Spine Segmentation with Multi-view GCN and Boundary Constraint.

Area of Science:

• Medical Imaging
• Computer Vision
• Spine Anatomy

Background:

• Accurate segmentation of vertebrae and inter-vertebral discs (IVDs) is essential for diagnosing spinal diseases.
• Current segmentation methods face challenges due to inter-subject variability and intra-subject similarities.

Purpose of the Study:

• To propose a novel framework for automatic multi-class segmentation of vertebrae and IVDs in MR images.
• To address the challenges in spine segmentation using advanced deep learning techniques.

Main Methods:

• A novel spine segmentation framework incorporating a Multi-View Graph Convolutional Network (MVGCN).
• Utilizes multi-view features and graph convolutional networks (GCN) to capture spatial relationships between vertebrae and IVDs.
• Incorporates a boundary constraint to enhance segmentation accuracy at vertebral-IVD interfaces.

Main Results:

• The proposed MVGCN framework demonstrated efficacy in segmenting vertebrae and IVDs on a public dataset of 172 MR images.
• Achieved accurate segmentation, outperforming existing methods by effectively handling variations in spine anatomy.
• The boundary constraint further improved the precision of the segmentation, particularly at critical anatomical junctions.

Conclusions:

• The developed framework offers a robust solution for automatic vertebrae and IVD segmentation in MR imaging.
• This method has the potential to significantly aid in the diagnosis and treatment planning for spinal conditions.
• Future work will involve releasing the code and models to facilitate further research and clinical application.