Deep learning models to classify skeletal growth phase on 3D radiographs

Abstract

Objectives: Cervical vertebral maturation (CVM) is widely used to evaluate growth potential in orthodontics. This study aims to develop an artificial intelligence (AI) algorithm that automatically predicts the CVM stages in terms of growth phases using cone-beam computed tomography images. Material and Methods: A total of 30,016 slices were obtained from 56 patients with an age range of 7–16 years. After cropping the region of interest, a convolutional neural network (CNN) was built to classify the slices based on the presence of a good vision of vertebrae. The output was used to train another model capable of categorizing the slices into phases of growth, which were defined as Phase I (prepubertal), Phase II (circumpubertal), and Phase III (postpubertal). After training the model, 88 new images were used to evaluate the performance of the model using multi-class classification metrics. Results: The average classification accuracy of the first and second CNN-based deep learning models was 96.06% and 95.79%, respectively. The multi-class classification metrics also showed an overall accuracy of 84% for predicting the growth phase in unseen data. Moreover, Phase I ranked the highest accuracy in terms of F1-score (87%), followed by Phase II (83%) and Phase III (80%). Conclusion: Our proposed models could automatically detect the C2–C4 vertebrae and accurately classify slices into three growth phases without the need for annotating the shape and configuration of vertebrae. This will result in the development of a fully automatic and less complex system with reasonable performance.