Wavelet-Driven Spatial Frequency Mamba Network for Spine Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate spine segmentation is vital for diagnosing and treating spinal diseases.
• Mamba-based methods show promise in medical image segmentation but struggle with fine spinal structures and global dependencies.
• Existing frequency-enhanced methods can lose spatial localization, hindering detailed anatomical analysis.

Purpose of the Study:

• To develop an advanced deep learning network for accurate spine segmentation in MRI.
• To address limitations of existing Mamba and frequency-enhanced methods in capturing spinal anatomy.
• To improve diagnostic capabilities for spinal diseases through enhanced segmentation.

Main Methods:

• Proposed the Wavelet-Driven Spatial Frequency Mamba Network (WDSFM-Net) integrating Discrete Wavelet Transform (DWT) with Mamba.
• Introduced Spatial-Frequency Mamba Block (SFMB) to capture global context and local details across frequency subbands.
• Developed Global Strip Pooling Attention (GSPA) and Multi-Scale Attention Enhancement (MSAE) modules for spinal morphology and scale variations.
• Implemented a Dual-Domain Loss (DDL) function for robust training in both spatial and frequency domains.

Main Results:

• WDSFM-Net demonstrated superior performance on two public spine MRI datasets (Spider and MRSpine).
• Achieved average Dice similarity coefficients of 0.8885 on the Spider dataset and 0.8669 on the MRSpine dataset.
• Outperformed other state-of-the-art methods in spine segmentation accuracy.

Conclusions:

• WDSFM-Net effectively addresses the limitations of previous methods for spine segmentation.
• The proposed network accurately captures both global context and fine anatomical details of the spine.
• This advancement holds significant potential for improving the diagnosis and treatment of spinal conditions.