Identification of dental restoration applications in panoramic radiographs using artificial intelligence

Today, the rapid evolution of artificial intelligence and deep learning technologies in healthcare has initiated a new era in dental diagnosis and treatment planning. Panoramic dental radiographs are an indispensable part of clinical routines, providing a comprehensive view of all teeth and surrounding tissues; however, manual interpretation is susceptible to subjective errors caused by factors such as fatigue, stress, and anatomical superpositions. This situation increases the need for AI-based decision support systems that can produce objective and standardized data from panoramic images. In this thesis, a hybrid architecture based on YOLOv11m and Mask R-CNN + Fast R-CNN was employed to achieve the automated detection, FDI numbering, and pixel-level segmentation of dental restorations in panoramic radiographs. The dataset consists of 300 panoramic radiographs collected from Fırat University. Containing a total of 9,783 object annotations (8,737 tooth labels and 1,046 treatment labels), the dataset defines four main treatment classes: filling, root canal treatment (RCT), crown, and implant. The YOLOv11m model was preferred for tooth detection and numbering; Spatial Rule Refiner (SRR) and Segment Anything Model (SAM) refinement modules were developed to ensure anatomical consistency and eliminate quadrant errors. During the treatment segmentation phase, restorative materials were separated at the pixel level using a hybrid structure of Mask R-CNN and Fast R-CNN. The system features an integration mechanism that automatically associates a treatment with the relevant FDI number when an IoU ≥ 0.90 overlap is achieved between a treatment mask and the tooth area. According to the findings, a success of 0.9754 mAP@0.5 and a 0.9563 F1-score were achieved in the tooth detection model. In the treatment segmentation model, means 0.9272 mDice and 0.8680 mIoU values were recorded. In class-based analyses, implant and crown treatments were detected perfectly with a 1.000 F1-score, while root canal treatment showed 0.9691 and filling showed 0.9387 F1-scores. The developed system analyzes a panoramic image in an average of 20-25 ms, producing results 79 to 200 times faster than the manual evaluation process. Consequently, this study provides a reliable, objective, and high-speed digital assistant for dental clinical workflows. Keywords: Artificial Intelligence, Deep, Learning, Panoramic Radiography, Dental Restorations, FDI Numbering System