Scholarlink

<p>Early diagnosis and accurate treatment planning are critical in dentistry, as impacted mandibular third molars (MTMs) can lead to complications such as infection, pain, and damage to adjacent teeth. Panoramic radiographs (PRs), routinely used in clinical practice, require precise classification of MTM status (erupted or impacted) to support effective surgical decision-making. This study proposes the explainable mandibular third molar YOLO (E-MTMYOLO) architecture, which integrates YOLOv5 with EigenCAM—a eXplainable Artificial Intelligence (XAI) technique—to detect and interpret MTMs in PR images. For this purpose, a novel expert-annotated dataset named ExAn-MTM, consisting of 973 PRs, was developed and publicly released in this study. Image preprocessing techniques, including median filtering and gamma correction, were applied to enhance image quality, resulting in a preprocessed dataset. An ablation study was conducted to determine the optimal YOLOv5 variant and hyperparameter configuration. The best performance was achieved using YOLOv5s on the preprocessed dataset, yielding 97.77% accuracy, 95.91% sensitivity, 97.67% precision, 96.78% F1 score, and 99.21% mAP@50. To assess interpretability and clinical relevance, expert dentists from multiple institutions evaluated the model's predictions and the XAI visualizations via MTMX-CDSS, a clinical decision support system developed. The expert evaluations were analyzed using statistical methods, and the results revealed high inter-rater reliability along with strong internal consistency, confirming the clinical applicability of the proposed method. In conclusion, the proposed method not only surpasses existing approaches in core performance metrics and provides clinically validated, interpretable outputs, but also demonstrates the computational scalability and efficiency required for high-performance AI-driven diagnostic systems in dentistry.</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

E-MTMYOLO: an explainable YOLOv5-based architecture for accurate detection of mandibular third molar using a novel expert-annotated dataset

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • İsmail Kayadibi,
  • Utku Köse,
  • Gür Emre Güraksın,
  • Bilgün Çetin

摘要

Early diagnosis and accurate treatment planning are critical in dentistry, as impacted mandibular third molars (MTMs) can lead to complications such as infection, pain, and damage to adjacent teeth. Panoramic radiographs (PRs), routinely used in clinical practice, require precise classification of MTM status (erupted or impacted) to support effective surgical decision-making. This study proposes the explainable mandibular third molar YOLO (E-MTMYOLO) architecture, which integrates YOLOv5 with EigenCAM—a eXplainable Artificial Intelligence (XAI) technique—to detect and interpret MTMs in PR images. For this purpose, a novel expert-annotated dataset named ExAn-MTM, consisting of 973 PRs, was developed and publicly released in this study. Image preprocessing techniques, including median filtering and gamma correction, were applied to enhance image quality, resulting in a preprocessed dataset. An ablation study was conducted to determine the optimal YOLOv5 variant and hyperparameter configuration. The best performance was achieved using YOLOv5s on the preprocessed dataset, yielding 97.77% accuracy, 95.91% sensitivity, 97.67% precision, 96.78% F1 score, and 99.21% mAP@50. To assess interpretability and clinical relevance, expert dentists from multiple institutions evaluated the model's predictions and the XAI visualizations via MTMX-CDSS, a clinical decision support system developed. The expert evaluations were analyzed using statistical methods, and the results revealed high inter-rater reliability along with strong internal consistency, confirming the clinical applicability of the proposed method. In conclusion, the proposed method not only surpasses existing approaches in core performance metrics and provides clinically validated, interpretable outputs, but also demonstrates the computational scalability and efficiency required for high-performance AI-driven diagnostic systems in dentistry.