Radiological biometrics for patient identification using deep learning

Abstract

Radiographs are a fundamental tool for visualizing the internal anatomy of patients in healthcare settings. Although intended for medical diagnosis, the unique anatomy captured in each image can also be used for identification. Forensic radiologists routinely utilize radiographs for identifying human remains, but existing methods are laborious and not scalable to the level of adoption by healthcare institutions. Meanwhile, fingerprint biometrics and facial recognition have seen mass adoption in smartphones and other security systems thanks to major advancements in deep learning. The scalability problem of radiographs could be solved if the success of deep learning-based biometrics can be replicated in medical imaging. This thesis presents an in-depth investigation into the automatic identification of subjects from radiographs using deep learning. Through a series of experiments, the feasibility of using radiographs of various body parts was demonstrated. Further experiments showed the generalizability of the deep learning models to variations in viewpoint, with final experiments revealing the ability of deep learning models to match images of completely different body parts from the same subject together. By establishing the viability of using radiographs for identification, this thesis aims to enable the development and eventual application of radiograph-based biometrics systems and lay the foundation for further work in other imaging modalities.