Department of Advanced General Dentistry, Inje University Sanggye Paik Hospital, Seoul, Korea.
Artificial Intelligence Research Center, Digital Dental Hub Incorporation, Seoul, Korea.
PLoS One. 2024 Oct 24;19(10):e0312537. doi: 10.1371/journal.pone.0312537. eCollection 2024.
The aim of this study was to propose a novel method to identify individuals by recognizing dentition change, along with human identification process using deep learning. Recent and past images of adults aged 20-49 years with more than two dental panoramic radiographs (DPRs) were assumed as postmortem (PM) and antemortem (AM) images, respectively. The dataset contained 1,029 paired PM-AM DPRs from 2000 to 2020. After constructing a database of AM dentition, the degree of similarity was calculated and sorted in descending order. The matched rank of AM identical to an unknown PM was measured by extracting candidate groups (CGs). The percentage of rank was calculated as the success rate, and similarity scores were compared based on imaging time intervals. The matched AM images were ranked in the CG with success rates of 83.2%, 72.1%, and 59.4% in the imaging time interval for extracting the top 20.0%, 10.0%, and 5.0%, respectively. The success rates depended on sex, and were higher for women than for men: the success rates for the extraction of the top 20.0%, 10.0%, and 5.0% were 97.2%, 81.1%, and 66.5%, respectively, for women and 71.3%, 64.0%, and 52.0%, respectively, for men. The similarity score differed significantly between groups based on the imaging time interval of 17.7 years. This study showed outstanding performance of convolutional neural network using dental panoramic radiographs in effectively reducing the size of AM CG in identifying humans.
本研究旨在提出一种新方法，通过使用深度学习来识别牙列变化，从而识别个体。将 20-49 岁成年人的近期和过去的牙科全景放射照片（DPR）分别视为死后（PM）和生前（AM）图像。该数据集包含 2000 年至 2020 年的 1029 对 PM-AM DPR。在构建 AM 牙列数据库后，计算并按降序排列相似度。通过提取候选组（CG）来测量与未知 PM 相同的 AM 的匹配等级。提取前 20%、10%和 5%的匹配 AM 图像的 CG 中的匹配等级的百分比计算为成功率，并根据成像时间间隔比较相似度得分。在提取前 20%、10%和 5%的 CG 中，20 秒的成像时间间隔的成功率分别为 83.2%、72.1%和 59.4%。成功率取决于性别，女性高于男性：提取前 20%、10%和 5%的成功率分别为 97.2%、81.1%和 66.5%，男性为 71.3%、64.0%和 52.0%。基于 17.7 年的成像时间间隔，组间的相似度得分差异显著。本研究表明，使用牙科全景放射照片的卷积神经网络在有效减小 AM CG 大小以识别人类方面表现出色。
