A fusion of cross-shaped window attention block and enhanced 3D U-Net for brain tumor segmentation

Abstract

Brain tumor diagnosis and treatment are primarily reliant on medical imaging, necessitating precise segmentation methodologies for practical clinical solutions. Tumor boundaries are difficult to consistently identify, even with breakthroughs in deep learning. To address this challenge, we propose a novel approach that combines an upgraded 3D U-Net architecture for brain tumor segmentation with cross-shaped window attention (CSWA-U-Net). Current segmentation techniques have limitations, particularly in capturing amorphous tumor shapes and fuzzy boundaries. Our strategy aims to overcome these constraints by combining the complementary capabilities of the expanded 3D U-Net, which is efficient at managing volumetric data and maintaining spatial features, with the cross-shaped window attention, which is well-known for capturing long-range relationships and contextual information. We evaluate our method's efficacy using a variety of performance measures, including specificity, sensitivity, and the Dice score. Our results demonstrate increased performance, with Dice scores of 94.7% for the whole tumor, 93.4% for the enhanced tumor region, and 90.5% for the tumor core. Furthermore, our technique has high sensitivity and specificity, highlighting its potential for improving medical imaging analysis.