科研之友
微信
新浪微博
Facebook
分享链接
摘要
The detection of molars in this paper is mainly for children around seven years old. The first molars of children in this age group have just erupted. We primarily check whether the teeth need to pit and fissure sealing to protect the teeth from caries. Our dataset comes from dental photos taken by mobile phones. We use these images to train the deep learning model and use the trained deep learning model to detect whether the teeth are healthy. However, this task has enormous challenges. The main difficulties are as follows: first, the teeth are closely arranged, and the individuals are relatively small, so the detection was a bit tricky. Second, the camera's shooting angles varied greatly, which might cause uneven image quality. Third, the image dataset is relatively small, which might result in the inability to obtain important features when training classification. By analyzing the dataset, we divided the task into two steps to build the model. First, object detection is used to detect the position of the first molar. Second, we classify the tested teeth into three categories. In response to the above problems, both of the two parts of the model are improved. An attention mechanism and a bounding boxes screening mechanism are added to the object detection part. For the classification part, we propose the MCGan model to extend the dataset. The dataset came from children's dental images collected in a stomatological hospital, and professional dentists annotate the dental images. For molar identification, the accuracy of our model is 98.5%, and the accuracy of tooth classification is 85.6%.
-
单位南方医科大学; 浙江师范大学
