A novel segmentation method for cervical vertebrae based on PointNet++ and converge segmentation

Area of Science:

• Medical Imaging
• Computational Anatomy
• Spine Surgery

Background:

• Cervical spine instability is a primary cause of cervical spondylosis, leading to nerve compression and potential paralysis.
• Accurate surgical planning for cervical spine reconstruction is crucial but challenging due to vertebral similarities.
• Vertebral segmentation is vital for automating surgical planning and improving treatment for cervical spondylosis.

Purpose of the Study:

• To develop and evaluate an accurate algorithm for segmenting cervical vertebrae from CT images.
• To enhance the precision of vertebral segmentation for improved surgical planning and diagnosis.
• To leverage deep learning for automated analysis of cervical spine anatomy.

Main Methods:

• A three-part algorithm combining adaptive threshold filtering, PointNet++ for single vertebra segmentation, and edge-based convergence for accuracy.
• Utilized CT images for segmenting cervical vertebra tissue structure.
• Employed PointNet++ and edge information for precise segmentation of individual vertebrae.

Main Results:

• Achieved a system accuracy of 96.15%, surpassing previous methods on the dataset.
• Demonstrated superior performance compared to Convolutional Neural Network (CNN) and PointNet methods on a separate dataset.
• Validated the robustness and promise of PointNet++ for medical image segmentation.

Conclusions:

• The proposed method offers superior classification performance for cervical spine image segmentation.
• Directly and effectively segments three-dimensional vertebral bodies.
• Precise vertebral segmentation supports automated biomechanical analysis and computer-aided diagnosis, enhancing cervical spondylosis treatment automation.