Development and Validation of Algorithm for Cervical Vertebral Bone Age Assessment in Maharashtra Population: A Cross-sectional Study

Introduction: The skeletal maturity of individuals has been shown to vary over age, gender, ethnicity, etc. The commonly used methods for Cervical Vertebral Bone Age (CVBA) assessments are subjective and prone to inter-operator bias. This has led to the quest for establishing objective techniques. One such objective method for bone age estimation using multiple regression formulae for males and females has been established by the authors for the population of Maharashtra. However, the practical use of the equations is questionable, encouraging us to develop an automated technique for more practicality. Aim: To establish and validate an algorithm for the quantitative assessment of CVBA in the population of Maharashtra. Materials and Methods: This cross-sectional study was conducted in the Department of Orthodontics and Dentofacial Orthopaedics, Terna Dental College and Hospital, Navi Mumbai, Maharashtra, India over a period of January 2024 to June 2025. This research developed an Artificial Intelligence (AI) algorithm for the objective estimation of CVBA. The research utilised 1000 radiographs from healthy individuals of Maharashtra origin (ages 9-20 years) to train a You Only Look Once (YOLO) v11 object detection algorithm, identifying C3 and C4 vertebral landmarks after data augmentation. A Python program later assessed bone age using these detections and gender-specific methodologies based on pre-existing Maharashtrian formulae, which are stepwise multiple linear regression formulae. The algorithm was tested with 150 independent lateral cephalograms and validated by comparing unused 150 lateral cephalograms with hand-wrist radiographs using the Tanner-Whitehouse (TW3) hand-wrist method (gold standard), employing statistical analyses such as the Mann-Whitney U test, Spearman's rank correlation, and Intraclass Correlation Coefficient (ICC). Results: The mean age estimated using AI technology and hand wrist age was 13.04±1.95 and 12.78±1.82 years, respectively and indicated a negligible difference in the age estimated by the AI and hand wrist age (TW3 method). There was an almost perfect agreement in the age estimated using AI and hand-wrist age (TW3 method) (ICC=0.994). There was an excellent positive correlation in the age estimated using AI and hand-wrist age (TW3 method) (Spearman’s rho=0.974). Conclusion: The average age of AI algorithm was very close to TW3’s hand-wrist method average estimated age. This was consistent across the genders. Thus, the algorithm that was built was dependable as it was closely related to the gold standard.