A Multi-stage Network with Self-attention for Tooth Instance Segmentation

Automatic and accurate instance segmentation of teeth from 3D Cone-Beam Computer Tomography (CBCT) images is crucial for dental diagnose. Although Convolutional Neural Networks (CNNs) are widely used for tooth instance segmentation, the limitations of CNNs in capturing global image information can impact model performance. Recently, Transformer models leveraging the Self-Attention mechanism have exhibited exceptional capabilities in modeling global relationships in images. In this paper, we propose a fully automated tooth instance segmentation model utilizing the Self-Attention mechanism. The model is primarily based on the Self-Attention UNETR++ network and consists of three stages. In the first stage, a V-Net is employed to identify the region of interest (ROI) containing the teeth. In the second stage, a multitask UNETR++ network is utilized to extract the centroid and skeleton of the teeth. In the third stage, another multitask UNETR++ is employed to simultaneously learn the tooth mask and boundary, leading to accurate tooth instance segmentation. Experimental results on a dataset consisting of 98 CBCT images demonstrate the efficacy of our method. It achieves a Dice score of 95.1 \(\%\) and reduces the average surface distance (ASD) to 0.14 mm.