Scholarlink

<p>Comprehensive analysis of orthopantomograms (OPGs) is essential for clinical diagnosis, treatment planning, and dental monitoring. This study presents an integrated deep learning framework for automated detection, segmentation, and labeling of teeth and common restorations—crowns, bridges, and implants—in OPGs. The system incorporates a YOLOv12m-based object detector followed by a modified U-Net with a ResNet-50 encoder for accurate segmentation. To address class imbalance and the low frequency of restoration instances, data augmentation techniques were applied. Model performance was rigorously validated through five-fold cross-validation, with the segmentation model achieving a mean Dice score of 94.53% and an IoU of 89.7%. The detection model consistently maintained F1-scores above 98% for most teeth and achieved perfect precision for implants, reaching an overall mAP of 99%. Despite challenges from anatomical overlaps and underrepresented restoration classes, the proposed comprehensive system demonstrated robust and consistent performance, highlighting its clinical potential for automated dental diagnostics and treatment planning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated detection, segmentation, and labeling of teeth and restorations in OPGs using deep learning

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Md Nafiur Rahman,
  • Shahbaz Ahmed,
  • Md Abu Shahid Chowdhury,
  • Muhammad Sharifujjaman,
  • Md Anas Ali

摘要

Comprehensive analysis of orthopantomograms (OPGs) is essential for clinical diagnosis, treatment planning, and dental monitoring. This study presents an integrated deep learning framework for automated detection, segmentation, and labeling of teeth and common restorations—crowns, bridges, and implants—in OPGs. The system incorporates a YOLOv12m-based object detector followed by a modified U-Net with a ResNet-50 encoder for accurate segmentation. To address class imbalance and the low frequency of restoration instances, data augmentation techniques were applied. Model performance was rigorously validated through five-fold cross-validation, with the segmentation model achieving a mean Dice score of 94.53% and an IoU of 89.7%. The detection model consistently maintained F1-scores above 98% for most teeth and achieved perfect precision for implants, reaching an overall mAP of 99%. Despite challenges from anatomical overlaps and underrepresented restoration classes, the proposed comprehensive system demonstrated robust and consistent performance, highlighting its clinical potential for automated dental diagnostics and treatment planning.