Linear Accuracy and Reliability of Cone Beam CT Derived 3-Dimensional Images Constructed Using an Orthodontic Volumetric Rendering Program

Recently, cone beam computed tomography (CBCT) systems have been developed specifically for the maxillofacial region.1 Many devices are capable of a large field of view imaging of the skull to include most anthropometric landmarks used in cephalometric analysis. Time and dose requirements have been suggested to be of the same order of magnitude as other dental radiographic modalities.2 Recently, maxillofacial applications of CBCT imaging have included craniofacial assessment in orthodontics.3–5 High dimensional accuracy has been reported for maxillofacial CBCT in measurement of facial structures.6–9 

CBCT derived 3D cephalometry has a number of potential advantages over conventional CT for cephalometric imaging including submillimeter resolution and reduced radiation exposure, while still permitting reconstruction of the soft tissue profile. Moreover, CBCT datasets can be imported as DICOM files into personal computer–based software to provide 3D reconstruction of the craniofacial skeleton. These possibilities and increasing access to CBCT imaging for orthodontics are enabling movement from 2D cephalometry to 3D visualization of craniofacial morphology.10 However, there is a lack of data regarding the reliability and accuracy of linear measurements obtained from 3D volumetric renderings of CBCT data constructed from orthodontic software. Studies indicate that 3D reconstructions of conventional fan beam CT datasets have a high degree of accuracy11–16 with differences between measured and actual dimensions being 2–3 mm.17,18 Recently, methodological approaches to cephalometric analysis of 3D CT images have also been described.19,20 We hypothesize that the accuracy of linear measurements obtained from 3D surface renderings of CBCT datasets is similar to the accuracy of measurements obtained from conventional CT datasets using orthodontic specific software.

Hence, the aim of the present study was to compare the reliability and accuracy of linear measurements made on 3D reconstructions generated using Dolphin 3D (Dolphin Imaging, Chatsworth, Calif) software applied to i-CAT (Imaging Sciences International, Hatfield, Pa) CBCT DICOM image datasets to direct measurements made on ex vivo human skulls.

The investigation was an observational cross-sectional ex vivo experiment, approved by the Institutional Human Remains Committee, Department of Anatomical Sciences and Neurobiology at our university.

Twenty-three anonymous dry human skulls with a stable and reproducible occlusion and full permanent dentition were available. Fourteen craniometric anatomic landmarks were identified on each skull using an indelible marker according to currently accepted operational definitions.21 Since it was not possible to locate coordinates in space, we defined Sella (S) to be the midpoint of rim between the anterior clinoid process in the median plane. The dimensions between specific points provided 20 linear distances commonly used in lateral cephalometric orthodontic analysis (Table 1). To minimize the inherent interrater differences in landmark identification and establish a fiducial anatomic location, the distance between each landmark was achieved by consensus of two observers (MM, AB). Measurements were made three times by both observers independently using an electronic digital caliper (27-500-90, GAC, Bohemia, NY). The mean of the measurements was designated as the dimensional truth.

To provide some degree of soft-tissue equivalent attenuation, two latex balloons filled with water were placed in the cranial vault prior to imaging.22 A 1.5-mm thick polystyrene foam wedge was placed in the joint space to separate the glenoid fossa and the condylar head. Prior to imaging, the dentition was placed in maximum intercuspation and the jaws held closed by bilateral metal springs. A custom plastic head holder was constructed to support the skulls during imaging.

Cone beam CT images were acquired using the i-CAT system. The skull was positioned according to manufacturer's instructions. Lateral scout radiographs were made and small adjustments to head position were made so that discrepancies between bilateral structures were less than 5 mm. A single 360° rotation, 20-second scan, comprising 306 basis projections was then made for each skull with a 17.0 cm (diameter) × 13.2 cm (height) field of view using XoranCat acquisition software (version 1.7.7, Xoran Technologies, Ann Arbor, Mich). Exposure parameters were controlled by automatic exposure control.

The CBCT data were exported from the XoranCat software in DICOM multi-file format and imported into Dolphin 3D (Prerelease version 2.3, Dolphin Imaging) on the same computer. All constructions and measurements were performed on a 20.1-inch flat panel color active matrix TFT screen (FlexScan L888, Eizo Nanao Technologies Inc, Cypress, Calif) with a resolution of 1600 × 1200 at 85 Hz and a 0.255 mm dot pitch, operated at 24 bit.

Three-dimensional reconstructions and measurements were undertaken in three stages:

• — 3D surface rendering was generated by manually adjusting the threshold of visible pixel levels. This was the method used for segmentation (Figure 1).

• — The cephalometric landmarks were located and marked on the 3D surface rendered volumetric image. The Dolphin 3D software provided various views using rotation and translation of the rendered image. Landmarks were identified by using a cursor-driven pointer. This was performed by a sequence of preset volumetric orientations (Figure 2). The volume was initially oriented in the posterior-anterior projection and Na, A point, ANS, B point, Gn, Pog and Or bilaterally located (Figure 2A). Next, the volume was oriented in the submentovertex projection, and Me, PNS and Ba were identified (Figure 2B). Then the volume was rotated to the lateral projection and Co, Po and, Go located (Figure 2C and 2D.). Finally the volume was oriented to demonstrate the cranial fossa and S was located (Figure 2E). As the volume rendering was reoriented, the positions of the previously positioned landmarks were verified and, if necessary, relocated.

• — Finally, measurements between specific landmarks were made. For the prerelease software version used, specific points and planes were unnamed. Therefore, it was necessary to select specific points to identify a linear plane. This was performed in a specific sequence such that linear measurement 1 corresponded to Na-A, linear measurement 2 corresponded to Na-B, and so forth. In this way the resulting analysis provided specific linear measurements (Figure 3) which could be exported as text data.

The version of the Dolphin 3D imaging software that was used in this investigation did not allow the user to “save” the location of the identified landmark on the 3D volumetric surface rendering, creating a digitally indelible landmark position. Therefore, for each of the three repeated measurements performed by the first author it was necessary to relocate landmarks and remeasure distances.

A standard statistical software package (SPSS version 12, Chicago, Ill) was used for data analysis. To determine intraobserver reliability, the average measures intraclass correlation coefficient (ICC) was determined for repeat measurements. In addition, absolute and percentage mean error, standard deviations, and 95% confidence levels were calculated. Absolute and percentage linear means of triplicate measurements between specified anatomic landmarks were calculated and standard deviations and 95% confidence levels calculated. The digital measurements of the skulls were taken as a dimensional truth and compared with the 3D CBCT measurements with a 2-tailed paired Student's t-test at an à priori level of significance of P ≤ .05.

The average measure ICC of triplicate skull measurements (0.996 ± 0.007; range: 0.981–0.991) was significantly higher than the mean ICC for 3D CBCT (0.976 ± 0.016; range: 0.941–0.993) (t = 5.468; P < .001) (Table 2).

The mean percentage measurement error for 3D CBCT (2.31% ± 2.11%; range: 1.07% ± 0 .72% to 3.86% ± 1.85%) was significantly higher than repeated measurements to directly determine anatomic skull dimensions (0.63% ± 0.51%; range: 0.29% ± 0.13% to 1.18% ± 0.58%) (t = 16.6, P < .001) (Table 3). For all measurements except Ba-Na, 3D CBCT mean absolute and percentage error was significantly higher than direct skull repeated measurements.

The 3D CBCT data was found to have significant differences to actual skull measurements for 13 of the 20 parameters: seven of the eight midsagittal measurements (all except Na-A) and six of the 12 bilateral measurements (Pog-Go_right, Go-Me, Go-Gn_left, and Po-Or) (Table 4). The greatest mean difference of 3.32 mm (3.56%) was found with Ba-ANS. In eight of 20 measurements (40%) an average difference of less than 1 mm, three of which were significantly different from actual measurements, was found. In six measurements (30%) a mean difference between 1 mm and 1.5 mm was found, and for five skulls the CBCT measurements were significantly different from actual measurements made directly on the anatomic specimens. In four measurements (20%) an average difference of 1.5 mm to 2 mm was found, all of which were significantly different from the assumed anatomic truth. Finally, for two measurements (10%) a mean difference greater than 2 mm was found, one of which was significantly different from the mean direct skull measurements. In all comparisons, except Na-B and Go-Gn bilaterally, 3D CBCT measurements were less than skull measurements. The mean percentage difference between the mean skull and 3D-based linear measurements was −1.13% (SD ± 1.47%) and ranged from −0.27% for Na-A to −3.44% for ANS-PNS.

The introduction of maxillofacial CBCT equipment provides clinicians with an opportunity to generate 3D volumetric renderings easily using relatively inexpensive third-party personal computer–based software. The rapidly emerging availability of this technology will undoubtedly expand the use and application of 3D imaging, particularly in the field of orthodontics.23–26 The aim of this study was to compare the reliability and accuracy of linear measurements made on 3D volumetric reconstructions generated from CBCT datasets using a proprietary orthodontic image and analysis program to direct measurements made on ex vivo human skulls.

It was found that for two-thirds of dimensions, CBCT measurements were statistically significantly different from actual measurements. However, analyzing the absolute and percentage differences, this statistical significance probably does not translate into clinical relevance. Statistical differences most likely resulted from small standard deviations within the measurements. In addition, the greater intraobserver variability demonstrated by the 3D CBCT measurements also may have so contributed.

These results are very similar in magnitude to those of Cavalcanti et al.15 However, they reported no statistically significant difference between imaging and physical measurements. Their mean difference between actual and 3D-based linear measurements was 0.83% as compared with −1.13 ± 1.47% in the present study. In the present study, 40% of measurements had an average difference of less than 1 mm, 70% had an average difference of less than 1.5 mm, and 90% had an average difference of less than 2 mm. These absolute differences compare favorably with those reported by additional authors.16,18 

There are numerous factors which should be considered when applying the results of this investigation to clinical situations. First and foremost, this study was performed on human skulls. The accuracy of measurement distances between three-dimensional landmarks on 3D volumes of patients may be affected by a reduction in image quality due to soft-tissue attenuation, metallic artifacts, and patient motion. Variation in scanning protocol such as voxel size and number of basis projection images may also influence measurement accuracy. Therefore, it would be expected that the dimensional accuracy of 3D measurements would be somewhat less on patient derived data.

There are also some potential limitations when using 3D images derived from CBCT data.³⁹ Three-dimensional volumetric depiction depends on appropriate segmentation—the thresholding of bone pixel values and suppression of surrounding tissue values to enhance the structure of interest. This process is dependent on the software algorithm, the spatial and contrast resolution of the scan, the thickness and degree of calcification or cortication of the bony structure, and the technical skill of the operator. In this study, the Dolphin 3D software provides a semimanual method of segmentation, dependent on the interaction of the operator with the data to produce a visually acceptable 3D rendering. These limitations result in deficiencies or voids in the surface of the image. These occur in regions that are represented by few voxels or have gray values still representing bone, but outside the threshold. These areas include the posterior and anterior superior walls of the maxillary sinus, bone overlying the dentition and cortical bone of the mandibular condyle. Consequently, this may lead to greater landmark identification error and subsequent measurement error. Anatomic landmarks whose accuracy may be affected by poor segmentation include A point, ANS, PNS, porion, and condylion.

In addition, the method of establishing dimensional truth could have potentially contributed to bias in the results. While the landmark identification and measurements on the 3D rendered images were repeated three times by a single observer, the landmark identification on the skulls was performed only once and measurements performed three times by two observers independently. This reduced the error of point identification on the skulls. However, the establishment of a consensus landmark location was necessary to provide a fiducial reference to which we could assess the inherent clinical inaccuracies of both landmark identification and measurement associated with the 3D image rendering.

We were unable to completely simulate soft tissue effects of attenuation on image quality in this study as this would have been problematic in positioning and orientation of the skulls. While the use of water balloons placed within the cranial cavity provided some degree of soft-tissue attenuation, the lack of peripheral attenuation material may have allowed easier identification of landmarks on 3D surface rendered images.

In this study it was also found that mean intrarater error was significantly greater for almost all 3D CBCT measurements compared to actual skull measurements. This was most likely due to the differences in which landmarks were identified between the skull and volumetric image. Previous authors have used metallic markers as fiducial points on skulls to identify landmarks with subsequent images providing radio-opaque landmarks. This methodology was not followed for the present investigation because of the desire to establish a control linear dimension on the skulls.

• While many linear measurements between cephalometric landmarks on 3D volumetric surface renderings obtained using Dolphin 3D software generated from i-CAT CBCT datasets may be statistically significantly different from anatomic dimensions, most can be considered to be sufficiently clinically accurate (−1.13% ± 1.47%) for craniofacial analyses.

The authors wish to thank Ms April Brown, 2005 summer research associate and graduate student candidate for the Masters in Oral Biology at The University of Louisville School of Dentistry, for her valuable assistance in this project.