enhanced further by a self-supervised Masked Autoencoder (MAE) pretraining strategy. Using a well-structured feature fusion mechanism, the proposed system focuses on efficiently classifying brain tumors into glioma, meningioma, pituitary, and no tumor categories. Experimental evaluations demonstrate that the hybrid SSL+CNN+Transformer model outperforms baseline architectures such as CNN-only, Transformer-only, and CNN+Transformer (no SSL) combinations. Specifically, the proposed model achieved an accuracy of 94% and an F1-score of 93%, significantly improving classification performance compared to traditional methods. Compared to the best existing non-hybrid models, the proposed solution offers an improvement of approximately 6% in accuracy and 7% in F1-score, highlighting its potential to enhance diagnostic reliability and support clinical decision-making in neuro-oncology applications.