A Relation-based Framework for Effective Teeth Recognition on Dental Periapical X-rays

Abstract Dental periapical X-rays are used as a popular tool by dentists for diagnosis. To provide dentists with diagnostic support, in this paper, we achieve automated teeth recognition of dental periapical X-rays by using deep learning techniques, including teeth location and classification. Convolutional neural network(CNN) is a popular method and has made large improvements in medical image applications. However, in our specific task, the performance of CNN is limited by lack of data and too many teeth positions in X-rays. Addressing this problem, we consider to utilize the prior dental knowledge, and therefore we propose a relation-based framework to handle the teeth location and classification task. According to the relation in teeth labels, we apply a special label reconstruction technique to decompose the teeth classification task, and use a multi-task CNN to classify the teeth positions. Meanwhile, for teeth location task, we design a proposal correlation module to use the information in teeth positions, and insert it into the multi-task CNN. A teeth sequence refinement module is used for the post processing. Our experiment results show that our relation-based framework achieves high teeth classification and location performance, which is a big improvement compared to the direct use of famous detection structures. With reliable teeth information, our method can provide automated diagnostic support for the dentists.