Abstract
With the advent of 3D printing and additive manufacturing of dental devices, IntraOral scanners (IOS) have gained wide adoption in dental practices and allowed for efficient workflows in clinical settings. Accurate automatic identification of dental landmarks in IOS is required to aid dental researchers and clinicians to plan and assess tooth position for crown restorations, orthodontics movements, and/or implant dentistry. In this paper, we present a new algorithm for Automatic Landmark Identification on IntraOralScans (ALIIOS), that combines image processing, image segmentation, and machine learning approaches to automatically and accurately identify commonly used landmarks on IOSs. Four hundred and five digital dental models were pre-processed by 3 clinician experts to manually annotate 5 landmarks on each dental crown in the upper and lower arches. Our approach uses the PyTorch3D rendering engine to capture 2D views of the dental arches from different viewpoints as well as the target 3D patches at the location of the landmarks. The ALIIOS algorithm synthesizes these 3D patches with a U-Net and allows accurate placement of the landmarks on the surface of each dental crown. Our results, after cross-validation, show an average distance error between the prediction and the clinicians' landmarks of 0.43 +/- 0.28mm and 0.45 +/- 0.28mm for respectively lower and upper occlusal landmarks, and 0.62 +/- 0.28mm for lower and upper cervical landmarks. There was on average a 5% error of landmarks more than 1.5mm away from the clinicians' landmarks, due to errors in landmark nomenclature or improper segmentation. In conclusion, we present and validate a novel algorithm for accurate automated landmark identification on intraoral scans to increase efficiency and facilitate quantitative assessments in clinical practice.