A Unified Adaptive Mri Enhancement And Dual-Path Transformer Framework With Radial Gan-Based Data Balancing For Brain Tumour Classification

Abstract

Brain tumor categorization through MRI has been proven to be difficult as tumors differ from person to person by varying shapes, sizes, and intensities. Most existing systems utilize CNN-based transfer learning techniques trained using natural images, resulting in decreased sensitivity to subtle medical properties. While numerous hybrid techniques have resulted in increased accuracies, most of these systems require manual pre-processing, inefficient feature extraction, and lack of proper solutions for imbalanced datasets. This reduces the effectiveness of such an automated system. The current deep learning techniques suffer from the loss of important features surrounding the tumor in different MRI scans, as well as the inability to generalize to a variety of tumor properties.The lack of a unified solution addressing both contrast enhancement and class imbalance further restricts clinical readiness. To address these issues, this study developed a unified pipeline combining adaptive MRI enhancement with a dual-path transformer architecture. The ATri–PLTri–HEA enhancement method improves local contrast, increases soft-tissue clarity, and preserves important tumor edges. A Radial GAN generates synthetic tumor samples to expand minority classes and stabilise dataset distribution. Feature extraction and classification are performed by a Dual-Path Vision Transformer with Deformable Cross-Attention (DPViT-DCA). The framework achieved 99.2% accuracy and strong benchmark performance.