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

BACKGROUND CONTEXT: Manual measurement of cervical sagittal parameters is time-consuming and exhibits significant interobserver variability. Existing artificial intelligence models fail when C7 is obscured by shoulder anatomy. PURPOSE: To develop and externally validate a deep learning model for automated cervical alignment measurements under clinical conditions, including C7-obscured cases. DESIGN: Retrospective observational study. PATIENT SAMPLE: A total of 5,604 lateral cervical radiographs were obtained from Chinese and Korean institutions. OUTCOME MEASURE: Intraclass correlation coefficient (ICC), Pearson correlation (r), and Bland-Altman agreement. METHODS: A Keypoint R-CNN with ResNet-50-FPN backbone was trained using multinational data, including C7-obscured cases. Model outputs were compared to consensus expert annotations using ICC, Pearson correlation, and Bland-Altman analysis. An independent dataset was used for external validation. RESULTS: In the external validation set (n=100), 62 patients (62.0%) had a partially obscured C7 and 20 patients (20.0%) had a fully obscured C7. The final model showed excellent reliability for the C2-C7 lordosis (ICC=0.95, r=0.95), C2 slope (ICC=0.99, r=0.99) and C7 slope (ICC=0.91, r=0.91). The mean errors for these parameters were clinically negligible at -0.44°, 0.06°, and -0.38°, respectively. The reliability for all disc height measurements were excellent in internal test set (ICC=0.97-0.99). Measurement errors slightly increased in few patients with complete C7 obscuration. CONCLUSION: The Keypoint R-CNN model enables rapid, accurate, and clinically generalizable automated cervical alignment measurements; however, C7 obscuration remains a critical limitation that requires targeted improvement.