Recently, a research paper titled "Cross-Center Model Adaptive Tooth Segmentation" by Assist Prof. Liu Zuozhu and his team at the Zhejiang University-University of Illinois Urbana Champaign Institute (ZJUI) was accepted by the top international journal in the field of medical imaging, Medical Image Analysis (MedIA, 5-year IF=11.9). The first author of this paper is Chen Ruizhe, a doctoral student at ZJUI. The corresponding author is Assist Prof. Liu Zuozhu of ZJUI. Other authors include Xiong Huimin, a 2021 doctoral student of ZJUI; Xu Ruiling, a 26’ undergraduate student of ZJUI; researchers from Nanyang Technological University in Singapore; and researchers from Angelalign, a leading company in the dental healthcare industry. The research was guided by Professor Wu Jian of Zhejiang University.
This paper innovatively proposes a Cross-Center Model Adaptive Tooth Segmentation (CMAT) framework. In practical applications, deploying existing tooth segmentation models across different medical centers often encounters challenges such as variations in data distribution and concerns about privacy protection. The CMAT framework is specifically designed to effectively address these issues. It offers significant advantages; it requires no additional annotated data and avoids the complexities and risks associated with data transmission. It enables source-free domain adaptation directly utilizing the model from the original center. Experimental validations have shown that CMAT delivers outstanding performance, and it holds broad application potential. This advancement is poised to provide critical technical support to the digital dental sector, facilitating further development and breakthroughs.
Highlights
- In response to the demand for cross-center 3D tooth segmentation, this study defines three clinical scenarios for deploying tooth segmentation models.
- This research has developed a source-free model adaptive framework for cross-center tooth segmentation, featuring three core modules: tooth-level prototype alignment, progressive pseudo-label transfer, and tooth-prior regularized information maximization.
- Utilizing real-world data, this study has constructed a large-scale cross-center tooth segmentation dataset, verifying the model's exceptional performance and potential applications across various deployment scenarios.
Introduction
In the digital oral treatment process, achieving precise automatic tooth segmentation on an Intra-oral Scanning (IOS) model is a crucial step. However, in practical application scenarios, deploying existing and well-trained IOS tooth segmentation models to different medical centers often faces two major challenges: differences in data distribution and privacy protection limitations. There are significant differences in IOS data collected by different medical centers in terms of patient age, race, and region, which may lead to a significant decline in the performance of the model. At the same time, due to privacy protection considerations and limitations on data sharing, data from the source center is often unable to be directly transmitted to the target center. To carry out new data annotation work in the target center, not only is the process cumbersome, but the cost is also high.
▲ The framework of CMAT
Therefore, this article innovatively proposes a cross-center model adaptive tooth segmentation framework (CMAT). This framework utilizes deep learning technology to successfully achieve efficient model transfer and tooth segmentation operations between different medical centers without the need for source data or additional annotated data. The CMAT framework mainly consists of three core modules, namely the tooth level prototype alignment module, the progressive pseudo label transfer module, and the tooth prior regularization information maximization module. In the specific operation process, the tooth level prototype alignment module introduces a unique tooth level prototype alignment strategy, which can accurately and effectively align the features of the target center with the tooth features of the source center, laying a solid foundation for subsequent model migration. The progressive pseudo label transfer module relies on a carefully designed progressive pseudo label transfer mechanism to gradually transfer the knowledge contained in high confidence areas to low confidence areas, achieving effective dissemination and utilization of knowledge. In addition, the module for maximizing prior regularization information of teeth combines the natural distribution pattern of teeth, cleverly introduces prior knowledge, and imposes information maximization constraints on it. This measure effectively enhances the adaptability and performance of the model when facing different data distributions, ensuring that the model can run stably and efficiently in various complex practical scenarios.
The research team conducted large-scale experiments and clinical applicability tests on the CrossTooth dataset containing 5 medical centers and the AbnTooth dataset covering cases of dental abnormalities, verifying the excellent performance of the CMAT framework in three typical scenarios (source data independent adaptation, multi-source data independent adaptation, and test time adaptation). The experimental results show that the CMAT framework not only significantly improves the accuracy of cross center tooth segmentation, but also significantly reduces the dependence on data annotation and privacy risks, providing an efficient and feasible solution for cross center digital dental diagnosis and treatment.
▲ Visualization results of cross-center tooth segmentation
This study is a collaboration led by Zhejiang University, involving a research team from various institutions, and partnered with Angelalign, a leading company in the oral healthcare device industry. It also includes collaborative efforts with teams from Nanyang Technological University, among others. The project aims to enhance the precision and efficiency of oral diagnostics and treatments through an integrated academic-hospital-industry partnership, while also reducing the costs of treatments. The goal is to deliver intelligent and efficient support tools for digital dentistry and to foster innovative research and application development at the intersection of computer science and dental medicine.
About the Author
Chen Ruizhe, a PhD candidate jointly cultivated by ZJUI, and the College of Computer Science and Technology at Zhejiang University, under the supervision of Assistant Professor Liu Zuozhu of ZJUI. Currently, he is exchange at the Singapore University of Technology and Design (SUTD) under the supervision of Professor Tony Q.S. Quek. His research primarily in fairness and personalization of LLMs and their cross application in the medical field. His research results have been published in top international conferences and journals such as NeurIPS, ACL, EMNLP, and AAAI.
Assist Prof. Liu Zuozhu is an Assistant Professor and PhD supervisor at ZJUI, and he is mainly engaged in research in the field of trustworthy artificial intelligence in medicine. He hosted multiple projects such as projects funded by the National Natural Science Foundation, Zhejiang Provincial Key Projects, and Zhejiang University Major Industrial Projects. In the past three years, as the first and corresponding author, he has published over 30 papers in flagship journals such as Cell sub journals Patterns, MedIA, IEEE TMI, and NeurIPS, or at top international conferences such as ICLR, ACL, and EMNLP. In addition, he serves as the Deputy Director of the Key Laboratory of Medical Imaging Artificial Intelligence in Zhejiang Province, a standing committee member of the Smart Medicine Branch of the Chinese Society for Stereology, and the IJCAI Area Chair of the CCF Class A Conference.
