Teeth segmentation and recognition on dental panoramic radiographs using improved Mask RCNN

Abstract

Accurately identify the type of tooth and its morphology is essential for the planning of dental implant surgery. However, manual recognition relies on the experience of dentists and usually has low recognition accuracy. And the existing instance segmentation models are difficult to realize accurately segmentation and identification at the same time due to the characteristics of highly imbalanced in teeth scale and low contrast of the real dental panoramic radiograph datasets. For this, this study proposes an automatic tooth recognition method based on improved Mask RCNN. In which, area detection module of Mask RCNN is improved through path enhancement and balancing mechanism is designed to solve the problem of low recognition accuracy caused by insufficient feature extraction of Mask RCNN. Experiments on real panoramic radiograph dataset show that the proposed method achieves fusion extraction of tooth type and morphology information. While the average accuracy of instance segmentation mAP(0.5) and mAP(0.5:0.95) for the testing set were 96.69% and 74.2%, respectively, and the recognition rate of the missing tooth reaches 93.81%. Ablation experimental results verify the effectiveness of the proposed path enhancement and balancing mechanism in increasing the accuracy of tooth classification and segmentation. The research results can promote the application of artificial intelligence-assisted diagnosis and treatment in the field of oral implants.