Exploring the Use of Generative Adversarial Networks for Automated Dental Preparation Design

Dental decay is an important disease that requires immediate treatment. Without proper medical care, it progressively affects the rest of the body. To remove a dental decay, dentists must prepare the tooth. The process of shaping extracoronal preparations demonstrates significant variability between dental practitioners. This task is time-intensive and consumes resources, often increasing the number of patient sessions. Current deep learning models have shown remarkable results on 3D shape generation. However, their use in digital dentistry has been limited to crown generation. In this study, we propose a dental prior deformation model (DPD), combining both the unsupervised generation of dental preparations from a tooth prior and point-to-mesh reconstruction. Specifically, DPD allows dentists to generate a variety of dental preparations from a dental prior to assessing the preparation choices automatically before the operation. Experiments demonstrate that a personalized global prior-with a maximum Hausdorff distance of 6.40 mm and a maximum Chamfer distance of 6.36 mm from the tooth prior across two resolutions-produces visually accurate results suitable for margin line detection. The variety of choices from DPD helps to generate and visualize position-specific preparations before clinical operations. Our code for these experiments is available at https://github.com/ImaneChafi/DPD.git