Akademik Veri Yönetim Sistemi
1st International Conference on Emerging Technologies and Engineering Systems, ICETES 2026, Hybrid, Amman, Ürdün, 7 - 09 Nisan 2026, ss.100-105, (Tam Metin Bildiri)
Recently, state-of-the-art artificial intelligence models become increasingly important for the automatic segmentation of dental structures in panoramic X-rays. Diagnostic support, quantitative analysis, and treatment planning can be easily done by this automated segmentation. This study compares six different deep learning models' (FCN-ResNet50, DeepLabV3, HRNet, U-Net, MedT, and SegNet) segmentation performance on panoramic radiographs obtained from the TUFT Dental Dataset. All of these deep learning models were trained with identical data, and evaluated using established metrics including Dice, IoU, Precision, Recall, F1, and mAP. Results indicate that the highest performance is achieved by U-Net (Dice: 0.8595, F1: 0.8775), followed closely by SegNet. These two models achieved their successful performance by their effectiveness in indentation detections. In boundary-sensitive evaluation, U-Net and SegNet achieved the highest Mean Surface Dice Scores (0.3651 and 0.3653, respectively), which indicates superior contour preservation. FCN-ResNet50 exhibited geometric inaccuracies, reflected by the highest MHDS (47.48) and near-zero surface overlap. It also produced the most false positives due to low precision, which makes FCN-ResNet50 the weakest overall performance. DeepLabV3 and HRNET models showed moderate boundary agreement and accuracy. Overall, encoder-decoder architectures are more effective for capturing fine anatomical details. This feature makes them advantageous for clinical applications. Future work will focus on real-time deployment, improved generalization across datasets, and integration with diagnostic workflows.