3D ST-Net: A Large Kernel Simple Transformer for Brain Tumor Segmentation

Glioblastoma is the most common primary malignant brain tumor, necessitating precise depiction through magnetic resonance (MR) imaging for effective treatment. Due to the tumor’s heterogeneous and elusive boundaries, appearance, and shape, automated segmentation remains a complex task. With the emergence of deep learning, researchers have significantly improved brain tumor segmentation using convolutional neural networks. In this article, we propose a novel volume-based 3D Simple Transformation Network (ST-Net), which utilizes large convolutional kernels (7 \(\,\times \,\) 7 \(\,\times \,\) 7) to map the original three-dimensional images into a lower-dimensional latent space, and employ depthwise convolutional scaling (DCS) to separate spatial and channel dimensions, reducing memory parameters and computational load while enlarging the receptive field. To enhance segmentation performance, our loss function combines cross-entropy and Dice loss. Through online validation, the enhanced Dice scores for Enhancing Tumor (ET), Tumor Core (TC), and Whole Tumor (WT) are 0.776, 0.790, and 0.880, respectively, while the Hausdorff distance measures are 36.7, 35.3, and 18.2.