Improving cervical maturation degree classification accuracy using a multi-stage deep learning approach.

Abstract

Purpose: Classifying cervical vertebral maturation (CVM) stages aids in determining the peak period of growth and in predicting growth rates and patterns. This study aimed to develop a multistage framework for the automated classification of CVM.

Materials and methods: The dataset consisted of 2325 lateral cephalograms. Two orthodontists independently classified these images into 6 categories. One object detection model (Faster RCNN) and 2 classification models (ResNet 101) were implemented using the Python programming language and the PyTorch library. The first classification model divided images into 2 primary groups (CS1-CS3 and CS4-CS6) based on the morphology of the C4 vertebra. The second model subsequently classified each primary group into their respective subcategories. Each classification model was trained and evaluated using a 10-fold cross-validation strategy. The learning process of the models was visualized with gradient-weighted class activation maps.

Results: The overall framework achieved an accuracy of 82.96%. Object detection for region-of-interest extraction reached mAP50 and mAP75 values of 100%. The first classification model demonstrated an accuracy of 99.10% on the hold-out test set. The classifier for CS1-CS3 images showed higher accuracy than the classifier for CS4-CS6 images (86.49% vs. 82.80%).

Conclusion: The accuracy achieved by this fully automated framework was promising.