Graduation Date

Fall 12-16-2023

Document Type


Degree Name

Master of Science (MS)


Medical Sciences Interdepartmental Area

First Advisor

Meenakshi (Minnie) Vishwanath

Second Advisor

Peter J. Giannini

Third Advisor

Gregory G. Oakley


Introduction: Skeletal maturation age is an important guide in growing adolescent patients to determine the proper time to start orthodontic treatment. The most practical and commonly used method to determine skeletal age in today’s clinical practice is the Cervical Vertebral Maturation (CVM) method [2]. Based on the morphology of the cervical vertebrae, six maturational stages can be determined – designated CS 1 to CS 6. The main drawback of this technique is the time required to determine CS number. Objective: In this research a machine learning (ML) algorithm to automatically determine CVM staging from the lateral cephalometric radiograph was developed. We hypothesize all preprocessing steps for Machine Learning based CVMS classification can be automated using Artificial Intelligence. Our sub-hypothesis are as follows: 1) deeper Convolutional Neural Network (CNN) improves classification accuracy compared to shallow CNN 2) entropy filter improves classification accuracy 3) large training dataset size improves classification accuracy. Methods: Cephalograms from the Michigan Growth Study collected from the American Association of Orthodontists Foundation (AAOF) Craniofacial Legacy were classified into one of 6 stages by two orthodontists. A new ML algorithm was created to take as input full radiographs and output the CVMS classification. The ML algorithm utilized an object detection (OD) model for automated vertebrae detection. The resultant images are output to a postprocessing stage for cropping and filtering [7] and finally passed through a ML classification algorithm. Results: The object detection model resulted in 100% accuracy in detecting the region of interest (C2, C3, C4 vertebrae). The overall CVMS classification accuracies ranged from 68 to 81% with 73% combined average. No difference was observed in classification accuracy between shallow and deep CNNs. The application of an Entropy filter showed a non-statistically significant improvement in classification accuracy (P = 0.2319). CNNs trained with larger datasets had a statistically significant improvement in classification accuracy compared to those trained with a smaller, balanced data distribution (P = 0.0274). The combined speed and accuracy of the model constitute a useful clinical tool to supplement practitioners with a quick second opinion for CVMS determination.


2023 Copyright, the authors

Available for download on Sunday, December 07, 2025