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

Abstract

Classifying the cervical vertebral maturation (CVM) degree is helpful in determining the peak period of growth and predicting the growth rate and pattern. The current study proposed a multistage framework for automated CVM classification.The dataset consisted of 2325 lateral cephalograms. Two orthodontists independently classified the images into six classes. One object detection (Faster RCNN) and two classification models (ResNet 101) were designed using the Python programming language and PyTorch library. The First classification model classified images into two main groups (i.e., C1-C3 and C4-C6) based on the C4 vertebrae shape. The second one classified each group into its subcategories. Each classification model was trained and tested using a 10-fold cross-validation strategy. The general framework reached an accuracy of 82.96%. The object detection of ROI extraction reached the mAP50 and mAP75 of 100%. The first classifier model had an accuracy of 99.10% on the hold out test set. The classifier of C1-C3 images had higher accuracy than the C4-C6 classification model (86.49% versus 82.80%) The accuracy of this fully automated framework was promising. Considering the gradual changes in cervical vertebrae morphology the use of visualized data by gradient-weighted class activation maps (Grad-CAM) is suggested to improve the model’s performance.