Deep Learning-Based Panoramic Radiograph Retrieval from Antemortem Images for Forensic Identification

The aim of this study was to evaluate the applicability of an expert-assisted, semi-automated, Convolutional Neural Network based deep learning framework developed using panoramic dental radiographs in large-scale retrieval-based forensic identification scenarios. During the model development process, 2,440 panoramic radiographs from 842 individuals were preprocessed through masking, intensity based cropping of the region of interest, and resizing. Positive pairs were generated from images belonging to the same individual, whereas negative pairs were generated from images belonging to different individuals, and class balance was maintained. Model performance was evaluated using subject level 5-fold cross validation with four different deep learning backbones, thereby preventing images from the same individual from being shared between the training and validation sets. The results showed that the best individual performance was achieved with the ConvNeXt-Tiny model. Top-1, Top-3, and Top-5 accuracy values obtained with the ConvNeXt-Tiny model were 75.00 ± 6.51, 82.22 ± 3.16, and 83.33 ± 1.96, respectively. These findings indicate that the proposed method is effective and applicable in forensic identification practices.