Automatic forensic identification based on dental radiographs

The purpose of forensic identification is to determine the identity of unidentified victims. Dental radiographs are often the only means of identifying victims, but manually comparing dental radiographs is inefficient and subject to errors due to fatigue in human experts. This thesis has designed and developed an automatic system for identifying human victims based on dental radiographs. There are three stages in the proposed automatic dental identification system: (1) Feature extraction. Salient features in the dental radiograph are identified, such as the contours of the tooth and the contours of crowns or fillings. (2) Registration. Each tooth is labeled according to its location within the mouth, and missing teeth are detected. (3) Matching. Each unidentified radiograph (post-mortem or PM) is compared to a database of known individuals (ante-mortem or AM). In the first stage, namely feature extraction, contours of teeth and dental work are extracted from dental radiographs. The radiographs are first segmented into regions such that each region contains only a single tooth. Contours of individual teeth are then extracted from each region of the radiograph using active shape models (ASMs). Dental work such as crowns, fillings, and root canal treatment is composed of radio-opaque material that appears as bright regions in the radiographs. An anisotropic diffusion-based technique is used to enhance the images, and regions of dental work are segmented using image thresholding. The second step in dental radiograph-based identification is to determine which teeth are (or are not) present in the radiograph. This step is important because the matching algorithm cannot properly align AM and PM radiographs if they do not contain the same number of teeth. Thus, the second step is to label each tooth in the radiograph based on the human dental atlas, which is a descriptive model of the shapes and relative positions of teeth. A hybrid model involving Support Vector Machines (SVMs) and a Hidden Markov Model (HMM) is developed for representing the dental atlas and classifying individual teeth in dental radiograph images. The final stage of the proposed automatic identification system is to match dental radiographs of unidentified subjects against dental radiographs of known subjects in the database. The extracted tooth contours are matched for every pair of teeth with the same index in the two radiographs. When both teeth in the pair to be matched contain dental work, then the shapes of dental work are also matched. An overall matching distance for each pair of teeth is computed by combining the matching distances for the tooth contours and for the dental work. Finally, the distances between the unidentified subject and all the subjects in the database are computed based on the distances between all pairs of teeth. The database subjects are then sorted in ascending order of the matching distances and the results are provided to forensic experts. Due to poor quality of dental radiographs and complexity of the matching problem, fully automatic methods may sometimes produce errors. Monitoring schemes are designed to evaluate the image quality and correctness of the automatic results at each processing stage. User interaction is requested when errors in intermediate processing steps are encountered. Experiments on retrieving PM radiographs of 29 subjects from a database of radiographs of 33 subjects show an accuracy of 72% for top-1 retrieval and 93% for top-2 retrievals. Experiments on retrieving PM radiographs of 29 subjects from a larger database of relatively poor quality radiographs of 133 subjects show a lower accuracy of 66% for top-1 retrieval and 90% for top-13 retrievals.