Classification of Cervical Vertebral Maturation Stages and Bone Age Assessment Using Transfer Learning–Based Deep-Learning Approaches

Abstract

In this study, an automatic classification of cervical vertebra maturation (CVM) stages was performed using raw lateral cephalometric radiographs to assess growth and development. A total of 4285 radiographs from the Department of Orthodontics at Van Yüzüncü Yıl University Faculty of Dentistry were utilized. Following detailed evaluations by specialist physicians, 3750 images meeting diagnostic accuracy and clinical suitability criteria were included. The selected images were categorized into six classes (CVMS 1–6), forming a balanced dataset for classification with the NFNet, ConvNeXt V2, EfficientNet V2, and DeiT3 models. The NFNet model achieved the highest overall performance, with 96% training accuracy and 85.7% test accuracy. ConvNeXt V2, attaining 95% training accuracy and 86.9% test accuracy, emerged as the most balanced in terms of generalization. Although EfficientNet V2 reached 94% training accuracy, its 80.7% test accuracy indicated limited generalization. With 93% training accuracy and 77.6% test accuracy, DeiT3 demonstrated the lowest capacity. Both NFNet and ConvNeXt V2 stood out as strong classification candidates based on their high accuracy and balanced performance. While NFNet showed a 10.3% gap between training and test accuracy, indicating somewhat reduced generalization, ConvNeXt V2’s narrower 8.1% gap suggested greater stability. In conclusion, NFNet and ConvNeXt V2 are promising models for CVM classification. Future studies should employ larger datasets and conduct hyperparameter optimization to enhance these models’ performance and strengthen their clinical applicability.