Transformer-Based Tooth Segmentation, Identification and Pulp Calcification Recognition in CBCT

Abstract

The recognition of dental pulp calcification has important value for oral clinic, which determines the subsequent treatment decision. However, the recognition of dental pulp calcification is remarkably difficult in clinical practice due to its atypical morphological characteristics. In addition, pulp calcification is also difficult to be visualized in high-resolution CBCT due to its small area and weak contrast. In this work, we proposed a new method of tooth segmentation, identification and pulp calcification recognition based on Transformer to achieve accurate recognition of pulp calcification in high-resolution CBCT images. First, in order to realize that the network can handle extremely high-resolution CBCT, we proposed a coarse-to-fine method to segment the tooth instance in the down-scaled low-resolution CBCT image, and then back to the high-resolution CBCT image to intercept the region of the tooth as the input for the fine segmentation, identification and pulp calcification recognition. Then, in order to enhance the weak distinction between normal teeth and calcified teeth, we proposed tooth instance correlation and triple loss to improve the recognition performance of calcification. Finally, we built a multi-task learning architecture based on Transformer to realize the tooth segmentation, identification and calcification recognition for mutual promotion between tasks. The clinical data verified the effectiveness of the proposed method for the recognition of pulp calcification in high-resolution CBCT for digital dentistry.