Automatic tooth segmentation method based on mask scoring RCNN in cone-beam CT images

Automated segmentation of the tooth is essential to accuracy analysis for clinical decision and treatment planning. Accurate tooth segmentation plays a vital role in digital orthodontic simulations for 3D reconstruction. In cone beam computed tomography (CBCT) images, the blurred borders of adjacent teeth or the similar intensity of teeth and bone make teeth segmentation tremendously tricky. We proposed a method focused on improving the Mask Scoring RCNN by combining feature enhancement and attention modules. First, we enhance the original image features by obtaining edge maps. Then, we adjust the weight of the feature maps by introducing an attention module to improve the accuracy. Experimental results show that compared with other models, our method can detect tooth root information more clearly and reduce the segmentation error of bones.