Evaluation of fixed restorations on panoramic radiographs using deep learning and auto-crop

The objective of this dissertation is to investigate the capabilities of deep learning techniques, particularly Convolutional Neural Networks (CNNs) in dental imaging. While CNNs have been extensively used in various fields, their application in dental imaging is still limited. This thesis seeks to bridge this gap by creating a unique dataset comprising 20,973 panoramic radiographs that have been categorized into five distinct groups by three dental experts. The dataset was used to train the CNN models, including AlexNet, VGG-16, and variants of ResNet models. The models were trained using 10-fold cross-validation and data augmentation techniques to ensure robustness. The evaluation results indicated that the ResNet-101 model achieved the highest accuracy of 92.7% and the highest macro-average AUC of 0.989. Additionally, other models performed well also with accuracy scores ranging from 75.5% for AlexNet to 92.1% for Inception ResNet V2. The best result was improved to 94.5% accuracy and 0.993 macro-average AUC with the introduced auto-crop optimization that emerged from efforts to reduce the difficulty level of the dataset. These findings clearly showcase the potential of CNNs in dental imaging and pave the way for the creation of computer-aided diagnosis systems that can provide valuable auxiliary information immediately to dentists upon obtaining a patients' panoramic radiograph. Furthermore, it should be noted that the proposed dataset has significant versatility as it can be re-labeled for different problems and utilized in various studies. Thus, it represents a valuable resource for advancements in dental imaging research. Similarly, the auto-crop may be utilized for many scenarios as an end-to-end network layer. Overall, this thesis highlights the potential of deep learning techniques in dental imaging to improve diagnostic accuracy and efficiency of dental care, along with the performance gain achieved by differentiable auto-cropping alteration, and lays a strong foundation for future research on the application of these techniques.