Estimating occlusal vertical dimension using facial landmarks and deep learning-based models

Akıncı I., Polat A. N., Mohemmed H., Köseoğlu M., Skutjo C., Wee A., ...Daha Fazla

48th European Prosthodontic Association (EPA) and 27th Turkish Prosthodontics and Implantology Association (TPID) Annual Congress, Nevşehir, Türkiye, 11 - 13 Eylül 2025, ss.1-5, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Nevşehir
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.1-5
  • Atatürk Üniversitesi Adresli: Evet

Özet

Statement of problem. The determination of Occlusal Vertical Dimension (OVD) using traditional methods is often subjective and inconsistent, highlighting the need for a deep learning-based approach to accurately and reproducibly predict OVD from facial landmarks. Purpose. The purpose of this study is to assess the efficacy of deep learning algorithms in predicting the OVD from selected facial landmark points. MATERIALS-METHODS: The study was conducted using standardized frontal and lateral photographs of 200 individuals, comprising 100 males and 100 females. Manual measurements of facial landmarks were taken from these digital images. These images and landmark data measurements were used to train and evaluate deep learning models based on the You Only Look Once version 8 (YOLOv8) architecture, including its sub-variants (YOLOv8s, YOLOv8m and YOLOv8l), each differing in complexity and parameter count. The performance of these models was evaluated in terms of their ability to accurately detect facial landmarks and predict the corresponding OVD values. RESULTS: Among the models tested, the highest prediction accuracy was achieved by YOLOv8l, with a mean absolute error (MAE) of 1.35 mm and a coefficient of determination (R²) of 0.786, indicating a strong correlation between the predicted and actual OVD values. The results demonstrate that the deep learning approach, particularly the YOLOv8 architecture, can effectively identify relevant facial landmarks and reliably estimate OVD with high accuracy. CONCLUSION: The findings of this study support the feasibility of using deep learning-based models for the prediction of occlusal vertical dimension from facial photographs. The YOLOv8 model family, especially the large-sized variant, offers a promising tool for enhancing diagnostic precision and reducing variability in prosthodontic practice.

Keywords: Artificial Intelligence, Vertical Dimension of Occlusion, Prosthodontics, Dentistry