Automated dental panoramic image segmentation using transfer learning based convolutional neural networks

X-Ray images used in radiological examinations are one of the basic data that dentists apply to understand and solve the patient's problem. Panoramic X-ray images from these images are more comprehensive in terms of diagnosis and treatment of the patient's problem, since they show the entire mouth region. In addition, since it includes many parts other than teeth, they are the most complex images. On the other hand, they have different image qualities due to the devices used. With these in mind, the correct processing of panoramic X-ray images is a challenging process and successful analysis is of great importance. Various classical image processing algorithms have been developed to assist experts by automating this process. Over the years, deep learning (DL) technology has become more popular and its application areas have expanded with the increase in the size of the data sets and the development of the devices used. Therefore, in addition to classical image processing algorithms, X-Ray images have been started to be analyzed using deep learning (DL) technology for many purposes. However, the dimensions of dental datasets are quite small compared to the sizes of datasets found in other application areas. The aim of this thesis is to find a solution to the small data set size problem with the transfer method and to perform a successful tooth segmentation with panoramic X-ray images. In order to eliminate the differences in the sizes of the panoramic X-ray images available, the whole data set was resized to ensure that they were in the same format. The image masks to be used in the training of the models were created with the selected image editing application. U-Net and InceptionResNet-V2 models were chosen as the pre-trained deep learning (DL) model needed for the transfer learning method and two different deep learning architectures were created. Since the InceptionResNet-V2 model is pre-trained with the public image dataset, it is available as a library. The U-Net model, on the other hand, was trained with a publicly published chest X-Ray dataset. U-Net architecture was chosen as the basic deep learning (DL) model, which will be trained on panoramic X-ray images by combining with pre-trained models. Thus, training was carried out by creating two different deep learning architectures. Data augmentation method against data memorization of the models was applied to the whole dental data set. In addition, with the early termination method, the data memorization problem is prevented. Using the created test data set, the accuracy of the models was measured with the dice coefficient and the results were compared.