Graph-guided frequency-enhanced state space network for 3D spine segmentation from MR images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate spinal MRI segmentation is crucial for diagnosing spinal diseases.
• Existing methods struggle with complex anatomy and artifacts, limiting global modeling and boundary delineation.

Purpose of the Study:

• To propose the Graph-Guided Frequency-Enhanced State Space Network (GF-SSNet) for accurate 3D multi-modal spine MRI segmentation.
• To address limitations in global semantic modeling, cross-modal perception, and fine boundary identification.
• To provide technical support for intelligent diagnosis and precision medicine in spinal diseases.

Main Methods:

• The GF-SSNet utilizes a dual frequency-spatial enhancement mechanism with Frequency Dynamic Convolution (FDConv) and Three-Directional Mamba (TD-Mamba).
• It incorporates Position-Aware Attention Fusion (PAAF) and Graph Convolutional Networks (GCN) for topological anatomical constraints.
• A Depth-aware Progressive Upsampling (DAPU) strategy is used for fine-grained spatial information reconstruction.

Main Results:

• GF-SSNet achieved superior performance on normal test sets, with Dice Mean of 92.04% and IoU Mean of 85.29%.
• It significantly reduced HD95 to 3.06 mm and ASSD to 0.612 mm compared to baselines.
• On pathological test sets, GF-SSNet maintained strong performance (Dice Mean 87.60%), demonstrating robustness despite segmentation challenges in degenerative conditions.

Conclusions:

• GF-SSNet effectively segments spinal MRIs by fusing frequency features and global dependencies.
• The method offers improved technical support for intelligent diagnosis of spinal diseases.
• Ablation studies and loss function analysis validated the contribution of each component.