3D dental biometrics: Alignment and matching of dental casts for human identification

Abstract

A 3D dental biometrics framework and a pose invariant dental identification (PIDI) technique are proposed for human identification in this study. As best as we can tell, this study is the first attempt at 3D dental biometrics. Using 3D overcomes a number of key hurdles that plague 2D methods in dental identification. 60 Postmortem (PM) samples and 200 Ante mortem (AM) samples taken from multi ethnic Asian groups (Chinese, Indian and Malay) are used in this study. The PIDI technique includes algorithms for feature extraction, feature description and correspondence. The proposed feature extraction algorithm can extract the salient points from the scanned model of dental cast. The proposed feature description and the correspondence algorithm have been tested and shown to be more robust to rigid transformations compared with the related work. Preliminary experimental result achieves 94% rank-1 accuracy in a human-assisted process, while in an automated identification process, the rank-1 accuracy decreases to 80%. In addition, the developed methodology, as it is also feasible to be applied to identifying severely corrupted dental, could promptly provide a potential candidate list in mass disasters before expert investigation. The high accuracy, fast retrieval speed and the facilitated identification process suggest that the developed 3D framework is more suitable for practical use in dental biometrics applications in the future. The limitations and future work are also presented. It could be used adjunctively with the traditional 2D X-ray radiograph identification scheme to improve the efficiency of current identification process.

Introduction

Dental records have been well recognized as the best information for human identification purposes in severe conditions and mass disasters due to the survivability and diversity of dental structures [1]. According to the INTERPOL disaster victim identification protocol, dental records have been regarded as one of the primary identifiers [2]. In forensic dentistry, the dental remains allow identification of deceased individuals when other characteristics, such as face, fingerprint, iris etc., of victims are not available. Dental identification outperformed DNA identification in 11/9 case in 2001 and Asian tsunami in 2004 [3]. It is more efficient, less time consuming and less expensive than DNA comparison in disaster victim identification such as plane crash, war, explosion, earthquake and tsunami. Traditionally, the dental chart construction is a prerequisite and the first step of postmortem dental identification. In this dental chart, general identifiers of the deceased (e.g. gender, height, estimated age, hair color, skin color) are recorded. The distinctive features are marked for every individual tooth such as missing tooth, crown and root morphology, pathology and dental restorations [2]. Subsequently, the forensic experts make comprehensive comparison of every detail in dental charts and the X-ray radiographs to confirm victims’ identities. The potential identities of victims are confirmed or rejected based on the number of matches.

For the huge volume of cases, such as the 9/11 terrorist attack (2001), Asian Tsunami (2004) and Japan earthquake (2011), the traditional manual comparison scheme using the dental chart is labor-intensive and time-consuming. It took 40 months to identify the 2749 victims of 9/11 disaster [4], and 9 months for 2200 victims out of 190,000 (2200/190,000 = 1.1%) in 2004 Asian tsunami identification. Therefore, an efficient identification process is desired for economical and psychological reasons in disaster victim identification.

The X-ray radiograph comparison is a well-established method that provides evidence in court to serve in justice purpose. Several computer-aided post-mortem (PM) identification systems have been developed to assist in the identification process, such as the CAPMI [5], WinID [6] and DAVID [7]. These systems have primarily automated the text searching. However, the X-ray radiographs were still manually compared. During last decade, research teams from West Virginia University (WVU), Michigan State University (MSU) and University of Miami (UM) in coordination with the Criminal Justice Information Services Division (CJIS) of FBI developed a benchmark prototype of an automated dental identification system (ADIS) from 2D X-ray radiographs, which aimed to efficiently find a candidate list with identical or similar dental features to those of the Missing and Unidentified Person (MUP) [8]. The 2D identification framework mainly involves four steps [9]: image segmentation for tooth isolation [10], tooth contour and dental work contour feature extraction [10], [11], atlas registration for teeth classification [12], [13], and matching of tooth contours and dental work contours [14], [15]. However, unsolved problems and challenges limit the identification capability and accuracy of the 2D methodology, including:

• Time-consuming tooth isolation process and inaccurate contour extraction from blurred X-ray radiographs due to the low image quality. Chen et al. [16] reported that 14 out of the 25 (56%) subjects in their database could not be identified because of the image quality problem, variation of the dental structures, and insufficient number of AM radiographs.

• Since 2D radiographs are projections of 3D teeth, distortions in tooth contour shape arising from different imaging angles are often significant, resulting in incorrect matching.

• Most of the 2D techniques were developed for bitewing X-ray radiographs [10], [11], [12], [13], [14], [15], [17] which only contain two types of teeth: premolar and molar. However, techniques dealing with other types of radiographs like periapical and panoramic radiographs may be required for identification due to the conditions of the human remains. The partial teeth identification is still a challenge and has not been explored extensively. It is reported that in the 2D identification framework, the matching algorithm cannot properly align AM and PM radiographs if they do not contain the same number of teeth [18].

• Insufficient and distorted features in X-rays radiographs. The dental arch shape is considered to be unique among individuals [19]. However, the shape of dental arch is distorted in 2D radiographs, thus making it an inapplicable feature in 2D identification scheme [14], [16], [17].

• Identification efficiency is not high enough and human interaction is required in each processing step in the 2D framework. It takes about 1.7 h to retrieve one subject from 33 subjects with 72% rank-1 accuracy and it takes 7 h to retrieve one subjects from 133 subjects (PC with a 2.99 GHz Pentium 4 processor) with 66% rank-1 accuracy [18].

3D biometrics is receiving increasingly more attention than 2D biometrics. For instance, 3D face and ear recognition [20], [21] have shown promising identification capability and accuracy. With the development of real-time scanning and 3D reconstruction technologies, the acquisition of 3D models has become effortless and faster. The most recent powerful scanning technology could be used to obtain the 3D digital dental surface within 100 s [22]. In addition, there are some recent dental research work on computer-assisted 3D reconstruction of teeth from CT images for human identification [23], 3D automatic teeth segmentation for dental biometrics [24], and online 2D and 3D dental databases for dental identification [25]. While different parts of 3D dental biometrics system are being developed, none of the aforementioned studies have investigated the identification. Therefore, the present study aims to develop an automated 3D dental biometrics system with the identification methodologies and algorithms for retrieving digitized 3D dental casts to examine the identification validity and accuracy as an adjunctive mean for 2D dental biometrics.

Our previous study [26] showed that the identification novelty with a preliminary experimental result achieving 71.4% rank-1 accuracy and 100% rank-4 accuracy by matching 7 postmortem (PM) samples against 100 ante-mortem (AM) samples toward a fully automated 3D dental identification testing. The present study expands the sample size for both PM and AM database to include partial and noisy PM samples. Different shape descriptors (saliency, Gaussian curvature, integral volume) are proposed and compared with regard to pose invariant characteristics.