DentaGAN: GAN-Based Synthetic Individual Dental Data Generation in Radiographic Images

Abstract

Panoramic radiographs are a low radiation exposure type often used as a data source for many deep learning algorithms. On the other hand, the operational structure of a traditional deep learning algorithm requires a large amount of data, which is a major problem for many researchers. It is aimed to overcome this problem through deep GAN models, many versions of which have been developed recently. The main purpose of the study is to generate a two-stage GAN model for data with the same image dimensions. The study is carried out in the form of inputting panoramic images containing a whole view, as well as single tooth data whose performance is desired to be measured, to the architecture. The generator model created for each tooth object in all panoramic radiographs generates new tooth objects that the model has yet to encounter in the dataset. Fréchet Inception Distance was used as a performance metric by measuring the distance for the Inception-v3 activation distributions for the real samples in the generated and training set. Thus, the statistical similarity of these two groups obtained from the experimental results was observed in the part of the experimental results. The cropped individual tooth classes were much more successful than the entire panoramic dataset.