MAUNet: a mixed attention U-net with spatial multi-dimensional convolution and contextual feature calibration for 3D brain tumor segmentation in multimodal MRI

Introduction Brain tumors present a significant threat to human health, demanding accurate diagnostic and therapeutic strategies. Traditional manual analysis of medical imaging data is inefficient and prone to errors, especially considering the heterogeneous morphological characteristics of tumors. Therefore, to overcome these limitations, we propose MAUNet, a novel 3D brain tumor segmentation model based on U-Net. Methods MAUNet incorporates a Spatial Convolution (SConv) module, a Contextual Feature Calibration (CFC) module, and a gating mechanism to address these challenges. First, the SConv module employs a Spatial Multi-Dimensional Weighted Attention (SMWA) mechanism to enhance feature representation across channel, height, width, and depth. Second, the CFC module constructs cascaded pyramid pooling layers to extract hierarchical contextual patterns, dynamically calibrating pixelcontext relationships by calculating feature similarities. Finally, to optimize feature fusion efficiency, a gating mechanism refines feature fusion in skip connections, emphasizing critical features while suppressing irrelevant ones. Results Extensive experiments on the BraTS2019 and BraTS2020 datasets demonstrate the superiority of MAUNet, achieving average Dice scores of 84.5 and 83.8%, respectively. Ablation studies further validate the effectiveness of each proposed module, highlighting their contributions to improved segmentation accuracy. Our work provides a robust and efficient solution for automated brain tumor segmentation, offering significant potential for clinical applications.