MULTITASK LEARNING FOR GENDER IDENTIFICATION AND AGE GROUP BASED ON THE MANDIBLE ON PANORAMIC RADIOGRAPHS

Forensic odontology is commonly applied for victim identification using comparing antemortem and postmortem dental radiographs. However, in cases where a victim's teeth are incomplete or missing, the mandible bone can also be used as a robust alternative for victim identification. Gender identification and age estimation are two tasks to assist in victim identification. For multiple related tasks, the multitask learning (MTL) approach has been proven to enhance generalization performance by concurrently learning the multiple related tasks and leveraging useful information across the tasks. Therefore, in this study, we propose an MTL approach for gender identification and age group based on the mandible. We propose a model, namely the mandible radiographs MTL model, that takes panoramic radiographs of the mandible as input. We built a dataset, namely the mandible radiographs dataset comprising 120 patients' panoramic radiographs of the mandible collected from Universitas Airlangga Dental Hospital, Surabaya, Indonesia, then augmented to 600 images. The experimental results show that the augmented mandible radiographs MTL model achieved the best performance for gender identification with a mean accuracy of 99.7% and an age group of 99.5%. Our research proposal is more practical because 1 model directly produces two outputs (gender and estimated age), so it is time efficient in creating models or testing.