Asymmetric encoder-decoder U-net with hybrid attention and residual dilated convolution for brain tumor segmentation

Inroduction The accurate and automatic segmentation of brain tumors from magnetic resonance imaging is regarded as essential for clinical diagnosis and surgical planning. Nevertheless, the concurrent extraction of both local fine details and global contextual cues remains challenging, primarily due to ambiguous tumor margins and marked disparities in tumor dimensions. Methods To address these issues, a novel HCSA-RMSDC U-Net is designed in this work, which is characterized by an asymmetric architecture with a computationally heavy encoder and a lightweight decoder. Specifically, a hybrid channel–spatial attention (HCSA) module is incorporated into the encoder, wherein axial decomposed convolution is utilized to capture directional spatial information, thereby enhancing spatial feature representation for tumor segmentation. Concurrently, a residual multi-scale dilated convolution (RMSDC) module is deployed in the bottleneck layer to capture semantic information across varying scales without increasing the spatial resolution. Results The proposed model was evaluated on the publicly available BRISC 2025 and Figshare datasets, achieving Dice coefficients of 87.39% and 82.84%, respectively Discussion These findings support the complementary roles of the HCSA and RMSDC mechanisms in enhancing feature representation and multi-scale contextual modeling. Overall, the proposed framework provides a fully automated approach for brain tumor segmentation.