CSM-WNet: A Multi-modal Feature Fusion Network for MRI Brain Tumor Segmentation
Year: 2026
Type: conference-paper
Abstract: A critical challenge in 3D multi-modal MRI analysis stems from the inherent information redundancy and coupling across different imaging modalities. This complexity makes it difficult to effectively integrate multi-modal information and extract meaningful features from the intricate modality-specific characteristics. Commonly used but simplistic fusion strategies, such as early fusion, which treats multi-modal data as multi-channel inputs, or late fusion, which aggregates results after independent processing, often fail to exploit the complementary relationships among imaging sequences fully. Meanwhile, brain tumors exhibit high intrinsic heterogeneity with sub-regions including enhancing tumor edema and necrotic core, demonstrating substantial variations in shape, size, location, and textural characteristics, while also possessing frequently ill-defined boundaries against surrounding healthy tissues. To address these challenges we propose CSM-WNet a novel framework designed explicitly for 3D multi-modal MRI segmentation. CSM-WNet reconfigures the traditional skip connections of the U-Net architecture by integrating up-sampling modules thereby enabling effective fusion of multi-scale image features across different resolutions. Furthermore, the introduced CSMix feature fusion mechanism performs refined extraction along both channel-wise and spatial dimensions, allowing the model to learn and leverage distinctive characteristics from various MRI modalities effectively. Experimental results on the BraTS2021 and BraTS2024 datasets demonstrate that CSM-WNet significantly outperforms existing state-of-the-art models in segmenting various tumor sub-regions, while maintaining a light-weight model structure that ensures computational efficiency and fast inference speed.
Institution: Shenzhen Technology University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203965521