Abstract
A fully automated personal identification method is crucial for mass disaster response. This study evaluated an approach using shape-based radiomic features of thoracic vertebral bodies on CT scans. This retrospective study included 66 individuals with both antemortem and postmortem CT scans (2007-2014) and 1018 antemortem cases from the Lung Image Database Consortium image collection. Thoracic vertebral bodies (T1-T12) were segmented, and 14 shape-based radiomic features were extracted. Outlier detection using the Mahalanobis distance excluded vertebral bodies with atypical feature distributions. Personal identification was performed using Euclidean distance-based similarity scores, ranking the ten most similar antemortem cases for each postmortem case. The Mann-Whitney U test compared similarity scores, and the Youden index determined the optimal similarity score threshold for one-to-one verification. Of the 66 individuals, five cases were excluded due to outlier detection. A top-1 match rate of 98.4% (60/61) was achieved for the remaining 61 postmortem cases. Similarity scores for the top-1 rank (median [interquartile range], 0.910 [0.899-0.918]) were significantly higher than those for the top-2 rank (0.834 [0.819-0.853], P < 0.001). The area under the receiver operating characteristic curve was 0.99938, with an optimal similarity score threshold of 0.832, enabling clear differentiation between matches and nonmatches. An automated identification method using shape-based radiomic features of thoracic vertebral bodies on CT scans achieved near-perfect top-1 accuracy, demonstrating its potential for victim identification in mass disasters.