Privacy-Preserved Federated Learning for 3D Tooth Segmentation in Intra-Oral Mesh Scans

Abstract

Semantic segmentation over three-dimensional (3D) intra-oral mesh scans (IOS) is an essential step in modern digital dentistry. Many existing methods usually rely on a limited number of labeled samples as annotating IOS scans is time consuming, while a large-scale dataset of IOS is not yet publicly available due to privacy and regulatory concerns. Moreover, the local data heterogeneity would cause serious performance degradation if we follow the conventional learning paradigms to train local models in individual institutions. In this study, we propose the FedTSeg framework, a federated 3D tooth segmentation framework with a deep graph convolutional neural network, to resolve the 3D tooth segmentation task while alleviating data privacy issues. Moreover, we adopt a general privacy-preserving mechanism with homomorphic encryption to prevent information leakage during parameter exchange between the central server and local clients. Extensive experiments demonstrate that both the local and global models trained with the FedTSeg framework can significantly outperform models trained with the conventional paradigm in terms of the mean intersection over union, dice coefficient, and accuracy metrics. The FedTSeg framework can achieve better performance under imbalanced data distributions with different numbers of clients, and its overall performance is on par with the central model trained with the full dataset aggregated from all distributed clients. The data privacy during parameter exchange of FedTSeg is further enhanced with a homomorphic encryption process. Our work presents the first attempts of federated learning for 3D tooth segmentation, demonstrating its strong potential in challenging federated 3D medical image analysis in multi-centric settings.