Identification using deep learning models in panoramic radiographs

Aim: Dental biometrics enables the identification of individuals using dental data and findings. Dental images of deceased individuals are analysed, especially in cases where soft tissues are damaged as a result of disasters or accidents. In this process, reliable identification is made by comparing features such as tooth contours, restorations, number, shape and position. Although there are existing studies on biometric identification from radiographic dental images, more research is needed. In this study, it is aimed to perform identification on dental panoramic radiographs using deep learning models, to compare the performance of these models and to determine the most successful model. Materials and Methods: In this retrospective study, 986 panoramic radiograph images of 450 patients were analysed. Ten datasets containing a total of 9860 ROIs were created by selecting 10 different Region of Interest (ROI) from panoramic radiographs. The pre-trained deep learning models DenseNet-201, VGG-16, VGG-19, MobileNet and ResNet-50 were used to extract features from these ROIs. Cosine similarity, Euclidean distance, Jaccard similarity, Sørensen-Dice coefficient and Pearson correlation metrics were applied to calculate the similarity between the feature vectors. As a result of these evaluations, the most similar images were identified and identification accuracy was calculated. Results: In this thesis, the highest performance for identification on panoramic radiographs was achieved with the DenseNet201 model in ROI 5, which includes the right half of the maxilla and mandible and the teeth in the right region of the maxilla and mandible. According to Jaccard and Sørensen-Dice similarity measures, the Rank-1 matching rate in this model was 90,67%. The highest performance results of all models were again observed in ROI 5. In the comparison of similarity metrics between ROI regions, the highest success was obtained with the MobileNet model and the lowest success was obtained with the ResNet-50 model. Conclusion: It was determined that deep learning models can be successfully used in identification processes in panoramic radiographs. The results of the study showed that the effectiveness of deep learning in the field of identification can make significant contributions to forensic identification systems, especially in mass disaster situations. Keywords: Panoramic radiograph, Deep learning, Identification, Forensic dentistry, Artificial intelligence