Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Machine learning for performance

Single-model deep learning approach for simultaneous cervical vertebral maturation staging and skeletal jaw relationship on lateral cephalograms using YOLOv8 and CNN

Alotaibi NM; Kamel AM; Khalil AT; Hammad SM

Originally published in BMC Oral Health · 2026-06-23

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

AbstractBackgroundThis study aimed to develop an integrated artificial intelligence (AI) pipeline for cervical vertebral maturation (CVM) staging and skeletal jaw relationships and to validate its output against the results of human observers.MethodsA total of 720 lateral cephalograms were collected from the archives of the orthodontic department. The participants’ ages ranged from 8 to 18 years. Cephalograms were categorized into six cervical stages based on McNamara’s criteria and classified into skeletal Class I, II, and III patterns based on their ANB angles. The dataset was divided into a training set (n = 540) and a test set (n = 180). The training set was used to train a convolutional neural network (CNN) and a YOLOv8 model. The test set of cephalograms was coded and randomly assigned to two orthodontists for comparison with…

Evaluation of the results – performance and statistical analysis

Cohen’s kappa and weighted kappa were analyzed in IBM SPSS Statistics (Version 27, IBM Corp, Armonk, NY, USA) to assess interobserver and intraobserver reliability. To evaluate inter-examiner reliability, 100 lateral cephalometric radiographs were randomly selected and read by the two examiners independently according to the evaluation criteria. For intra-examiner reliability, the time interval between two independent assessments of the same image was 3 weeks. Several criteria were applied to assess the algorithm’s robustness. The classification model’s performance was evaluated quantitatively using accuracy, specificity, sensitivity, F1 score, and training and validation curves. Another way to measure the network’s performance was through the confusion matrix, which shows the correctly and incorrectly predicted classes. The confusion…

Discussion

Because of the significant variation in growth patterns among individuals, an accurate assessment of an individual’s skeletal classification and growth status is essential for effective diagnosis and treatment planning. The general CNN architectures and the specialized YOLOv8 model, which were used in the current study, are revolutionizing orthodontics by automating and improving diagnostic image analysis. YOLOv8 is fast and accurate, capable of identifying and segmenting objects. Therefore, it is useful for the initial detection and localization of anatomical structures and pathologies before performing further classification and detailed analysis with other more specialized CNN models [23, 24]. Therefore, CNN and YOLOv8 were deemed the most appropriate choices for our study’s goals. The use of artificial intelligence (AI) for…

Editorial reading guide

  • What data created and tested the model, and was evaluation independent?
  • What is the cost of a false positive, false negative or out-of-distribution decision?