WGAN-based multi-structure segmentation of vertebral cross-section MRI using ResU-Net and clustered transformer

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spine Anatomy

Background:

• Accurate segmentation of lumbar spine structures (vertebral body, foramen, lamina) is crucial for diagnosing spinal diseases.
• Existing segmentation methods may lack the precision required for detailed pathological analysis.

Purpose of the Study:

• To develop and evaluate a novel multi-structure semantic segmentation method for vertebral transverse section MRI slices.
• To enhance the accuracy and efficiency of segmenting critical lumbar spine components.

Main Methods:

• Utilized a Generative Adversarial Network (WGAN) framework incorporating a residual U-Net and a clustered Transformer.
• The generator network was replaced with a hybrid segmentation network combining enhanced U-Net and clustered Transformer components.
• Employed dilated convolutions and residual structures in the U-Net encoder for improved multi-scale feature extraction.

Main Results:

• Achieved statistically significant improvements in segmentation accuracy (p < 0.05).
• Demonstrated a 3.1% increase in Dice coefficient for vertebral body segmentation.
• Showcased a 4.1% improvement in mean Intersection over Union (mIOU) and a 2.0% increase in Positive Predictive Value (PPV).
• Reduced Hausdorff distance by 0.6 mm, indicating higher boundary accuracy.

Conclusions:

• The proposed method, combining residual U-Net and clustered Transformer, significantly enhances semantic segmentation of vertebral structures in MRI.
• The enhanced modules effectively improve segmentation accuracy for the vertebral foramen, lamina, and vertebral body.
• This advanced segmentation technique holds promise for improved diagnosis and understanding of spine-related pathologies.