Learning-based tooth segmentation and analysis in digital dentistry

Abstract

With the development of computer techniques, digital dentistry systems have been widely used in orthodontics, dental implants and restorations. In the systems, 3D tooth models are essential to provide the reference for dentists to make diagnosis or treatment planning. Typically, these tooth models can be captured by different data modalities, including 3D cone-beam CT (CBCT) images, 3D intra-oral scanning model, and 2D panoramic image. How to accurately segment and reconstruct 3D tooth models from these multi-modality data is of great importance in clinical practice.

The first part of this thesis aims to make instance tooth segmentation from CBCT images. It is 3D volumetric data including tooth crown, root and other oral tissues. Existing methods often adopt handcrafted features to delineate tooth labels, which are semi-automatic and usually sensitive to tooth boundaries with limited intensity contrast. We propose a deep learning framework, named ToothNet, that takes 3D CBCT images as input, and predicts the label and ID of each tooth object. It is the first learning-based framework for automatic tooth segmentation from CBCT images. Moreover, to capture the complicated tooth shape more faithfully, we further design a hierarchical morphology-guided network to more accurately extract tooth from CBCT images, especially on the details (i.e., tooth apices). The second part of this thesis addresses the problem of tooth segmentation from intra-oral scanning data. It is a surface representation with high-resolution tooth crown information. Previous approaches can achieve satisfactory segmentation results on normal cases; however, they fail to robustly handle challenging clinical cases such as dental models with missing, crowding, or misaligned teeth before orthodontic treatments. we propose a novel end-to-end learning-based method, called TSegNet, for robust and efficient tooth segmentation on 3D scanned point cloud data of dental models.

The third part of this thesis aims to reconstruct complete tooth models from 3D intra-oral scanning data and 2D panoramic images. However, the panoramic image is a projection image with tooth shape information in 2D, and the 3D intra-oral scanning data only includes tooth crown information. We propose an implicit function based method to a learn continuous latent space for the complete tooth shape reconstruction from both the 3D and 2D inputs. It can reconstruct patient-specific tooth models with 3D root information if CBCT images are not accessible in many countries due to massive radiation.