TransUNetB: An advanced Transformer–UNet framework for efficient and explainable brain tumor segmentation

This study presents TransUNetB, a hybrid architecture that combines Transformer and UNet for multi-class brain tumor segmentation. This model integrates global context modeling with precise spatial localization. A lightweight Transformer encoder at the bottleneck captures long-range dependencies, while the U-Net's skip pathways preserve fine anatomical details. Additionally, a multi-scale decoder fusion module consolidates features at various resolutions, enhancing the clarity of tumor boundaries in heterogeneous, low-contrast conditions. Our contributions are threefold: (1) a simple, efficient design that integrates bottleneck self-attention with multi-scale fusion for robust ED/TC/ET segmentation; (2) a comprehensive ablation of design choices—attention type, positional encoding, fusion strategy, loss formulation, and patch size—quantifying their impact on accuracy and efficiency; and (3) an explainability analysis using Grad-CAM with quantitative focus/entropy measures to verify that salient regions align with clinical tumor substructures. Evaluated on the BraTS 2020 and BraTS 2021 datasets, TransUNetB achieves a Dice score of 98.90% and an Intersection over Union (IoU) score of 96.10%. It outperforms strong CNN and vision-transformer baselines while maintaining a competitive runtime of approximately 63 ms per image. These results suggest that combining global attention with spatially faithful decoding provides a favorable trade-off between accuracy and efficiency for clinical deployment. We also discuss the generalization of our model beyond MRI cohorts, practical constraints in resource-limited settings, and future research avenues, including attention-guided fusion and broader multi-center validation.