A fully automatic Cobb angle measurement framework of full-spine DR images based on deep learning

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Scoliosis affects millions of children globally, requiring accurate Cobb angle measurement for assessment.
• Manual Cobb angle measurement is time-consuming, labor-intensive, and prone to observer variability.
• Accurate identification of vertebrae in both AP and LAT views of spinal DR is challenging.

Purpose of the Study:

• To develop a deep learning framework for fully automated Cobb angle measurement from full-spine digital radiography (DR).
• To overcome the limitations of manual Cobb angle assessment, including time, labor, and observer variations.

Main Methods:

• A deep learning network was employed to differentiate AP and LAT views.
• Spine region of interest (ROI) was located and extracted.
• A detection network (YOLOv8 with CBAM) identified vertebrae boundaries, types, and corner points for automated Cobb angle calculation.

Main Results:

• The framework achieved mean Cobb angle errors of 2.56° (AP) and 2.498° (LAT).
• High intra-class correlation coefficients (0.956 AP, 0.925 LAT) and Pearson correlation coefficients (0.961 AP, 0.930 LAT) were observed.
• A reader study reported a mean error of 3.918° for the major curve with strong ICC (0.943) and correlation (0.960).

Conclusions:

• The proposed framework demonstrates significant accuracy and consistency in Cobb angle measurement.
• The AI-driven approach validates its effectiveness and offers strong support for clinical scoliosis diagnosis.
• Automated Cobb angle measurement enhances diagnostic efficiency and reduces variability.