TuSegNet: A Transformer-Based and Attention-Enhanced Architecture for Brain Tumor Segmentation

Abstract

Brain tumor segmentation is crucial in medical imaging, allowing informed diagnosis and treatment planning. In this study, we propose TuSegNet, a new transformer-based and attention-enhanced architecture for robust brain tumor segmentation. The model combines convolutional layers with transformer blocks for global context awareness, incorporates Atrous Spatial Pyramid Pooling (ASPP) for multi-scale feature extraction, and employs channel attention mechanisms to concentrate on tumor-relevant parts. Evaluated on three datasets—Dataset A, Dataset B, and a combined dataset—TuSegNet achieves state-of-the-art performance with a Dice Similarity Coefficient (DSC) of 0.895, 0.910, and 0.930, respectively, and an Intersection over Union (IoU) of 0.820, 0.835, and 0.860. Ablation studies validate the importance of ASPP and attention mechanisms, while comparative analysis demonstrates outstanding performance over existing SOTA models such as Swin UNet and TransUNet. The proposed methodology improves segmentation accuracy and highlights the importance of hybrid architectures in handling complex medical imaging tasks. These developments underscore the potential of TuSegNet for real-world healthcare applications in brain tumor diagnosis.