Dental Reasearch
Bagdy-Bálint Réka
Semmelweis University, Department of Paediatric Dentistry and Orthodontics
Dr. Réka Bagdy-Bálint1
1: SE Department of Paediatric Dentistry and Orthodontics
Introduction
Artificial intelligence (AI) has been increasingly used in medicine, including orthodontics, where it can automate time-consuming tasks such as identifying anatomical landmarks on lateral cephalograms. Although AI improves time efficiency, it generates an unexpected volume of data, for which storage and reuse solutions are still lacking.
Aims
This study aimed to investigate the training process of a cascaded Convolutional Neural Network (CNN) for landmark detection on lateral cephalograms of varying quality, and to determine the speed, reliability, and clinical accuracy of the algorithm for orthodontic diagnosis. Another goal was to support comprehensive orthodontic diagnostics by structuring the data to enable its future reuse.
Method
The CNN model was trained using a total of 1600 lateral cephalograms. After adding each new dataset (400, 800, 1200, and 1600 images), the model was evaluated on a test set containing 78 images of mixed quality. Accuracy was assessed through statistical analysis of intra- and inter-examiner distance errors, examiner-versus-model comparisons, and prediction of diagnostic failures. Additionally, a structured SQL database was developed that communicates both with general analysis software and standard orthodontic diagnostic workflows.
Results
The integration of AI substantially improved time efficiency. Manual cephalometric tracing took on average 5.25 minutes longer per case compared to AI-assisted analysis, even when human correction time (104–167 seconds) was included. The best-performing model demonstrated more consistent landmark detection than either two different examiners or the same examiner on separate occasions. Angular deviations (0.05°–1.86°) and proportional errors (3.14%) remained within clinically acceptable ranges.
Conclusion
Implementing a robust AI algorithm for cephalometric analysis in orthodontics reduces variability compared to human examiners while improving both speed and accuracy. AI predictions increasingly supported clinicians in accurately locating landmarks as training progressed. Furthermore, structured data storage enables the efficient integration of AI-generated outputs into data-driven healthcare systems, bridging the gap between prediction and clinical decision-making. Future research should explore the long-term therapeutic impact of AI-assisted orthodontic diagnostics.
Funding
DI