Automated cutaneous landmark prediction based on CBCT-derived skeletal landmarks

This study explored the potential of artificial intelligence (AI) to predict cutaneous landmark positions from skeletal landmarks, with applications in maxillofacial surgery planning and forensic reconstruction. A dataset of 137 cone-beam computed tomography scans was annotated with skeletal landmarks (Nasion, Sella, Menton, and bilateral Gonion) as model inputs and facial landmarks (Sellion and bilateral Ala) as prediction targets. Five AI models were evaluated: linear regression (LR), random forest and three feed-forward neural network architectures. The LR model outperformed other approaches, achieving a mean Euclidean error between the predicted and actual landmark positions of 2.65±1.36 mm for Sellion and 2.16±0.79 and 1.79±0.72 mm for right and left Ala, respectively. This study confirms the feasibility of using AI to estimate cutaneous landmarks from skeletal data. The approach shows promise for improving maxillofacial surgical planning and forensic analyses, offering an automated prediction of the treatment outcome and data for facial reconstruction. Future work could refine model architectures for broader clinical adoption and test the model on clinical case.