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

BACKGROUND

Accurate spinal MRI segmentation is essential for computer-aided diagnosis of spinal diseases. Existing methods have limitations in global semantic modeling and boundary delineation due to complex anatomy and imaging artifacts.

PURPOSE: Our work aimed to propose a novel Graph-Guided Frequency-Enhanced State Space Network (GF-SSNet) method to achieve more accurate 3D multi-modal spine MRI automatic segmentation, addressing the limitations of existing algorithms in global semantic modeling of high-dimensional voxel space, cross-modal information synergistic perception, and fine boundary identification of anatomically similar tissues, thereby providing technical support for intelligent diagnosis and precision medicine of spinal diseases.

METHODS

The proposed network is based on the GF-SSNet architecture. During encoding, a dual frequency-spatial feature enhancement mechanism is employed, which adaptively fuses local frequency dynamic features and global spatial long-range dependencies through Frequency Dynamic Convolution (FDConv) and Three-Directional Mamba-based state space model (TD-Mamba). At the bottleneck, Position-Aware Attention Fusion (PAAF) and Graph Convolutional Networks (GCN) are integrated to explicitly encode topological anatomical constraints between vertebrae, enhancing the global perception capability of spinal continuity structures. During decoding, a Depth-aware Progressive Upsampling (DAPU) strategy is introduced to effectively alleviate the reconstruction loss of fine-grained spatial information. The entire framework achieves end-to-end automatic segmentation of multi-modal MR images.

RESULTS

On the normal test set, GF-SSNet outperformed all baselines across all metrics. Specifically, Dice and IoU Means reached 92.04 ± 0.06% and 85.29 ± 0.10%, exceeding the best baseline results of 89.81 ± 0.61% and 81.51 ± 0.91%. HD95 and ASSD were significantly reduced to 3.06 ± 0.46 mm and 0.612 ± 0.018 mm, compared to top-tier baseline values of 4.76 ± 1.09 mm and 1.14 ± 0.01 mm, respectively. On an independent pathological test set with various spinal pathologies, GF-SSNet maintained superior performance with Dice Mean of 87.60 ± 0.10%, still outperforming all baseline methods. The 4.4 percentage point performance decline from normal cases primarily stemmed from intervertebral disc segmentation challenges in degenerative conditions, while vertebrae segmentation remained robust. Ablation studies confirmed significant contributions of all proposed components. The proposed HFD-Tversky loss outperformed conventional losses. All performance differences were statistically significant after correction for multiple comparisons.

CONCLUSION

GF-SSNet demonstrates performance in spinal segmentation through adaptive fusion of frequency features and global dependencies, providing technical support for intelligent spinal disease diagnosis.