Research and development of deep learning based regional detection methods for the analysis of panoramic dental images

This thesis aims to develop innovative deep learning-based models for the automatic detection of three different dental pathologies—impacted teeth, dental caries, and dental infections—in panoramic dental radiographs. Due to the time-consuming, subjective, and fatigue-prone nature of traditional manual visual evaluation methods, deep learning architectures specifically designed and optimized for each pathology type have been proposed. For impacted tooth detection, an ensemble approach combining GAN-based image enhancement with YOLOv8-L and RT-DETR-L architectures was employed. In experiments conducted on a dataset consisting of 407 panoramic images, the optimized model achieved 98.3mAP@0.5, and 99.2% recall values. For tooth decay detection, ConvNeXtV2 blocks were integrated into the YOLOv8 architecture, and hyperparameter optimization was performed using a genetic algorithm. In studies conducted on a dataset consisting of 474 images, the optimized model achieved 62.6% mAP@0.5 and 78.4% precision values; this provided an 8.7% increase in mAP@0.5 compared to the basic YOLOv8. For dental infection detection, ECA, SGE, and SKConv attention mechanisms were integrated into the YOLOv10 architecture, and genetic algorithm optimization was applied. In experiments conducted with 422 panoramic images, the optimized model achieved 64.5% mAP@0.5, 76.1% precision, and 61.8% recall, providing a 6.6% increase in mAP@0.5 compared to the baseline YOLOv10. The findings reveal that the proposed architectural improvements, optimization techniques, and ensemble approaches provide significant performance gains compared to baseline models. This study contributes significantly to deep learning applications in dental imaging and demonstrates the potential of the developed systems to support dentists' diagnostic processes, enabling faster, more objective, and accurate diagnoses.