Identification of unknown individuals through a feature-based comparison of ante- and postmortem radiological images using Computer Vision

• Applicant: Dr. Andreas Heinrich
• Subject Area: Toxicology, Laboratory Medicine ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Endocrinology, Diabetology, Metabolism ; Criminology ; Medical Physics, Biomedical Technology ; Nuclear Medicine, Radiotherapy, Radiobiology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 514572362

For victims of crimes, accidents and mass disasters, identification of unknown bodies is a complex and challenging task. The goal of this project is a systematic and fundamental comparison of feature extraction and feature matching methods for unambiguous identification of persons using computer vision (CV). In the project, fundamental research is to be conducted into the extent to which CV features can contribute to narrowing down the group of persons of an unknown corpse and are thus suitable as a new tool, in addition to the public search of the criminal investigation department, for the purposeful collection of reference materials for a legally binding identification procedure. The identification method should allow a unique identification of a single person from orthopantomograms (OPG) of at least 100,000 possible individuals. Different CV feature extraction methods, such as SIFT, SURF, AKAZE, KAZE, BRISK, BRIEF, and ORB will be systematically compared in terms of match points obtained between same/different individuals, identification success rate, signal processing duration, and a significance threshold to assess identification confidence. Using artificial neural networks (KNN), the identification method will be both further accelerated and made more robust. First of all, a KNN will be used to evaluate a possible improvement in accuracy and signal propagation time due to a reduction of the area to be evaluated to the tooth rows and a segmentation of the teeth, possibly also tooth implants, according to the FDI tooth scheme. Furthermore, the aim is to estimate the approximate age of the cadaver objectively and automatically with another KNN. Computed tomography (CT) virtopsy is a simple procedure compared to postmortem OPG acquisition. For this reason, the postmortem CV features of virtopsia will be compared with CV features from antemortal CT datasets and later with radiographs (cross-modality identification). In this context, we will fundamentally explore whether a KNN can support the computation of a curvilinear multiplanar reformatting (MPR) from CT datasets of the head to obtain an OPG-like representation. Furthermore, the applicability of the identification method will be evaluated on CT and X-ray images of other body regions, such as thorax, jaw and skull, in addition to OPG data sets. For this purpose, different KNNs will automate a segmentation of defined regions in X-ray images and CT slice images in order to subsequently apply the CV.

• DFG Programme: Research Grants
• Co-Investigators: Professorin Dr. Gita Mall; Professor Dr. Ulf Karl-Martin Teichgräber