T-GAN: Transformer Generative Adversarial Network for Brain Tumor Segmentation

The brain, an intricate organ that oversees body processes and intellect, is susceptible to tumors, which are abnormal growths that impede regular brain activity and pose substantial health risks. Since tumor structures are multifaceted and varied, precisely segmenting brain tumors in multimodal Magnetic Resonance Imaging (MRI) still remains a problematic endeavor, despite being essential for efficient diagnosis and treatment planning. This research unveils Transformer-based Generative Adversarial Network (T-GAN), a new Deep Learning (DL) framework that combines the T-GAN architecture. The proposed strategy uses a Multi-Scale Cross-Attention Transformer (MCAT), which uses axial-slice attention mechanisms, cross-attention fusion and an attention U-Net to capture both local and global contextual variables. These components enhance the model’s capability to handle intermodal interdependence, which means the mutual reliance and interaction among different network modules to extract complementary features and to effectively capture spatial heterogeneity in 3D brain MRI data. Furthermore, a Patch-GAN discriminator is employed that improves structural consistency by enforcing anatomical priors, resulting in refined and realistic segmentation masks. For the purpose of enhancing modal generalization, the method incorporates data enrichment techniques in addition to strong preprocessing methods, such as skull stripping, resizing and Z-score normalization. Simplicial Convolutional Neural Network (SCNN) is employed for the fine feature extraction of the MRI image data. Utilizing the Brain Tumor Segmentation (BraTS) dataset for evaluation, T-GAN exhibits remarkable enhancements over both conventional and modern models, attaining exceptional results in segmentation metrics, including 99.8% accuracy, 99.7% Dice Similarity Coefficient (DSC) and 99.8% Intersection over Union (IoU). These findings support the framework’s ability to produce reliable, precise and clinically relevant segmentation results.