Heterogeneous data fusion and loss function design for tooth point cloud segmentation

Tooth point cloud segmentation plays an important role in the digital dentistry, and has received much attention in the past decade. Recently, methods based on the graph neural network have made significant progress. However, the development has been hindered by two challenges: (1) the heterogeneous geometry data are analyzed separately or combined linearly which leads to a semantic gap in different streams; (2) there is mis-alignment between the loss function and evaluation metrics in the segmentation task. In this paper, a novel interacted graph network is presented that combines cues from heterogeneous geometry data by extending the graph attention architecture to propagate information among the different graphs. Moreover, in this paper, an approach is designed to search the segmentation loss function based on the computation graphs according to the evaluation metrics, and the evolution algorithm is revised to avoid potential loss and equivalent loss functions. Our method and other methods use the Shining3D Tooth Segmentation dataset, with experimental results compared in terms of accuracy.

Automated summary

Tooth point cloud segmentation is crucial in digital dentistry, but has challenges of analyzing heterogeneous geometry data and aligning loss function with evaluation metrics. This paper presents an interacted graph network to address these challenges, and compares experimental results using the Shining3D Tooth Segmentation dataset.