Bridging the Patient Distribution Gap for Robust 3D Tooth Segmentation

Hao Chen, Jianfei Yang, Huimin Xiong, Yang Feng, Jin Hao, Zuozhu Liu

Keywords: Tooth Segmentation, Domain Adaptation, 3D Point Cloud

TL;DR: Bridge the gap of patient distribution difference by unsupervised domain adaptation for robust 3D tooth segmentation.

Abstract: Empowered by deep learning, 3D semantic segmentation algorithms have been applied to computer-aided dental systems in recent years. Existing models yield satisfactory performance for 3D point cloud data. However, there exists a common assumption that the data distribution of training and testing scenarios is the same, which limits the model capacity to deal with cross-domain situations. In real-world clinical scenarios, the patients for evaluation could be quite different from those for training in terms of their teeth symptoms, such as teeth defects and eruption. To deal with this problem, we borrow the idea of Domain Adaptation (DA) and propose a Domain-Invariant Tooth Segmentation (DITS) framework that bridges the clinically symptomatic domain gap. DITS leverages a dynamic graph convolutional neural network for 3D semantic segmentation backbone, while maximizing the probabilities of well-classified samples during evaluation by maximum square loss, thus adapting the 3D segmentation model to a realistic domain with different teeth symptoms. In the experiment, the real-world datasets are collected including 4272 3D IOS scans which are annotated with tooth-ID and three common tooth symptoms by experts. Extensive experiments have shown that DITS leads to a significant improvement for the large-scale cross-domain 3D tooth segmentation.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: both

Primary Subject Area: Transfer Learning and Domain Adaptation

Secondary Subject Area: Segmentation

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.