UFTrans: A Hybrid 3D Vision Transformer and U-Net Model for Brain Tumor Segmentation
Usman Sadiq, Aqsa Butt, Labiba Gillani Fahad, Ahmad Raza Shahid, Asad Safi · IEEE Access · 2026
Gliomas represent a highly common and malignant form of brain tumor. Precise segmentation plays a pivotal role in supporting clinical diagnosis and guiding treatment strategies. However, many existing techniques struggle to integrate both local and global information, which limits their segmentation accuracy. Local features such as tumor texture, boundaries, and subtle structural details are important to delineate subregions like the necrotic core and enhance tumor boundaries. Global features, such as tumor size, shape, and spatial relationships within the brain, provide crucial contextual information for precise localization. To address these challenges, we propose UFTrans, a hybrid model that effectively fuses the local feature extraction capabilities of the 3D attention U-Net with the global contextual understanding of a transformer-based model. Our approach employs Hadamard product-based feature fusion to selectively emphasize important features and suppress redundant information, creating a composite feature representation. An inherent attention mechanism in the 3D U-Net continues to refine segmentation by eliminating unwanted activations to reduce computational complexity and accelerate model convergence. We further employ tumor-centered patching to address the severe class imbalance in brain tumor datasets. Extensive experiments on the BraTS 2021 dataset demonstrate the model’s effectiveness, achieving Dice scores of 0.90, 0.91, and 0.89 for the whole tumor, enhancing tumor and tumor core regions, respectively, on the validation set. Comprehensive ablation studies validate the contribution of each architectural component. These results surpass state-of-the-art approaches, indicating the ability of the model to enhance brain tumor segmentation and clinical decision making.