Graphics and Vision Research Group (GraVis), University of Basel, Basel, Switzerland.
Institute of Forensic Medicine, University Hospital of Mainz, Mainz, Germany.
Forensic Sci Med Pathol. 2024 Mar;20(1):23-31. doi: 10.1007/s12024-023-00590-w. Epub 2023 Mar 9.
The identification of teeth in 3D medical images can be a first step for victim identification from scant remains, for comparison of ante- and postmortem images or for other forensic investigations. We evaluate the performance of a tooth detection approach on mandibles with missing parts or pathologies based on statistical shape models. The proposed approach relies on a shape model that has been built from the full lower jaw, including the mandible and teeth. The model is fitted to the target, resulting in a reconstruction, in addition to a label map that indicates the presence or absence of teeth. We evaluate the accuracy of the proposed solution on a dataset consisting of 76 target mandibles, all extracted from CT images and exhibiting various cases of missing teeth or other cases, such as roots, implants, first dentition, and gap closure. We show an accuracy of approximately 90% on the front teeth (including incisors and canines in our study) that decreases for the molars due to high false-positive rates at the wisdom teeth level. Despite the drop in performance, the proposed approach can be used to obtain an estimate of the tooth count without wisdom teeth, tooth identification, reconstruction of the existing teeth to automate measurements taken as part of routine forensic procedures, or prediction of the missing teeth shape. In comparison to other approaches, our solution relies solely on shape information. This means it can be applied to cases obtained from either medical images or 3D scans because it does not depend on the imaging modality intensities. Another novelty is that the proposed solution avoids heuristics for the separation of teeth or for fitting individual tooth models. The solution is therefore not target-specific and can be directly applied to detect missing parts in other target organs using a shape model of the new target.
在 3D 医学图像中识别牙齿，可以作为从残骸中识别受害者的第一步，用于比较生前和死后的图像，或用于其他法医调查。我们评估了一种基于统计形状模型的方法在具有缺失部分或病变的下颌骨上检测牙齿的性能。该方法依赖于从包括下颌骨和牙齿在内的完整下颌构建的形状模型。该模型被拟合到目标上，除了指示是否存在牙齿的标签图之外，还会产生重建。我们在一个由 76 个目标下颌骨组成的数据集上评估了该方法的准确性，这些下颌骨全部从 CT 图像中提取，并且表现出各种缺失牙齿的情况，或者其他情况，如根、植入物、第一齿列和间隙闭合。我们在前牙（包括我们研究中的切牙和尖牙）上获得了大约 90%的准确率，而磨牙的准确率由于智齿水平的高假阳性率而降低。尽管性能有所下降，但该方法可以用于获得没有智齿的牙齿计数估计值、牙齿识别、现有牙齿的重建以自动测量作为常规法医程序一部分的测量值，或预测缺失牙齿的形状。与其他方法相比，我们的解决方案仅依赖于形状信息。这意味着它可以应用于从医疗图像或 3D 扫描获得的案例，因为它不依赖于成像模式的强度。另一个新颖之处在于，该解决方案避免了用于分离牙齿或拟合单个牙齿模型的启发式方法。因此，该解决方案不是针对特定目标的，可以直接应用于使用新目标的形状模型来检测其他目标器官中的缺失部分。
