Automated measurement of cervical sagittal and local parameters using a generalizable deep learning model: a...

Area of Science:

• Spine surgery and imaging analysis
• Artificial intelligence in medical diagnostics
• Radiographic parameter measurement automation

Background:

• Manual measurement of cervical sagittal parameters is time-consuming and prone to interobserver variability.
• Existing AI models struggle with C7 obscuration due to shoulder anatomy.

Purpose of the Study:

• To develop and externally validate a deep learning model for automated cervical alignment measurements.
• To address clinical conditions, including cases with C7 obscuration.

Main Methods:

• Retrospective observational study involving 5,604 lateral cervical radiographs from Chinese and Korean institutions.
• A Keypoint R-CNN model with ResNet-50-FPN backbone was trained on multinational data, including C7-obscured cases.
• Model performance was evaluated against expert annotations using ICC, Pearson correlation, and Bland-Altman analysis, with external validation on an independent dataset.

Main Results:

• The model demonstrated excellent reliability for C2-C7 lordosis (ICC=0.95), C2 slope (ICC=0.99), and C7 slope (ICC=0.91) in the external validation set.
• Mean errors for these parameters were clinically negligible (-0.44°, 0.06°, -0.38°).
• Reliability for disc height measurements was excellent (ICC=0.97-0.99), with slight error increases in cases of complete C7 obscuration.

Conclusions:

• The Keypoint R-CNN model provides rapid, accurate, and generalizable automated cervical alignment measurements.
• C7 obscuration remains a significant limitation requiring targeted improvements for enhanced clinical utility.