The impact of CBCT-head tilting on 3D condylar segmentation reproducibility

Abstract

Objectives:

To investigate whether variations in head positioning may influence the reproducibility of cone-beam CT (CBCT) three-dimensional (3D) segmented models of the mandibular condyle.

Methods:

Five fresh frozen cadaver heads were scanned in four different positions: reference position (RP) and a set of three tilted alternative head positions (AP) in anteroposterior direction (AP1: 2 cm anterior translation, AP2: 5° pitch rotation, AP3: 10° pitch rotation). Surface models of mandibular condyles were constructed and compared with the condylar reference position using voxel-based registration. Descriptive statistics and a linear mixed-effects model were performed to compare condylar volumetric differences and root mean square (RMS) distance between surfaces of AP vs RP.

Results:

The mean differences in condylar volumes of AP vs RP were 14.1 mm³ (95% CI [−79.3, 107.4]) for AP1, 1.0 mm³ (95% CI [−87.2, 89.2]) for AP2 and 0.1 mm³ (95% CI [−88.3, 88.4]) for AP3. Mean and absolute volumetric differences did not exceed earlier reported intraoperator differences of 30 mm³. The RMS distance values obtained per group were 0.12 mm (95% CI [0.05,0.20]) for AP1, 0.17 mm (95% CI [0.10, 0.22]) for AP2 and 0.17 mm for AP3 (95% CI [0.10,0.22]). The confidence intervals (CI) for RMS distance remained far below the threshold for clinical acceptability (0.5 mm).

Conclusions:

Within the limits of the present study, it is suggested that tilted head positions may affect the reproducibility of 3D condylar segmentation, thereby influencing outcome in repeated CBCT scanning. Nevertheless, observed differences are unlikely to have a meaningful impact on clinical patient diagnosis and management.

Introduction

Cone-beam CT (CBCT) has become an increasingly used imaging modality for three-dimensional (3D) assessment of the craniomaxillofacial (CMF) complex and dentoalveolar structures. ^1,2 A lower radiation dose, an increased spatial resolution, a seated patient position and lower machine-investment cost, compared to multislice CT, has propelled its application in dentistry and oral & maxillofacial surgery forward. Next to assessing anatomical and pathological information on two-dimensional (2D) reconstructions of the volumetric data, segmentation techniques can render 3D surface models of different CMF structures. These surface models can be used for 3D diagnostics and printing, virtual surgical planning and design of patient specific implants. ^3–5 Virtual anatomical models also provide possibilities for novel methods of patient follow-up such as evaluation of morphological changes of the maxillofacial complex following CMF surgery. Objective evaluation of the treatment results allows for a feedback loop in treatment planning and eventually improves patient care. ⁶

To allow for this high-quality feedback loop, accurate 3D models of the CMF complex are essential. As described by Stamatakis et al, three main categories are known to influence segmentation accuracy: scanner-related (scanner type, field of view size, voxel size settings), operator-related (software employed, segmentation parameters, operator performing the segmentation) and patient-related (patient positioning, patient movement, metal artifacts, covering soft tissue) variables. ⁷ Several studies investigated the influence of scanner- or operator-related factors on CBCT imaging. ^8–16 Studies examining the influence of patient-related factors on CBCT imaging seem less prevalent. However, patient-related factors are also important during scanning, as they may attribute to variability in image quality, radiation dose, 3D surface rendering and anatomical segmentation. The few studies investigating patient-related factors often use dry human skulls where soft tissue is absent. ^17–19 Soft tissues may induce scattered radiation, thereby affecting anatomical segmentation, leading to deviations from clinical reality. ¹⁶

In comparison to CBCT, multislice CT (MSCT) offers a better contrast-to-noise ratio, hence a higher contrast resolution, higher tube voltage and stable Hounsfield units, resulting in better segmentation results. ^20–22 However, the adequate contrast resolution obtained using lower radiation doses in addition to its compact chairside and low-cost equipment, makes CBCT the modality of choice for long-term follow-up of oral and maxillofacial surgery patients. ²⁰ The mandibular condyle remains an area sensitive to segmentation errors due to the proximity of the glenoid fossa, connected soft tissues such as the articular disc, beam attenuation by the petrous part of temporal bone and its low inherent bone density. ²³ If image distortion due to patient positioning would occur, one could expect it to be noticeable in this region. ²⁴ Knowing the importance of image distortion is essential when one starts interpreting volumetric differences over time of mandibular condyles due to condylar remodeling in orthognathic surgery. The aim of this study was to assess if tilted head positions in a CBCT unit may affect the resulting 3D segmented models of the mandibular condyle.

Methods and materials

Data acquisition and CBCT imaging

For this study, five fresh frozen cadaver heads derived from a Caucasian elderly population were obtained. Each head was fixated in expanded polystyrene (EPS) boxes with isolation foam, making any non-intentional movement impossible. A custom-made plexiglass platform was manufactured to fixate the EPS boxes. Tripod positioning was standardized using markers on the floor. This construction was stable and provided a standardized and reproducible positioning workflow (Figure 1). The platform allowed detailed and measurable rotation and translation movements. The scans were performed by the Newtom VGi evo CBCT machine (QR Verona, Verona, Italy) following the same imaging protocol for orthognathic patients (FOV 24 × 19 cm, voxel size 0.3 mm³, 110 kVp, 4.3 mA, rotation time of 16 s and effective radiation time of 9,5 s). ²⁵ Each head was scanned in four positions, which resulted in a total of 20 scans. A reference position (RP) at 0° pitch of the Frankfort horizontal (FH) plane towards the floor was set as a reference. The Frankfort horizontal plane was estimated visually to be as parallel as possible to the floor using the horizontal reference lines. A set of alternative head positions (AP) were simulated, starting from the RP: 20 mm anterior translation (AP1), 5° clockwise (CW) pitch rotation (AP2) and 10° CW pitch rotation (AP3) (Figure 2). For the translated positioning (AP1), the platform was moved 2 cm anteriorly, while maintaining the Frankfort horizontal plane. An electronic spirit level was used to perform the pitch rotations (AP2 and AP3). The cadaver heads were wrapped in a plastic bag for hygienic reasons. The orbits and ears were visible and palpable, which made it possible to estimate the positioning in the Frankfort horizontal plane.

Figure 1.

Setup of the scanning procedure. (a) Anterior view. (b) Sagittal view.

Figure 2.

Four pre-determined head positions. (a) Reference position, RP. (b) 2 cm translated position, AP1. (c) 5° pitch rotation, AP2. (d) 10° pitch rotation, AP3. FH, Frankfort horizontal plane; cm, centimeter, °, degrees; RP, reference position.

Segmentation and registration

The digital imaging and communication in medicine (DICOM) files from each scan were exported and imported into an artificial intelligence (AI)-driven segmentation platform (Virtual Patient Creator, Relu BV, Leuven, Belgium) (Figure 3). ²⁶ The AI tool is based on a deep learning algorithm by means of a convolutional neural network using a 3D UNet architecture. First, a rough segmentation of the mandible is performed using a downsampled version of the input image. This rough segmentation is subsequently used to create patches of the input image considering only the regions where the rough AI detected mandible voxels. These patches are fed to a second convolutional neural network based on the same architecture as the first model that provides a much finer segmentation. Finally, all segmented patches are combined to obtain the full segmentation. Manual corrections for over- and undersegmentations are possible, but were not made since the proven accuracy of the validated AI-driven tool. Using an AI-based automatic approach limits the operator-related variability and delivers accurate 3D segmentations. For each head, this approach resulted in four Surface Tessellation Language (STL) files (ASCII format), corresponding to RP, AP1, AP2 and AP3. The segmentations were performed by one Oral and Maxillofacial surgeon with expertise using this software and analyzing these structures. Intra- and interexaminer reliability of the segmentations were not assessed since the automated AI-driven segmentation process. This tool has been validated and is accurate when benchmarked to semi-automatic segmentation, which is the current clinical standard. ²⁶

Figure 3.

Segmentation process: automated segmentation was performed using an AI-driven segmentation platform (Virtual Patient Creator, Relu BV, Leuven, Belgium). An initial segmentation of the mandible was obtained by the AI tool, marking the voxels corresponding to the mandible in red. Then, manual corrections were made in the coronal (a), sagittal (b) and axial (c) plane. Finally, a full segmentation of the mandible was obtained (d). AI, artificial intelligence.

To be able to compare the 3D condyles between the four scanning positions, the AP scans were registered on the RP scan using voxel-based rigid registration in Amira software (v. 6.7.0, Thermo Fischer Scientific, Merignac, France) via a stepwise wizard which allowed to transpose pre-constructed 3D models according to the transformation matrix derived from the registration process. The region for registration was the complete mandible excluding both the condyles (i.e. the region of interest for analysis) cut-off by the operator above the lowest point of the incisural notch.

The transformation matrix of the voxel-based rigid registration procedure was subsequently applied to the STL’s created out of AP’ scans (Figure 4).

Figure 4.

Registration process: voxel-based registration was performed using Amira software. (a) The scans in reference position (gray) and in alternative position (green; 2 cm translation scan) were imported in the software. (b) The mandible excluding both condyles was used as reference for registration. (c) Voxel-based registration. (d) This panel shows the initial position of the 2 cm translation scan (red) and the transformed position of the STL based on the VBR transformation matrix (blue). STL, Surface Tessellation Language; VBR, voxel-based rigid registration.

Surface model analysis

For each head, the RP STL and the three registered AP STL’s were imported in 3-matic software (v. 13.0, Materialise, Leuven, Belgium) for volumetric analysis and closest point surface analysis (Figure 5). Condyles were cut from the mandible using a shared analytical sphere. The root mean square (RMS) distance was calculated to measure the match between the condylar surfaces of the AP models and the RP model. Volumes of the cut condyles were also calculated. Volume differences of 30 mm³ or less were regarded as acceptable, as these are typical for intraoperator segmentation errors of the condyle. ²⁷ For surface distance mapping, an RMS distance value of up to 0.5 mm is acceptable in clinical conditions. ²⁸

Figure 5.

Condylar volume analysis and surface distance mapping using the Materialise 3-matic software. (a) Set of four superimposed condyles, anterior view. (b) Color map illustrating the surface distance mapping of two superimposed condylar models, posterior view.

Statistical analysis

For statistical analysis of the results, the software package RStudio for Windows, v. 4.1.2 (Boston, M: RStudio) was used. Statistical analysis and interpretation were supported by a professional statistician. Descriptive statistics were performed to calculate the mean and standard deviation (SD) of condylar volumes for each head position. This was also done for the absolute volumetric differences. A linear mixed-effects model was performed to calculate the volumetric differences between the RP and the AP’s and their 95% confidence interval (CI) considering experimental positions and condylar side (left condyle (L), right condyle (R)). The RMS distance values between the RP and the AP’s were also assessed using descriptive statistics as well as a linear mixed-effects model to calculate the mean RMS distance and its 95% CI taking into account experimental position and condylar side.

Results

Table 1 shows the condylar volumes of the heads, sides and experimental positions as well as the mean volume of each experimental position and its standard deviation. A linear mixed-effects model was conducted to assess the differences in condylar volumes between the different head positions considering the possible effect of side and its interaction with the experimental positions. Since the assumption of homogeneity of variance was violated due to a different variability of condylar volume for each experimental position, we attributed weights to the different positions in the linear mixed model. Different models were assessed for a best fit using the Akaike information criterion (AIC). ²⁹ The best fit was found for the full model with position and side as individual factors as well as the interaction of the independent variables. However, no significant effect for the interaction of position and side on condylar volume was found using a likelihood ratio test (p = 0.97). This allowed for an evaluation of the difference in condylar volume between the different positions irrespective of condylar side. The mean volumetric difference of the translated AP vs RP was the largest, measuring 14.1 mm³. The pitch AP’s showed lower mean volumetric differences of 1.0 mm³ for AP2 and 0.1 mm³ for AP3. The confidence intervals obtained using a lined mixed-effects model were [−79.3, 107.4] for AP1, [−87.2, 89.2] for AP2 and [−88.3, 88.4] for AP3. On the other hand, the absolute volumetric difference of the pitch rotation groups was the largest, measuring 27.3 mm³ for AP2 and 22.1 for AP3. The translated AP had a lower absolute volumetric difference of 14.8 mm³. The average condylar volume in this study sample was 1801.8 mm³.

Table 1.

Condylar volumes assessed for five different heads comparing a reference position with three tilted alternative head positions (Figure 2)

Condylar volumes (mm³)
Head	Side	RP	AP
			AP1				AP2				AP3
				RP-AP1		Abs Δ		RP-AP2		Abs Δ		RP-AP3		Abs Δ
H1	L	1649.59	1656.56	−6.97		6.97	1647.82	1.77		1.77	1655.02	−5.43		5.43
	R	1427.94	1443.49	−15.55		15.55	1460.56	−32.62		32.62	1450.47	−22.53		22.53
H2	L	1887.82	1886.48	1.34		1.34	1902.17	−14.35		14.35	1903.63	−15.81		15.81
	R	1771.98	1769.59	2.39		2.39	1785.33	−13.35		13.35	1765.58	6.4		6.4
H3	L	1738.84	1751.58	−12.74		12.74	1735.05	3.79		3.79	1761.22	−22.38		22.38
	R	1823.67	1830.47	−6.8		6.8	1840.6	−16.93		16.93	1836	−12.33		12.33
H4	L	1638.15	1656.9	−18.75		18.75	1567.25	70.9		70.9	1569.78	68.37		68.37
	R	1585.87	1601.78	−15.91		15.91	1530.65	55.22		55.22	1550.36	35.51		35.51
H5	L	2228.09	2261.23	−33.14		33.14	2240.22	−12.13		12.13	2242.54	−14.45		14.45
	R	2227.95	2262.48	−34.53		34.53	2280.26	−52.31		52.31	2245.81	−17.86		17.86
Mean		1797.99	1812.06	−14.07		14.81	1798.99	−1.00		27.34	1798.04	−0.05		22.11
(±SD)		(±260.82)	(±267.56)	(±12.58)		(±11.58)	(±280.09)	(±37.57)		(±24.13)	(±272.39)	(±29.71)		(±18.43)
[95% CI]			[−79.3, 107.4]					[−87.2, 89.2]			[−88.3, 88.4]

H: cadaver head; L: left condyle; R: right condyle; RP: reference position; AP1: 2 cm anterior translation; AP2: 5° clockwise pitch rotation; AP3: 10° clockwise pitch rotation; Abs Δ: absolute difference; SD: standard deviation; CI: confidence interval.

Table 2 shows the individual RMS distance values, each time comparing the surface of the AP to the RP. A linear mixed-effects model was performed to calculate the mean RMS distance and its 95% CI considering experimental position and condylar side. Since the assumption of homogeneity of variance was violated due to a different variability of RMS distance for each experimental position, we attributed weights to the different positions in the linear mixed model. Different models were assessed for a best fit using the AIC and the best fit was found for the model not taking any factor into account. This already hinted towards the conclusion that nor head position, nor condylar side played a significant role in the RMS distances. Using this model, the average RMS distances and their confidence intervals were calculated for each AP-RP comparison. None of the average RMS distances nor their confidence intervals exceeded the clinical standard of 0.5 mm.

Table 2.

RMS distance values: measurement of the accuracy between the condylar surfaces of the AP vs RP

RMS distance (mm)
Head	Side	AP1	AP2	AP3
H1	L	0.14	0.12	0.13
	R	0.12	0.14	0.14
H2	L	0.12	0.11	0.12
	R	0.13	0.13	0.11
H3	L	0.1	0.12	0.12
	R	0.1	0.12	0.09
H4	L	0.13	0.43	0.36
	R	0.12	0.27	0.29
H5	L	0.14	0.12	0.15
	R	0.13	0.14	0.14
Mean (±SD)		0.12 (±0.01)	0.17 (±0.10)	0.17 (±0.09)
[95% CI]		[0.05, 0.20]	[0.10, 0.22]	[0.10, 0.22]

AP1: 2 cm anterior translation; AP2: 5° clockwise pitch rotation; AP3: 10° clockwise pitch rotation; CI: confidence interval;H: cadaver head; L: left condyle; R: right condyle;RMS, root mean square; RP: reference position; SD: standard deviation; Δ: difference.

Discussion

Patient head positioning has been suggested as an influential factor of CBCT image quality and subsequent accuracy of maxillofacial 3D segmented model generation. Since these 3D models are increasingly used for diagnostics, virtual surgical planning, or follow-up, assessing the influence of variable head positions and tilting is necessary. In this study, we focused on 3D models of mandibular condyles. The condylar region especially is susceptible to image distortion due to its location near the highly dense petrous part of the temporal bone, causing beam attenuation and its low inherent bone density. If head positioning has an impact on the accuracy of surface model generation, this should be noticeable in the condylar region.

When comparing the obtained RMS distance values, the upper limit of normal of the AP did not exceed 0.2 mm, which is significantly smaller than the previously set limit of 0.5 mm. ²⁸ By calculating the confidence intervals in this study, an attempt was made to extrapolate these findings to a population level beyond this study sample. The upper border of the confidence levels for RMS distance stayed below the clinical standard of 0.5 mm. As to the volumetric differences, we compared both mean and absolute volumetric differences. The mean volumetric difference was largest in the translation group (AP1). On the other hand, the absolute volumetric difference was smallest in the latter and greater in the pitch rotation groups (AP2 and AP3). This slight discrepancy between mean and absolute volumetric differences is an interesting observation. However, both the mean and absolute differences did not exceed the known intraoperator errors of 30 mm³ and can therefore be considered as clinically negligable. ²⁷ When analyzing the confidence intervals, the highest limit was observed in the AP1 group and measured 107.4 mm³. When taking the average condylar volume in this study sample of 1801.8 mm³ into consideration, this means a maximum of 6% condylar volume difference could be found between AP’s and the scanning position of reference, based on this small study sample. This number remains very low and could be considered as clinically irrelevant. Overall, the results in this study suggest that the simulated variations in head positioning in a CBCT may influence the reproducibility of 3D segmented models of the mandibular condyle, but on a clinically non-relevant level.

Only one previous study was found investigating the effect of patient positioning in a CBCT unit on the accuracy of segmented 3D models of the maxillofacial complex. ⁷ Their results suggest head positioning does affect the accuracy of segmented 3D models of the maxillofacial complex on statistically significant level, however not to a clinically relevant extent. Our findings, although based on a small experimental sample, seem to concur with the conclusion of Stamatakis et al. ⁷ There are differences between 3D models of different scanning positions, but these are minor and could also be regarded as clinically insignificant.

Previous studies have examined head position and the effect on CBCT image quality, organ radiation dose and effective dose ranges. ^{19,25,30–34} Recent articles have reported head positioning to significantly influence the image quality of CBCT, but few have studied the effect on creating of 3D segmented models from these data. Contrast-to-noise ratio (CNR) was used in these studies as an objective measure of image quality. Lindfors et al suggested that patient positioning in a CBCT unit influences the image quality in the mandibular region, as hyperextension of the head rendered higher CNRs regardless of the CBCT device or field of view (FOV) used, indicating an improved image quality. ³² These results were attributed to the amount of tissue volume in the cone-shaped radiation field affecting the uniformity of the density values. Koivisto et al suggested that the highest CNRs were obtained in the prone imaging position, resulting in the best image quality of the maxillofacial region. According to the authors, this is due to the closer proximity of the head to the scanner sensor in the prone position. ³⁰ The differences in image quality were also determined between different anatomical structures in the maxillofacial area. Prone and supine positions resulted in good image quality from the mandible, whereas the image quality around the ear was only assessed as adequate. The lower part of the mandible is separated from the dense area of the skull base and is therefore less prone to the influence of superimposition of surrounding anatomical structures. Moreover, the mandibular corpus has a thick cortication offering better contrast resolution and therefore easier segmentation as compared to the maxilla. Mandibular condyles, however, are smaller and less corticated, or even fully lacking in case of condylar resorption after orthognathic surgery, hampering segmentation results. ²³

Besides patient-related factors, scanner-related variables are also of importance when assessing image quality from CBCT imaging. ⁷ In general, CBCT offers a lower radiation dose compared to MSCT for orthognathic patients, except for temporal bone procedures. ²⁵ This illustrates the importance of kilovoltage (kV) for penetration in thick bone structures. MSCT or CBCT of 120 kV is required to facilitate penetration in dense bone structures such as the petrous part of the temporal bone, which is in close proximity to the mandibular condyle. Other scanner-related variables such as the size and position of the FOV also influence radiation doses. ³⁵ When focusing on the condylar area, the horizontal diameter of the FOV should be larger than 15 cm to avoid truncation artifacts due to exclusion of the condyles from the FOV. ³⁶ Therefore, the condyles must be completely included in the FOV to ensure segmentation accuracy.

A first limitation in this study is the relatively small study sample. This is due to the limited availability of cadaver heads. Secondly, only a limited amount of different head positions were simulated in this study. The positioning platform only allowed translation and pitch rotation in a reproducible way. These positions were further supported by our clinical experience on patient positioning errors. However, Shokri et al reported horizontal head positioning to affect cephalometric measurements more than tilted or pitched head positions. Moreover, they observed that horizontal rotations affected the measurements of distances made on CBCT posteroanterior (PA) cephalograms more than on conventional PA cephalograms. Therefore, special attention for horizontal head rotations in a CBCT unit should be considered and tested in the future. ³⁷ Secondly, the current study was carried out on a single CBCT unit. Results cannot be extrapolated to other CBCT units and varying exposure protocols, as these may lead to different image quality outcome. ^7,32 Finally, the sample only consisted out of five cadaver specimens. Future studies on this topic should consider the use of additional fresh frozen cadaver heads, a larger set of head positions including horizontal head positions, the application of different exposure protocols and FOVs, the use of different CBCT units to rule out scanner-related interference and the inclusion of different patient categories such as condylar remodeling, condylar resorption or rheumatoid arthritis.

Conclusion

The results in this study seem to suggest that, within the limits of this study, slight alterations in patient positioning during CBCT scanning do affect the reproducibility of the resulting 3D segmentation of condylar models. Nevertheless, observed differences are unlikely to have a meaningful impact on clinical patient diagnosis and management.