Reliability and accuracy of a semi-automatic segmentation protocol of the nasal cavity using cone beam computed tomography in patients with sleep apnea

The objectives of this study included using the cone beam computed tomography (CBCT) technology to assess: (1) intra- and inter-observer reliability of the volume measurement of the nasal cavity; (2) the accuracy of the segmentation protocol for evaluation of the nasal cavity. This study used test–retest reliability and accuracy methods within two different population sample groups, from Eastern Asia and North America. Thirty obstructive sleep apnea (OSA) patients were randomly selected from administrative and research oral health data archived at two dental faculties in China and Canada. To assess the reliability of the protocol, two observers performed nasal cavity volume measurement twice with a 10-day interval, using Amira software (v4.1, Visage Imaging Inc., Carlsbad, CA). The accuracy study used a computerized tomography (CT) scan of an OSA patient, who was not included in the study sample, to fabricate an anthropomorphic phantom of the nasal cavity volume with known dimensions (18.9 ml, gold standard). This phantom was scanned using one NewTom 5G (QR systems, Verona, Italy) CBCT scanner. The nasal cavity was segmented based on CBCT images and converted into standard tessellation language (STL) models. The volume of the nasal cavity was measured on the acquired STL models (18.99 ± 0.066 ml). The intra-observer and inter-observer intraclass correlation coefficients for the volume measurement of the nasal cavity were 0.980–0.997 and 0.948–0.992 consecutively. The nasal cavity volume measurement was overestimated by 1.1%-3.1%, compared to the gold standard. The semi-automatic segmentation protocol of the nasal cavity in patients with sleep apnea and by using cone beam computed tomography is reliable and accurate. This study provides a reliable and accurate protocol for segmentation of nasal cavity, which will facilitate the clinician to analyze the images within nasoethmoidal region.


Introduction
The nose is an important human organ with multiple functions, including olfaction, filtering, humidifying inhaled air [1].Analysis of nasal cavity anatomy can help detect certain disorders, such as obstructive sleep apnea (OSA) [2,3].Evidence from population studies shows the role of nasal obstruction in the pathophysiology of OSA [4,5] Often, the treatment plan for patients who suffer from OSA and have chronic nasal obstruction could be nasal surgery, such as septoplasty [6,7].Thus, it is recommended that patients with OSA undergo further nasal analysis as an aid in diagnostics and in treatment decision making [8].
The anatomy of the nasal cavity is very complex [9] and varies among different races and ethnic groups [10].To the best of our knowledge, there are a limited number of studies using 3D analysis to analyze the anatomy of the nasal cavity [2,10,11].Furthermore, these studies mainly used computerized tomography (CT) images, which expose patients to a high radiation dose.More recently, cone beam computed tomography (CBCT) has been used as an alternative tool in the diagnostics of diseases and anatomic variations with a lower radiation exposure risk for patients [12][13][14].There are Extended author information available on the last page of the article several studies on 3D analysis of upper airway using CBCT [15][16][17].However, studies using CBCT for the 3D analysis of the nasal cavity in OSA patients of various ethnicities, are still very limited [18][19][20][21].In this study, we develop a semiautomatic segmentation protocol for 3D analysis of the nasal cavity in OSA patients.It will be interesting to investigate whether our protocol for 3D analysis of the nasal cavity is reliable for different ethnic groups.
The accuracy of the 3D analysis of the nasal cavity is another important issue.To test accuracy, it is necessary to create a phantom of the nasal cavity to serve as a "gold standard".Recent developments in the field of three-dimensional (3D) printing offer new opportunities for manufacturing lifelike anthropomorphic phantoms [22], allowing us to use them as a gold standard.
Therefore, this study aimed to test the reliability and accuracy of a semi-automatic segmentation protocol for the 3D analysis of the nasal cavity.

Data source
Date source consisted of CBCT images of patients with diagnosed sleep apnea from two different countries.This included CBCT images of 15 Chinese and 15 Canadian patients (age: 57.8 ± 18.3) diagnosed with OSA by polysomnography (PSG), which were randomly selected from scans available at the Department of Oral Radiology, Shandong University, China and the Department of Oral Radiology, University of Montreal, Canada.These images were obtained during previous studies [23] and IRB approvals were obtained.
The inclusion criteria were as follows: age > 18 years and availability of CBCT images covering the entire nasal cavity from the nasion point to the level of the hard palate.The exclusion criteria were presence of a palatal cleft, presence of a craniofacial syndrome, presence of space occupying lesions in the nasal cavity and craniofacial surgery in the past.
The CBCT images were obtained using NewTom 5G (QR systems, Verona, Italy) machines in both oral radiology departments, according to the standard imaging protocol.During the CBCT, patients were positioned in the supine position, with the Frankfort horizontal (FH) plane perpendicular to the floor.They were instructed to maintain maximum intercuspation and to avoid swallowing and other movements during the scanning period.The exposure settings were 110 kV, 4 mA, 18*16 cm field of view, 0.3 mm voxel size, 3.6 s exposure time (pulsed radiation), 18 to 36 s scanning time, depending on the size of the patient.For further analysis, the images were saved as Digital Imaging and Communications in Medicine (DICOM) files.To get a standardized head position of each CBCT image, re-orientation was performed by adjusting the palatal plane (the plane crossing anterior nasal spine (ANS)-posterior nasal spine (PNS)) being parallel to the global horizontal plane in the sagittal view, and perpendicular to the global horizontal plane in the axial view [24,25].These data sets were imported into Amira software (v4.1, Visage Imaging Inc., Carlsbad, CA) for nasal cavity measurements.

Measurement and semi-automated nasal cavity segmentation protocol
Using five CBCT imaging data sets not related to this study, two orthodontists were calibrated and trained as observers.After calibration, each observer performed the nasal cavity measurements on all images twice within a 10-day interval.The observers were blinded to patients' information during the measurements.
The Amira software was used for the semi-automatic segmentation of nasal cavity as follows: first, a voxel set was built to include all nasal cavity information.second, a new mask was built with its thresholds ranging from -1000 to -300.third, the boundaries of the nasal cavity were selected in the corresponding axial plans and put into the voxel set.The boundaries were set as follows (Fig. 1): the upper boundary (i.e., the plane across the anterior point of the sphenoid parallel to the FH plane); the lower boundary (i.e., the last axial slice to show the nasal cavity); the left and right boundaries (i.e., the sagittal plane across the medial wall of the left and right maxillary sinus); the anterior boundary (i.e., the nostril not connected with the outside); the posterior boundary (i.e., the coronal plane across posterior nasal spine point).Fourth, all of the slices within the above boundaries were selected and put into the voxel set.Finally, all slices were checked to make sure no area was missing and, if necessary, modified properly using the tools integrated in the software.The software then calculated the total volume of the nasal cavity.

Accuracy of nasal cavity measurement
A CT data set (Discovery CT 750, General Electric Healthcare, Milwaukee, USA) of a 40-year-old Chinese male previously was used to design an anthropomorphic phantom of the nasal cavity.The aforementioned CT data set was converted into a virtual 3D surface, i.e., a standard tessellation language (STL) model of the nasal cavity.The STL model of the nasal cavity was subsequently used to build the phantom.It was then printed using polylactic acid (PLA) and a 3D printer (Aurora, Shenzhen, China) (Fig. 2a).Then an 1 3 impression of the PLA-phantom was acquired using liquid silicone (Wacker, Germany).After taking the PLA-phantom out of the silicone, the phantom of the nasal cavity was obtained (Fig. 2b).
The next step was to perform a CBCT scan (NewTom 5G, Verona, Italy) on the silicone phantom.The acquired CBCT data set was saved as DICOM files and was imported into Amira® software (v4.1, Visage Imaging Inc., Carlsbad, California, USA) (Fig. 2c).The segmentation procedure on the DICOM data sets of the nasal cavity was performed five times by the two calibrated orthodontists and was subsequently repeated after a 10-day interval following the procedure as follows: First, a voxel set was built to include all of the nasal cavity information; second, a new mask was built with its thresholds ranging from -1000 to -300; finally, the entire slice was scrolled to check if any area was missing, and corrections were applied, if needed, using the tools integrated in the software.This resulted in a total of 20 values for nasal cavity volume (experimental nasal cavity measurements).The volume of the PLA-phantom was used as our gold standard value.

Sample size
The power calculation recommended by Walter et al. for reliability studies was followed [26].The null hypothesis was defined as H0: ρ0 ≦ 0.6, and the alternative hypothesis was defined as H1: ρ1≧0.8.The rate of type I error (a), which equates to the criterion for significance, was set at 0.05.The rate of type II error (b), which is related to the power of a test (1-b), was set at 0.2.After reviewing Table II in Walter et al.'s study, the proposed sample size was set at 15 patients [27].The accuracy of the protocol was calculated as the ratio of the experimental nasal cavity measurements to the gold standard in percent.To evaluate the accuracy of the protocol, a one-sample t-test was used to test the difference between the experimental nasal cavity measurements and the gold standard value.The measurement error (%) was calculated as the difference between the experimental nasal cavity measurements and the gold standard value.Data were analyzed using the Statistical Package for Social Sciences for Windows (version 21, SPSS Inc., Chicago, IL).Statistical significance was established at α = 0.05.

Reliability
Table 1 presents the nasal cavity measurements and the ICCs of intra-and inter-observer reliability.Both intra-and interobserver reliability of the nasal cavity measurements were excellent (ICC = 0.94-1.00).

The volume measurement
The 3D printed anthropomorphic phantom and the CBCT images of this phantom are shown in Fig. 2. A statistically significant difference (p < 0.05) was found between the gold standard (18.9 ml) and the measurement of the nasal cavity (Mean ± SD: 18.99 ± 0.066 ml).The measurements of the volume of the nasal cavity were larger than those of the gold standard.The measurement error in volume was 1.1%-3.1%.Bland-Altman plots showed the agreement between observers during the accuracy measurement (Fig. 3).

Discussion
In our study, both intra-observer reliability and interobserver reliability were excellent for the nasal cavity measurement on CBCT images of OSA patients.The nasal cavity volume measurement methodology used in this study can be applied in future research in a reliable way.
A phantom of a nasal cavity was used in this study to test the accuracy of the measurements.Our results showed that 3D measurement, such as the volume of the nasal cavity, obtained from CBCT images gives an accurate representation of the anatomic dimension of the phantom.

Nasal cavity
CBCT is playing an increasingly important role in the diagnosis of morphologic abnormalities in the oral and maxillofacial region [28].Clinician must also be able to interpret the anatomical regions outside the dentomaxillary complex, such as the nasoethmoidal region [29].The nose is a complex and important structure in facial aesthetics and respiratory physiology [1].Besides, the nose plays an important role in the physiology of sleep by regulating nasal airway resistance and stimulating ventilation [20].Nasal obstruction is common in sleep apnea, contributes to OSA, and interferes with tolerance of OSA treatment with continuous positive airway pressure (CPAP) or oral appliances [30,31].Moreover, the anatomy of the nasal cavity varies among different ethnic groups, such as Chinese versus Caucasian [10].Therefore, it is necessary to develop a protocol for nasal cavity measurement, which could help clinicians diagnose and develop a treatment plan and which takes into consideration the characteristics of different ethnic groups.

Reliability
In previous studies, several segmentation procedures were developed to segment the nasal cavity with paranasal sinuses based on CBCT images [18,19].This can be a very time-consuming task when one chooses to exclude the sinuses from the simulation domain [21].To facilitate computational fluid dynamics (CFD) analysis of the nasal cavity, Keustermans et al., developed a method to segment the nasal cavity based on active shape modelling [21].However, this method is difficult for clinicians.In our study, we developed a protocol for segmentation of the nasal cavity suitable for use by clinicians.In this protocol, the boundary was defined using anatomical landmarks with high reliability, such as the posterior nasal spine point [27].In this way, the clinician was able to segment the nasal cavity without knowing the complex algorithm integrated in the software.
Although the observers in our study had different experiences with the use of the software, they showed good inter-observer reliabilities (ICC = 0.94-0.992) in the volume measurement of the nasal cavity.This finding can be explained by several factors.First, the landmark localization was well defined, and the observers were trained before performing the measurements.Second, a fixed threshold ranging from -1000 to -300 was used for the selection of the nasal cavity.However, in certain cases with a rather narrow nasal cavity, due to small detail, areas near the border of nasal cavity could not be selected only using threshold segmentation.In such cases, the observers had to scroll through all the slices to check if any area was missing and make modifications when needed using the tools integrated in the software.In this study, it is shown that there is a high interobserver reliability.Observer's subjectivity in nasal cavity segmentation could be eliminated.Therefore, it is recommended to use semi-automatic segmentation of the region of interest in studies assessing the characteristics of the nasal cavity to improve the reliability of the measurements.

Accuracy
To test accuracy, a phantom with known dimensions is often used as gold standard.In our pilot study, different approaches were tried to obtain such a phantom.At the beginning, a skull model was used as the phantom.However, since the nasal cavity is partially enclosed by soft tissue, and there is no soft tissue in the skull model, the complexity of the nasal cavity cannot be represented by the skull model.Nowadays, 3D printing applications are expanding.Using this technique, the nasal cavity and its surrounding tissues can be segmented and directly printed by a 3D printer.As the anatomy of the nasal cavity is very complex, during 3D printing, supporters have to be added to the phantom to hold the nasal cavity walls.This does not resemble the true structure of the nasal cavity.To address this problem, in this study, after removing the supports from the PLAphantom, we took an impression of the PLA-phantom using liquid silicone.Therefore, the silicone phantom used in the present study offered the unique possibility of assessing the accuracy of nasal cavity measurements under real, lifelike conditions [32].
In this study, nasal cavity volume was overestimated by 1.1%-3.1%.This phenomenon can be explained by the CBCT image acquisition process and/or the subsequent image segmentation by means of thresholding.During CBCT image acquisition, anatomical structures are differentiated based on their radiographic density.However, voxels residing on tissue boundaries can represent more than one tissue type.This phenomenon is known as the partial volume effect.The result of the partial volume effect is that voxels could erroneously be allocated to "nasal cavity" instead of "soft tissue" during the image segmentation process [33].Another reason is that as the anatomy of nasal cavity is very complex, the observer's experience in assessment of the nasal cavity could influence the results.Therefore, it is advisable to train the observers before assessment.In previous studies, thresholding methods were widely used for upper airway segmentation [25,34].However, thresholding methods alone are not sufficient to handle the segmentation of the nasal cavity, especially the nasal turbinate region [18,19,35,36].In our study, to make the segmentation more reliable and accurate, skeletal landmarks such as anterior nasal spine (ANS) was used to define the anterior boundary of the nasal cavity.In this way, the most anterior part of the nostril could not be included, which can be further investigated in future study.

Limitations
This study has several limitations.Firstly, while reliability of the protocol was shown to be excellent for both Chinese and Caucasian groups, only OSA patients were included in this study.In future, the reliability of the protocol for a larger population with different range of anatomy or nasal cycle needs further investigation.Seasonal impact on the nasal cavity could also be considered.Secondly, only one phantom based on CT images of a Chinese OSA patient rather than a Canadian OSA patient were used as a gold standard.However, to test the accuracy of the protocol, it is common to have one phantom as a gold standard.To make the protocol more feasible for the clinician, we set up specific boundaries for the nasal cavity based on skeletal landmarks.Following the protocol, part of the nasal cavity could be out of the boundary.This could be a limitation in this study.

Conclusions
The intra-observer and inter-observer reliability of 3D measurement of the nasal cavity using CBCT images were excellent.Therefore, the methodology of nasal cavity measurement used in this study is recommended in future studies on CBCT images.The 3D anthropomorphic phantom that was used in this study offered a feasible method to validate the measurement of the nasal cavity on CBCT images.

Fig. 1
Fig. 1 Boundaries of nasal cavity segmentation.a. Lower boundary of nasal cavity; b.Left boundary of nasal cavity; c.Right boundary of nasal cavity; d.Posterior boundary of nasal cavity; e. Anterior boundary of nasal cavity; f.Upper boundary of nasal cavity; g.Posterior, right and upper boundaries of nasal cavity.Green line = posterior boundary of nasal cavity (i.e., the coronal plane across posterior nasal spine point); blue line = right boundary of nasal cavity (i.e., the sagittal plane across the medial wall of the right maxillary sinus); red

Fig. 2 .
Fig. 2. a. Polylactic Acid (PLA)-phantom of nasal cavity; b.Phantom of nasal cavity made of silicone; c.CBCT images of the silicone phantom of the nasal cavity