Dental cast acquisitions
Subjects who met the inclusion criteria have engaged in this study after voluntarily providing written informed consent. The inclusion criteria of this study were healthy subjects without severe dental crowding, dental caries, attrition/abrasion, missing teeth, and orthodontic appliances, and agreed to participate as a subject. All experiment designs were conducted in accordance with a protocol reviewed by the Independent Ethics Committee of Tohoku University School of Dentistry, Japan (permit no: 2017-3-17). The number of subjects enrolled in this study was determined after consulting an expert in statistics.
The maxillary dental arch was considered to be used in this study because it is easier for forensic dentists to reach and record in the actual identification cases. For example, the complication sometimes arises when opening the mouth of the dead body. In such situations, the labial surface of maxillary teeth will be easier to reach and register than the mandibular teeth.
The maxillary dental condition of the participant was recorded using a dental impression tray and the alginate impression material powder Aroma Fine Mixer Type® (GC Corp., Tokyo, Japan) to fabricate the dental cast. Afterward, each dental cast was trimmed and cleaned on its surfaces from small bubbles to obtain optimum scanning results.
3D scanner and software
The 3D surface data collection of the dental cast in this study was performed using a 3D scanner Vivid 910. Vivid 910 is a non-contact 3D digitizer that employs a triangulation method. The Vivid scanner has been used for several years by clinical and forensic applications, such as examining changes in patients with cleft palates (Kitagawa et al. 2004), investigating facial changes (Kau and Richmond 2008), and analyzing bite-mark evidence (Flora et al. 2009; Evans et al. 2013).
The Polygon Editing Tool (PET) software that comes with Vivid 910 was used to align and merge multiple scanned data into a single polygonal model. The polygonal model acquired from the PET then processed in the Rapidform XOS/SCAN software (INUS Technology, Inc., Seoul, South Korea) to generate the 3D point cloud data (Fig. 1).
3D image registration
The dental cast scanning process was done by placing the dental cast on a rotating table to simplify its 360-degree rotation, with the biting edges facing upwards. Vivid 910 positions towards the dental cast was approximately 45° allowing the scanning of the entire dental cast surfaces. Each dental cast was scanned from six points of view (a field of approximately 60° could be captured in a single scan), resulting in a single 3D data from each point of view. The dental cast scanning was done twice in separate sessions.
Some algorithms have been developed in several studies to automatically align and estimate the similarities between two 3D images in a process commonly called image registration (Williams and Bennamoun 2001; Williams et al. 2003; Mian et al. 2006). A similar approach of the image registration was used in this study, starting with the adjustment of the transformation parameters of each scanned 3D data in PET software, i.e., 3D rotation and translation.
After the adjustment process in PET, all of the scanned 3D data were processed using the Rapidform XOS/SCAN software. Multiple 3D data acquired from six points of view were aligned for each subject based on the transformation parameters obtained by PET and merged into a single 3D dataset, resulting in 3D point cloud data.
Two sets of experiments were designed by superimposing the genuine and the imposter pairs, termed as experiment 1 and experiment 2 (Fig. 2). The genuine pair is a concept of information science to define the superimposition between two 3D datasets derived from the same subjects from the first and the second scanning. The imposter pair is the superimposition of two 3D datasets from the different subjects, i.e., subject 1 (first scanning) vs subject 2 (second scanning).
The 3D dataset superimposing groups were classified as follows: eight groups in experiment 1 and three groups in experiment 2. Afterward, 3D datasets between subjects (both the genuine and the imposter pairs) were aligned following the midline of the reference image as the guideline. Especially, for the right and left posterior groups in experiment 1, the 3D dataset alignment was based on the position of the second premolar teeth of the reference images. The 3D dataset from the first scan was used as the reference images, and the 3D dataset from the second scan was referred to as the moving images.
Root mean square errors (RMSEs) in both experiments were calculated using the iterative closest point (ICP) algorithm, performed in MATLAB (MathWorks, Inc., Natick, MA, USA). The ICP is an algorithm employed to minimize the discrepancies between two 3D data point clouds (Fig. 3). The RMSE can be defined as the representation of the standard deviation of the discrepancies between the observed subjects (Zhang 2014). The RMSE distribution patterns of each group were investigated in experiments 1 and 2. An overview of 3D data acquisition, processing, and data analysis are explained in Fig. 4.
Experimental design
Experiment 1
The first experiment was performed to verify the effective number of teeth and the involved dental arch segment for forensic identification. All surfaces of teeth (labial, palatal, and occlusal) were included in experiment 1.
The groups made for 3D dataset superimposition in experiment 1 were classified as follows: full arch (A); partial arch from the left to the right second incisor (B), the left to the right canine (C), the left to the right first premolar (D), the right central incisor to the right last molar (E), the left central incisor to the left last molar (F), the right second premolar to the right last molar (G), and the left second premolar to the left last molar (H). The number of included teeth and the involved dental arch segment (bilateral and unilateral arches) for each group are illustrated in Fig. 5.
Experiment 2
The second experiment was conducted by superimposing the labial surfaces of the anterior teeth between subjects. The labial surface was selected as an approach to the development of actual identification tasks, i.e., when there are difficulties in opening the victim’s mouth, recording the labial surfaces of the anterior teeth will be easier for the forensic dentist. The 3D point cloud data of the anterior teeth and bilateral dental arch groups (groups B, C, and D in experiment 1) were divided by excluding the palatal surfaces and retaining only the labial surfaces. Groups for experiment 2 (Fig. 6) were classified as follows: labial surfaces from the right to the left second incisor (B′), the right to the left canine (C′), and the right to the left first premolar (D′).
Statistical analysis
The statistical analysis in this study was carried out using IBM SPSS Statistics version 23.0 (International Business Machines Corp., Armonk, NY, USA). The scattered plot diagram described the distribution of the RMSE value of the genuine and the imposter pairs. The Mann–Whitney U test was used to analyze the significance of the difference of the RMSE value between the genuine and imposter pairs, and between experiments 1 and 2. The threshold of statistical significance was set at p < 0.05.