Scholarlink

Background <p>Dental radiographs are a primary modality for human identification in disaster victim identification, when fingerprint and visual recognition are not feasible. As antemortem (AM) databases grow, exhaustive AM–postmortem (PM) comparisons by expert visual inspection scale poorly to large reference databases. An automated method that pre-screens an AM database into a ranked shortlist of candidates for expert verification would address this bottleneck.</p> Methods <p>The framework integrates a Detection Transformer (DETR) with a Circle Loss objective. DETR detects individual teeth and outputs 256-dimensional embeddings from the detection queries. Circle Loss then optimizes these embeddings to minimize same-individual pair distances and maximize different-individual pair distances. Evaluated on 1,029 individuals with paired AM–PM DPRs, identification performance was assessed by deriving the retrieval indices including Rank-K accuracy, mean average precision (mAP), and normalized discounted cumulative gain (nDCG). This detection-based approach was benchmarked against representative convolutional neural network (CNN)-based detectors, including Faster Region-based CNN (Faster R-CNN), RetinaNet, and You Only Look Once (YOLOv9), as well as image-level representations that bypass tooth localization.</p> Results <p>The framework achieved 65.4% Rank-1 and 87.4% Rank-10 accuracy, 72.8% mAP, and 75.3% nDCG, consistently outperforming CNN-based detection architectures. Detection-based structural representations improved Rank-1 accuracy by 44.4% over image-level counterparts, with embeddings capturing fine-grained anatomical details such as root curvature and arch configuration.</p> Conclusions <p>The proposed detection-based metric learning framework provides a practical pre-screening tool that compresses exhaustive AM–PM database comparisons into a ranked shortlist of candidates for forensic expert verification. The system is intended to operate upstream of expert verification rather than to replace in disaster victim identification and missing person investigations.</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 forensic human identification on dental panoramic radiographs using transformer-based detection with supervised metric learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Hye-Ran Choi,
  • Sang-Heon Lim,
  • Ji Yong Han,
  • Sujeong Kim,
  • Su Yang,
  • Jo-Eun Kim,
  • Kyung-Hoe Huh,
  • Sam-Sun Lee,
  • Won-Jin Yi,
  • Min-Suk Heo

摘要

Background

Dental radiographs are a primary modality for human identification in disaster victim identification, when fingerprint and visual recognition are not feasible. As antemortem (AM) databases grow, exhaustive AM–postmortem (PM) comparisons by expert visual inspection scale poorly to large reference databases. An automated method that pre-screens an AM database into a ranked shortlist of candidates for expert verification would address this bottleneck.

Methods

The framework integrates a Detection Transformer (DETR) with a Circle Loss objective. DETR detects individual teeth and outputs 256-dimensional embeddings from the detection queries. Circle Loss then optimizes these embeddings to minimize same-individual pair distances and maximize different-individual pair distances. Evaluated on 1,029 individuals with paired AM–PM DPRs, identification performance was assessed by deriving the retrieval indices including Rank-K accuracy, mean average precision (mAP), and normalized discounted cumulative gain (nDCG). This detection-based approach was benchmarked against representative convolutional neural network (CNN)-based detectors, including Faster Region-based CNN (Faster R-CNN), RetinaNet, and You Only Look Once (YOLOv9), as well as image-level representations that bypass tooth localization.

Results

The framework achieved 65.4% Rank-1 and 87.4% Rank-10 accuracy, 72.8% mAP, and 75.3% nDCG, consistently outperforming CNN-based detection architectures. Detection-based structural representations improved Rank-1 accuracy by 44.4% over image-level counterparts, with embeddings capturing fine-grained anatomical details such as root curvature and arch configuration.

Conclusions

The proposed detection-based metric learning framework provides a practical pre-screening tool that compresses exhaustive AM–PM database comparisons into a ranked shortlist of candidates for forensic expert verification. The system is intended to operate upstream of expert verification rather than to replace in disaster victim identification and missing person investigations.