FRSK-UNet: Enhanced MRI Brain Tumor Segmentation with Feature Refinement and Selective Kernel Attention

Accurate feature extraction is essential for identifying tumor boundaries in brain tumor MRI segmentation. This paper introduces FRSK-UNet, a U-shaped network that integrates Selective Kernel Attention (SKA) and a newly designed Feature Refinement Module (FRM). The FRM enhances feature maps by combining channel and spatial attention, improving context and feature fusion, thus enhancing U-Net's ability to segment complex and diverse brain tumor regions. SKA captures multi-scale information through variable-sized convolutional kernels, enhancing the network's comprehension of intricate structures. Evaluated on the BraTS2020 dataset, FRSK-UNet achieved Dice coefficients of 0.8426, 0.8995, and 0.7948, and Hausdorff distances of 1.3516mm, 0.9152mm, and 1.4847mm for WT, TC, and ET regions, respectively. These findings indicate that FRSK-UNet outperforms other leading networks, markedly improving segmentation accuracy.