DuS-Net: Spatial-Semantic Dual-Collaborative Architecture for MRI Glioma Segmentation
Automatic and precise segmentation of brain gliomas from magnetic resonance imaging (MRI) is of great significance for clinical auxiliary diagnosis and treatment planning. To address the high heterogeneity of brain gliomas, blurred boundary localization caused by infiltrative growth, and difficulty in extracting local fine-grained features, this study focuses on tumor-region feature extraction and global network modeling, and proposes a spatial-semantic dual-collaborative network (DuS-Net). In the encoder stage, a multi-branch dynamic convolution module (MDConv) is designed. By integrating standard depthwise separable convolution and dilated convolution in parallel and combining them with a selective kernel attention mechanism, MDConv achieves precise extraction of fine-grained local details in heterogeneous tumors. Meanwhile, a KAN nonlinear proxy generation module (KAB) is constructed at the bottleneck layer, which collapses the high-dimensional feature space into a global proxy vector and uses B-spline basis functions for nonlinear channel disentanglement to dynamically model complex relationships among pathological features, thereby enabling efficient global semantic modeling. In the decoder stage, a three-dimensional coordinate-attention-guided adaptive feature fusion module (AFF) is proposed. Through the fusion gate, spatially adaptive encoder-decoder weight allocation is learned from concatenated features, and encoder features are attention-refined along three-dimensional directions to enhance spatial localization ability, achieving adaptive integration and alignment of spatial fine-grained information and high-level semantics. Experimental results on the BraTS 2021 and BraTS-Africa datasets demonstrate that DuS-Net achieves higher segmentation accuracy than recent mainstream segmentation methods, validating its effectiveness for complex brain glioma segmentation and indicating its potential value for computer-assisted glioma segmentation.