Investigating the effect of image preprocessing methods on the segmentation of radiographic dental ımages using deep learning

Dentists use various imaging techniques to diagnose diseases related to oral and dental health. However, elements such as artifacts, blurring, and lighting issues in the obtained images can negatively affect interpretation and automatic analysis processes. In this thesis study, an innovative, domain-based framework is proposed to improve the segmentation performance of panoramic dental radiographs. The proposed framework consists of five main stages: preprocessing, region of interest extraction, image segmentation, U-Net-based segmentation, and postprocessing. In the preprocessing step, a total of 14 methods, including traditional and hybrid approaches, were applied, and the integration of these methods with segmentation was systematically investigated. The three most successful methods were adapted to the original panoramic radiographs. The regions of interest obtained through region of interest extraction were divided into 512×512 patch structures to make them suitable for input into the model. During the segmentation phase, each patch was processed using custom U-Net architectures based on ResNet50, VGG19, and EfficientNetB4 backbones, as well as the classic U-Net model. In the final phase, the predicted patches were combined to obtain holistic masks. The effectiveness of the proposed framework was demonstrated through experiments conducted on the Tufts Dental Database. The results show that the developed framework significantly improves segmentation performance. In particular, it was determined that the EfficientNetB4-based U-Net architecture exhibits higher performance in the segmentation of panoramic dental radiographs compared to other approaches. The findings emphasize the impact of dental image quality on segmentation performance and demonstrate that the proposed framework will make a significant contribution to the automatic analysis of dental radiographs.