DKA-U-Net: Dynamic-Kernel Attention U-Net for Brain Tumor Segmentation From MRI

The precise delineation of tumor boundaries and the accurate segmentation of small tumor subregions are two pivotal factors influencing the performance of brain tumor segmentation from MRI.To enhance boundary accuracy and improve sensitivity to small tumor subregions, we propose a novel U-shaped segmentation network enhanced with dynamic kernel attention convolution, abbreviated as DKA-U-Net. First, the second convolutional layer in the standard two-layer block is replaced with a parallel combination of a dynamic convolution path and a standard convolution path, which optimizes the representation of multi-scale features. Second, the dynamic convolution path replaces fixed kernels with a radial basis function-based kernel generator, which enables differentiable parameterization. Third, the outputs of the two convolutional paths are fused and then passed through a lightweight hybrid attention mechanism, which refines them while maintaining computational efficiency. Extensive experiments implemented on the BraTS2023 dataset confirmed the superior performance of the proposed DKA-U-Net achieves competitive performance, with Dice scores of 92.0% for the whole tumor (WT), 83.2% for the tumor core (TC), and 81.6% for the enhancing tumor (ET), reflecting a 4.1% average improvement over the U-Net baseline. Sensitivity results reach 92.2% (WT), 81.5% (TC), and 81.6% (ET), corresponding to an average gain of 4.1%. These findings demonstrate that DKA-U-Net enhances boundary precision and improves segmentation accuracy for small tumor subregions in complex neuroimaging scenarios.