Automated Measurement of Occipito-Axial Angle on Cervical Radiographs Using a Deep Learning Object Detection Model: A Proof-of-Concept Study

Objective: Maintaining the occipito-axial (O-C2) angle following occipitocervical fusion is crucial to prevent postoperative complications. Although automated O-C2 measurement has been reported, practical methods that provide rapid results for routine practice remain limited. This study aimed to develop a deep learning model using the YOLO (You Only Look Once) object detection algorithm to automatically identify anatomical landmarks and rapidly calculate the O-C2 angle. Methods: A retrospective analysis was conducted using cervical spine radiographs from 2 independent facilities. The internal dataset comprised 574 lateral cervical radiographs from 271 patients for model development, while the external validation dataset included 100 radiographs from 100 patients. Model performance was evaluated against manual measurements by 3 expert raters. Results: The model demonstrated excellent detection performance, achieving perfect metrics for the hard palate (F1 score: 1.00) and high performance for the occipital bone (F1 score: 0.97), anteroinferior corner of C2 (F1 score: 0.99), and posteroinferior corner of C2 (F1 score: 0.99). For O-C2 angle estimation, the mean absolute error was 2.35° and root mean squared error was 2.98°, with an accuracy of 94.7% for determining the presence or absence of the O-C2 angle (i.e., whether all 4 anatomical landmarks were simultaneously detected). Bland-Altman analysis revealed minimal bias (0.57°; 95% confidence interval, -0.06° to 1.12°) with limits of agreement from -5.19° to 6.33°. Inference time was approximately 0.14 s per image. Conclusion: Our deep learning model enables rapid and accurate O-C2 angle measurement on lateral cervical radiographs, demonstrating performance comparable to expert raters and potential clinical utility.