Towards Efficient Brain Tumor Segmentation via a Transformer-Driven 3D U-Net

Abstract

Accurate brain tumor segmentation is critical for clinical diagnosis and treatment. The rapid development of deep neural networks (DNNs) in computer vision offers an automated solution for segmentation tasks. However, convolutional neural networks (CNNs) cannot model long-range dependencies, hindering their perception of global information on tumors. Moreover, vision Transformers (ViTs) require extensive annotated data for optimal segmentation performance, leading to high computational costs and overfitting on small datasets. To address these challenges, we propose TDU-Net, an efficient and accurate brain tumor segmentation scheme using Transformer-driven 3D U-Net. In TDU-Net, improved inverted residual bottlenecks with large kernels are employed in both downsampling and upsampling blocks, optimizing memory efficiency while maintaining global semantic richness in 3D multimodal tumor data. Inspired by ViT, fewer activation functions and normalization layers are used in downsampling and upsampling blocks. GELU activation, group normalization, and larger convolution kernels are employed to improve the global perception and segmentation capability on small datasets. Additionally, orthogonal regularization is introduced during training to mitigate overfitting and enhance generalizability. Experimental results demonstrate that TDU-Net achieves superior brain tumor segmentation accuracy with fewer model parameters, thereby improving generalizability and reducing performance degradation due to overfitting.