ALIDDM : Automatic Landmark Identification in Digital Dental Models

Objectives: To develop automated methods for landmark identification in digital dental models via FlyByCNN a multi-view approach for 3D shape analysis that combines rendering techniques and state of the art neural networks such as Unets.
Methods: 80 digital dental models were pre-processed by 3 clinicians who placed landmarks on each tooth crown. An algorithm (DentalModelSeg and Universal Labeling, Merging and Splitting (ULMS)) isolates each dental crown so that each tooth has its own associated landmarks. We train a Unet via the FlyByCNN approach by capturing 2D renderings of each dental crown from different view points and using the landmarks renderings as targets.
Results: Training of the algorithm is The Medical Open Network for AI (MONAI) enabled high-performance open-source training workflows and reproducible implementations of automated approaches for landmark identification in digital dental models. Prediction of the landmarks with the trained model and utilization of the trained model to predict the landmarks on new patients allows clinicians to verify and adjust landmark location when needed.
Conclusions: Our preliminary results indicate the clinical applicability and reproducibility of the ALIDDM tool. The automatically identified landmarks can be used to plan and quantify tooth move-ment, as well as to achieve standardized orientation and registration of longitudinal intra-oral scans.