Detection and numbering of teeth in panoramic radiographic images using Deep Neural Networks

Abstract

Objectives: Panoramic images are used as the most common method of the initial evaluation of dental treatments. Today, with the advancement of the artificial intelligence methods, it is possible to automatically evaluate these images in order to save the clinician's time. For this purpose, the aim of the present study is to design the Convolutional Neural Networks (CNNs) for segmentation and numbering of teeth in the panoramic radiographic images. Materials and Methods: The data set includes 527 panoramic images that were selected from the archives of the Radiology Department of the Faculty of Dentistry of Tabriz. After that, the images were labeled by an oral and maxillofacial radiologist by using a bounding box for each dental area, and the number of each dental area was determined by a number according to the FDI numbering system. Then we changed the size of the images to 512x512. We used the CNNs to segment and number the teeth. The segmentation was done by using the U-Net architecture and its output after post-processing entered the VGG-16 network for numbering. Eighty percent of the data was used for network training and 10% for validation and another 10% for network testing. The test data had not been shown to the network during any of the training phases. Results: The results obtained from the U-Net network for tooth segmentation, based on the original data; sensitivity, specificity, and Dice, are 98.9, 98.4, and 95.4, respectively. Also, for teeth numbering by using the VGG-16 Network Architecture, we obtained sensitivity, specificity and accuracy equal to 98.58, 99.93 and 96.8, respectively. In the examination and diagnosis of the implant, the retained roots, and the extracted teeth, the accuracy of 98.45, 97.1, and 98.2 was obtained, respectively. Conclusion: Considering the limitation in the number of images, the obtained results are favorable compared to similar studies, and in the future, with the development of these methods, it can be a useful help in the automatic analysis of panoramic images and other dental images.