Tumor-Swin Transformer model for brain Tumor classification using MRI

Abstract

Brain tumor detection using magnetic resonance imaging (MRI) remains a critical challenge in medical image analysis due to variations in tumor appearance and the limited inclusion of normal brain images in existing studies. To address this issue, a multi-source MRI dataset containing 4484 images was compiled, including both tumor and normal brain classes to improve model generalization under realistic diagnostic conditions. The dataset consisted of 3133 training images and 1351 testing images, with balanced distributions across tumor and non-tumor categories. The proposed study introduces a fine-tuned Tumor-Swin-Transformer (TST) framework for automated brain tumor classification. Image preprocessing included normalization, resizing, duplicate removal, and stratified data splitting to reduce bias and improve robustness. The TST model was optimized using task-specific training configurations to ensure stable convergence and efficient learning. Experimental results demonstrate that the proposed framework achieved an accuracy of 0.98 within only three training epochs, along with a Matthews Correlation Coefficient (MCC) of 0.9576 and a Hamming Loss of 0.0215. Furthermore, Grad-CAM-based explainable artificial intelligence (XAI) was incorporated to improve interpretability and clinical relevance. Comparative analysis indicates that the proposed framework outperforms several existing machine learning and deep learning approaches while providing improved generalization, computational efficiency, and interpretability for brain tumor detection.