MSAF: A Cardiac 3D Image Segmentation Network Based on Multiscale Collaborative Attention and Multiscale Feature Fusion

Accurate segmentation of cardiac structures is essential for clinical diagnosis and treatment of cardiovascular diseases. Existing Transformer-based cardiac segmentation methods mostly rely on single-scale token-wise attention mechanisms that emphasize global feature modeling, but they lack sufficient sensitivity to local spatial structures, such as myocardial boundaries in cardiac 3D images, resulting in ineffective multiscale feature capturing and a loss of local spatial details, thereby negatively impacting the accuracy of cardiac anatomical segmentation. To address the above issues, this paper proposes a cardiac 3D image segmentation network named MSAF, which integrates Multiscale Collaborative Attention (MSCA) and Multiscale Feature Fusion (MSFF) modules to enhance the multiscale feature perception capability at both microscopic and macroscopic levels, thereby improving segmentation accuracy for complex cardiac structures. Within the MSCA module, a Collaborative Attention (CoA) module combined with hierarchical residual-like connections is designed, enabling the model to effectively capture interactive information across spatial and channel dimensions at various receptive fields and facilitating finer-grained feature extraction. In the MSFF module, a gradient-based feature importance weighting mechanism dynamically adjusts feature contributions from different hierarchical levels, effectively fusing high-level abstract semantic information with low-level spatial details, thereby enhancing cross-scale feature representation and optimizing both global completeness and local boundary precision in segmentation results. Experimental validation of MSAF was conducted on four publicly available medical image segmentation datasets, including ACDC, FlARE21, and MM-WHS (MRI and CT modalities), yielding average Dice values of 93.27%, 88.16%, 92.23%, and 91.22%, respectively. These experimental results demonstrate the effectiveness of MSAF in segmenting detailed cardiac structures.