Vertebrae and intervertebral discs segmentation using deep learning-based model in disability analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate segmentation of vertebrae and intervertebral discs (IVDs) is crucial for diagnosing and treating spinal disorders.
• Current deep learning methods often segment spinal components independently, neglecting valuable anatomical relationships.
• Existing approaches may lack the precision needed for comprehensive clinical applications.

Purpose of the Study:

• To develop and validate a two-stage deep learning framework for automated spine segmentation in T2-weighted MR images.
• To incorporate structural dependency modeling to improve segmentation accuracy by leveraging anatomical relationships.
• To provide a robust and clinically applicable automated system for spinal analysis.

Main Methods:

• A two-stage deep learning framework combining a 3D Graph Convolutional Segmentation Network (GCSN) and a 2D ResNet refinement network.
• Modeling spine components as graph nodes with anatomical relationships captured in an adjacency matrix.
• Utilizing T2-weighted MR images from 218 subjects for model training and testing.

Main Results:

• Achieved an average Dice Similarity Coefficient (DSC) of 87.32% for vertebrae, 87.78% for intervertebral discs, and 87.49% for the entire spinal column.
• Demonstrated significant improvements in segmentation consistency and accuracy through graph-based learning of anatomical dependencies.
• Exemplary segmentation performance indicating high reliability.

Conclusions:

• The proposed graph-based deep learning approach significantly enhances automated spine segmentation accuracy and consistency.
• The framework effectively utilizes anatomical relationships, outperforming methods that treat spinal components independently.
• The system offers a safe, highly effective, and clinically viable tool for spinal disorder diagnosis and treatment planning.