The aim of this investigation was to evaluate the reproducibility of a voxel-based 3-dimensional superimposition method and the effect of segmentation error on determining soft tissue surface changes.
A total of 15 pairs of serial cone-beam computed tomography images (interval: 1.69 ± 0.37 years) from growing subjects (initial age: 11.75 ± 0.59 years) were selected from an existing digital database. Each pair was superimposed on the anterior cranial base, in 3 dimensions with Dolphin 3D software (version 2.1.6079.17633; Dolphin Imaging & Management Solutions, Chatsworth, Calif). The reproducibility of superimposition outcomes and surface segmentation were tested with intra- and interoperator comparisons.
Median differences in inter- and intrarater measurements at various areas presented a range of 0.08-0.21 mm. In few instances, the differences were larger than 0.5 mm. In areas where T0-T1 changes were increased, the error did not appear to increase. However, the method error increased the farther the measurement area was from the superimposition reference structure. For individual images, the median soft tissue segmentation error ranged from 0.05 to 0.06 at various areas and in no subject exceeded 0.13 mm.
The presented voxel-based superimposition method was efficient and well reproducible. The segmentation process was a minimal source of error; however, there were a few cases in which the total error was more than 0.5 mm and could be considered clinically significant. Therefore, this method can be used clinically to assess 3-dimensional soft tissue changes during orthodontic treatment in growing patients.
The tested method is user-friendly and reproducible in detecting 3D soft tissue changes.
Superimposition and segmentation error are both considered minimal.
The tested method can be used reliably for clinical and research purposes.
The paradigm shift in modern orthodontics toward more soft tissue–based diagnostics has expanded the social role of the specialty to one that manages a greater spectrum of patient needs than straight teeth. Facial and smile esthetics have gained increased attention in the professional orthodontic community and among patients whose decision to pursue treatment is primarily driven by their wish to improve their appearance. This motive is also well supported by a large body of literature emphasizing the significant social impact of facial and smile attractiveness on various aspects of professional and personal life and on overall well-being.
Although facial esthetics have traditionally played an important role in treatment planning, , there is no widely used reliable tool for measuring soft tissue dimensions or soft tissue changes in clinical practice. Direct anthropometric measurements are considered reliable but are not easy to perform and are time-consuming. By contrast, 2-dimensional (2D) photography or radiography, which is used routinely, does not provide safe information on facial dimensions nor allows for reliable comparisons of different time points. Furthermore, the dimensional reduction of a 3-dimensional (3D) object leads to a significant loss of information that limits the imaging value of 2D modalities. Three-dimensional photography seems quite promising but is not yet incorporated in clinical practice.
As a result, clinicians are mostly limited in using 2D (lateral cephalometric images) and 3D (cone-beam computed tomography [CBCT] images) radiographs to assess facial soft tissues at a certain time point. However, cephalometric assessments of soft tissue structures are prone to various sources of error ; and the availability of CBCT images requires that they are justified and have been obtained following the As Low As Diagnostically Acceptable principle. Nevertheless, if available, they depict both hard and soft tissues and also provide stable reference structures for assessing facial changes after treatment or growth. Technological advances in 3D imaging techniques have extended the use of computed tomography and CBCT. , This development has increased the potential of studying the craniofacial complex and its surrounding tissues in a more thorough manner. However, this requires the adaptation of currently used diagnostic analyses into 3 dimensions. For example, the anterior cranial base, which remains stable after an early age, , has been used traditionally for studying changes on cephalometric images but can also be used reliably in 3D superimpositions. ,
Superimpositions of serial 3D images are performed with landmark-based, surface-based, or voxel-based methods, each of which presents its advantages and disadvantages. Landmark-based methods are highly dependent on the number of selected landmarks and can significantly become time-consuming when high accuracy is required. , Surface-based techniques are faster and more user-friendly; however, they require previous segmentation of the bony structures from the entire volume. Therefore, they are subject to the error generated by selecting different threshold values when performing this process. In CBCT images, even if the same threshold is used, there is no guarantee for precise segmentation owing to the absence of correspondence of gray scale values to Hounsfield units.
Voxel-based methods use the best-fit approach to superimpose original volumetric data (voxel gray scale values) and are thus not subject to segmentation error. Nevertheless, they also require bone segmentation to create a surface model that can then be used for a thorough outcome assessment. Several studies have investigated voxel-based superimposition techniques, , some of which reported the technique to be user-friendly and highly reliable. The voxel-based Dolphin 3D (version 2.1.6079.17633; Dolphin Imaging & Management Solutions, Chatsworth, Calif) superimposition technique has been recently tested on a growing population, and it proved suitable for everyday use in clinical practice to evaluate hard tissue treatment outcomes and skeletal changes owing to growth.
Nevertheless, CBCT volumes also include soft tissue data, and software programs allow for the construction of a surface rendering from the original volumetric information. The aim of the present investigation was to assess the reliability and reproducibility of the voxel-based Dolphin 3D superimposition technique to detect soft tissue changes in a growing patient population. The effect of soft tissue surface segmentation on the superimposition outcome was also tested.
Material and methods
This prospective methodological study, using pre-existing patient data, was registered and approved by the Swiss Ethics Committee (protocol no. 2018-01670).
The electronic archives of a single orthodontic clinic, between 2008 and 2018, were reviewed. This clinic does not take CBCTs routinely on all incoming patients; therefore, all CBCTs in the archives were originally obtained in subjects for whom they were considered to be essential for proper diagnosis and treatment planning. All scans were performed with the same x-ray machine (KaVo 3D eXam; KaVo Dental, Hatfield, Pa) under the following settings: scanning area 170 mm × 232 mm; 0.4 mm 3 -voxel size; 5 mA; 120 kV; scan time, 8.9 seconds; exposure time, 3.7 seconds, which allowed for lower dose scans. The data were saved and exported in a Digital Imaging and Communications in Medicine format. The scans of patients with craniofacial syndromes, malformations, or severe facial asymmetries, as well as low-quality scans, were disregarded.
The final sample comprised serial CBCT images of 15 (8 males and 7 females) growing orthodontic patients. This number of scans represents all available scans in the archives that fulfilled the inclusion criteria and was considered adequate for the purpose of the study. Two time points were recorded; T0 (first CBCT volume) and T1 (second CBCT volume). The mean ages of participants were 11.75 ± 0.59 years and 13.44 ± 0.96 years at T0 and T1, respectively. Two researchers (S.T.H, N.G) visually inspected all criteria independently and agreed on them before proceeding to any data generation process.
All pairs of Digital Imaging and Communications in Medicine datasets were imported in Dolphin 3D software. Voxel-based superimpositions of each pair of CBCTs were performed on the anterior cranial base, which is a standard reference to evaluate changes in craniofacial structures over time. , , The first CBCT volume (ie, taken at T0 and mentioned as CBCT T0 thereafter) was repositioned according to the position of the CBCT taken at T1 (named CBCT T1 thereafter). The exact process followed is displayed in Supplementary Figure 1 and has been described previously. Consequently, soft tissue surfaces were extracted from the superimposed volumes through segmentation to evaluate the superimposition outcome. The soft tissue segmentation of the CBCT T0 was conducted with the automated function of Dolphin, using the same threshold for a single dataset. Therefore, the surface generated from CBCT T0 was not affected by the segmentation factor allowing for comparisons between and within operators. By contrast, the segmentation of the CBCT T1 was performed manually, through real-time visual inspection of the resulting surface with different threshold values, until the segmented surface was smooth, with minimum artifacts or holes. All segmented soft tissue surfaces were then extracted into STL format for further analyses. All aforementioned procedures were performed twice, by 2 independent operators (S.T.H, G.K) to test interrater agreement. Furthermore, 1 operator (G.K) repeated the whole superimposition process for 10 sets of 3D models to assess intraoperator agreement. To assess the effect of the segmentation process on the superimposition outcome, the same operator repeated the manual segmentation of 10 surfaces and performed a manual segmentation of 10 surfaces that were previously extracted automatically.
All generated STL files were imported and further processed with Viewbox 4 Software (version 220.127.116.11, BETA 64; dHAL software, Kifisia, Greece), which has been thoroughly tested in surface model processing. , To evaluate the reproducibility of the method, intra- and interoperator agreements were measured on color-coded maps by determining the mean absolute distances (MAD) between corresponding models in each set. The MADs between T0 and T1 surface models were measured on areas with a predetermined size of 100 mesh points, which were selected on the T0 models and transferred to all other models generated from the same set, for consistency reasons. The following 7 areas were selected: N-point, A-point, pogonion, zygomatic arch right and left, and gonial angle right and left ( Supplementary Fig 2 ).
The intraoperator agreement was tested through calculating the MADs between 2 surface models both extracted from CBCT T0 at different time points, by the same operator (operator 1). T0 surface models were chosen because their segmentation was performed with an automated software function, and, thus, there was no manual threshold definition error. Similarly, the MAD between the surface models extracted by 2 operators at T0 was measured to assess interrater agreement. A MAD of 0 mm indicated perfect agreement. The intra- and interrater agreement in measured T0-T1 changes were also tested by comparing the respective MADs between registered models, on the 7 areas previously mentioned. This comparison is influenced by the segmentation error related to the manual T1 model extraction. The magnitude of this error was explored by measuring the MADs between 10 sets of repeatedly extracted surface models. Those were compared with each other, and one of them was also compared with another automatically segmented model. Any deviation between the pairs of extracted models represented the segmentation error.
For consistency reasons, the statistical methodology of the present study was identical to that applied to a previous similar study on hard tissue outcomes. Statistical analysis was performed using the SPSS software (version 25.0; IBM Corp, Armonk, NY).
Raw data were tested for normality through the Shapiro-Wilk test and were not normally distributed in all cases. Thus, nonparametric statistical tests were used. Intra- and interoperator agreement on superimposition outcome was shown with box plots. Any deviation from 0 indicated a superimposition error. Differences in the amount of error among the selected areas were tested in a paired manner through the Friedman test. In the case of significant results, pairwise comparisons were performed through the Wilcoxon signed rank test.
Segmentation error was tested in the same manner as the superimposition error.
In all cases, a 2-sided significance test was carried out at an alpha level of 0.05. In the case of multiple comparisons, a Bonferroni comparison was applied to the level of significance to avoid false-positive results.
The Bland-Altman method (difference plot) was also used to evaluate interoperator agreement in the detected morphologic changes. A 1-sample t test was used to assess if there was a systematic error between the operators.
The intra- and interoperator agreement on superimposition outcomes, assessed through the MAD between the relocated T0 models is shown in Figure 1 . The median error ranged from 0.08 to 0.15 mm and from 0.17 to 0.21 mm for the intra- and interoperator comparisons, respectively. This amount of error was considered clinically acceptable in both cases. In a few cases, it exceeded 0.5 mm, mainly for interoperator comparisons. No significant differences were evident between the error assessed at different areas for both intra- ( P = 0.260) and interoperator ( P = 0.095) comparisons. In both cases, the measurements showed the highest values and the largest variance on the pogonion region, which is located at the furthest distance from the superimposition reference. Visual assessment of the respective color maps confirmed these outcomes ( Fig 2 ). These results are free of any segmentation error, and, thus, they show the pure superimposition process error.