TransU-Net: Glioma Sub-region Segmentation

Gliomas, the utmost prevalent form of malicious primary brain tumors, are characterized by their extreme heterogeneity. Precise glioma sub-region segmentation in to Enhancing Tumors, Tumor core, and Whole Tumor are essential for planning treatment and monitoring tumor progression. Automating this process is a formidable job owing to the complexity of tumor shapes and textures. This study presents a U-Net Transformer architecture that incorporates a self-attention mechanism and skip-connections. The model leverages the global contextual capabilities of Transformers and the precise localization from U-Net. Using BraTS 2021 dataset, we demonstrated the efficacy of our approach. The TransU-Net architecture outperforms existing methods, particularly in segmenting challenging ET and TC regions, achieving Dice similarity coefficients (DSC) of 0.855 (ET), 0.904 (WT), and 0.915 (TC), with corresponding sensitivities of 81.9%, 83.4%, and 81.6%, respectively. Our results point out that the proposed TransU-Net model not only excels in capturing complex tumor patterns but also shows promise in enhancing the accuracy of medical image segmentation, potentially aiding in better clinical outcomes.