RASpine: Regional Attention Lateral Spinal Segmentation based on Anatomical Prior Knowledge

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Spine segmentation from X-rays aids in diagnosing spinal disorders.
• Current methods struggle with lateral X-rays due to low contrast and overlapping vertebrae, particularly thoracic ones.
• Overlapping segmentation results hinder accurate visualization of spinal structure and morphology.

Purpose of the Study:

• To develop an improved spine segmentation network for lateral X-ray images.
• To address the challenge of overlapping vertebrae in segmentation results.
• To enhance the accuracy and reliability of spinal imaging analysis.

Main Methods:

• Proposed a novel segmentation network: Region Attention and Spine Prior-based Network (RASpine).
• Incorporated anatomical prior knowledge of non-overlapping vertebral regions.
• Developed an overlap detector and a specialized loss function to penalize overlapping segmentation.
• Utilized region attention to improve segmentation in challenging areas.

Main Results:

• RASpine effectively identifies and penalizes overlapping vertebrae in lateral X-ray spine segmentation.
• The network demonstrated superior performance compared to existing mainstream medical image segmentation algorithms.
• Achieved more satisfactory results across multiple evaluation metrics on a clinical dataset.

Conclusions:

• RASpine offers a significant advancement in segmenting lateral X-ray spine images.
• The method successfully mitigates the problem of overlapping vertebrae, enhancing diagnostic clarity.
• This approach holds promise for improving the clinical diagnosis and treatment of spinal disorders.