Image Anal Štereol 2015;34:199-208 Original Research Paper doi:10.5566/ias.1334 VARIABILITY OF MANUAL AND COMPUTERIZED METHODS FOR MEASURING CORONAL VERTEBRAL INCLINATION IN COMPUTED TOMOGRAPHY IMAGES Tomaž Vrtovec^, Franjo Pernuš and Boštjan Likar University of Ljubljana, Faculty of Electrical Engineering, Tržaška cesta 25, SI-1000 Ljubljana, Slovenia e-mail: tomaz.vrtovec@fe.uni-lj.si, franjo.pernus@fe.uni-lj.si, bostjan.likar@fe.uni-lj.si (Received May 14, 2015; accepted June 11, 2015) ABSTRACT Objective measurement of coronal vertebral inclination (CVI) is of significant importance for evaluating spinal deformities in the coronal plane. The purpose of this study is to systematically analyze and compare manual and computerized measurements of CVI in cross-sectional and volumetric computed tomography (CT) images. Three observers independently measured CVI in 14 CT images of normal and 14 CT images of scoliotic vertebrae by using six manual and two computerized measurements. Manual measurements were obtained in coronal cross-sections by manually identifying the vertebral body corners, which served to measure CVI according to the superior and inferior tangents, left and right tangents, and mid-endplate and mid-wall lines. Computerized measurements were obtained in two dimensions (2D) and in three dimensions (3D) by manually initializing an automated method in vertebral centroids and then searching for the planes of maximal symmetry of vertebral anatomical structures. The mid-endplate lines were the most reproducible and reliable manual measurements (intra- and inter-observer variability of 0.7° and 1.2° standard deviation, SD, respectively). The computerized measurements in 3D were more reproducible and reliable (intra- and inter-observer variability of 0.5° and 0.7° SD, respectively), but were most consistent with the mid-wall lines (2.0° SD and 1.4° mean absolute difference). The manual CVI measurements based on mid-endplate lines and the computerized CVI measurements in 3D resulted in the lowest intra-observer and inter-observer variability, however, computerized CVI measurements reduce observer interaction. Keywords: computed tomography, computerized measurements, coronal vertebral inclination, manual measurements, measurement variability. INTRODUCTION The evaluation of spinal deformities from medical images is essential for diagnosis and treatment of pathological conditions affecting the spine. Scoliosis, which can be observed as a deviation of the spinal curve from the straight line in the coronal (frontal) plane, is one of the most frequent manifestations of spinal deformities. Accurate and objective measurement of coronal vertebral inclination (CVI) is therefore of significant importance, and the angle of inclination between the superior and inferior vertebral endplates in coronal radiographs, i.e., the Cobb angle (Cobb, 1948), is the most established measurement technique. However, such measurements are biased by observer interpretation, anatomical deformations of vertebrae and image acquisition (Capasso et al., 1992), as radiographs represent a two-dimensional (2D) projection of the observed anatomy. With the development of modern three-dimensional (3D) imaging techniques, it is possible to measure CVI from 2D cross-sectional images extracted from 3D volumes, as well as from original 3D volumetric images, which better display the 3D nature of spinal deformities. To the best of our knowledge, a systematic analysis of CVI measurements from 3D images has not been performed yet. The purpose of this study is to systematically evaluate the reproducibility and reliability of segmental CVI measurements from computed tomography (CT) images. MATERIALS AND METHODS IMAGES Fourteen vertebrae from one normal CT spine image (male subject, 47 years, Cobb angle around 1° between T5 and T12) and 14 vertebrae from one scoliotic CT spine image (female subject, 36 years, Cobb angle around 60° between T5 and T12, right thoracic curve), both including levels between T1 and L2, were included in this study. The images were acquired by the Tomoscan AVE and MX 8000 CT scanners (Philips Medical Systems, The Netherlands) for diagnostic purposes not related to this study, and the institution of origin anonymized the images before handing them over. 199 Vrtovec T et al: Variability of for measuring coronal vertebral inclination in CT images Fig. 1. Manual coronal vertebral inclination (CVI) measurements in two-dimensional (2D) oblique coronal cross-sections. OBSERVERS Manual and computerized measurements were performed by three observers (observer 1: a postgraduate biomedical engineering student; observer 2: a medical imaging researcher; observer 3: a spine surgeon) with different experience in medical imaging and orthopedic surgery, who were familiar with basic tools for visualization of spine and vertebrae, and encountered spine and vertebrae images on daily basis either for research, software development, treatment planning or evaluation purposes. Each observer independently performed a set of measurements twice, leaving a two-week period between the first and second set of measurements, resulting in six sets of measurements for each vertebra. MANUAL CVI MEASUREMENTS For the purpose of manual measurement of CVI, we developed a dedicated computer program that guided each observer step-by-step through the measurement procedure. For each vertebra, the observers first manually identified the vertebral centroid in 3D and estimated the axial and sagittal vertebral tilt, which served to automatically extract the oblique 2D coronal cross-section from the CT image. Oblique cross-sections were used instead of orthogonal to obtain the best possible coronal views and therefore reducing measurement errors that may be introduced by axial and sagittal vertebral tilt. In the oblique coronal cross-section, the observers manually identified the four vertebral body corners, which were used to evaluate CVI according to six different measurements (Fig. 1). The superior tangents and the inferior tangents represent the segmental Cobb angle (Cobb, 1948) at the superior and inferior vertebral endplate, respectively. The left tangents and the right tangents describe the inclination of the left and right vertebral body wall, respectively. The mid-endplate lines are defined between the central points of the left and right vertebral body wall, while the mid-wall lines are defined between the central points of the superior and inferior vertebral endplate. The angles of CVI were computed from the inclinations of the obtained lines against horizontal or vertical references. The average time required to perform manual measurements for one vertebra by each observer was estimated to around 4 min. Fig. 2. The planes of maximal symmetry of vertebral anatomical structures and their relation to coronal vertebral inclination (CVI). COMPUTERIZED CVI MEASUREMENTS The computerized measurements were based on a method that determines the sagittal, coronal and axial angle of vertebral rotation in 3D images from the inclination of the planes of maximal symmetry (Vrtovec et al., 2008), which divide the vertebral body into symmetrical left and right, anterior and posterior, and cephalic and caudal halves (Fig. 2). The planes of symmetry are manually initialized so that they are parallel to the axes of the CT image, centered in the vertebral centroid in 3D that represents the center of rotation, and 50 mm in size to encompass the whole thoracic or lumbar vertebral body. By rotating these planes in 3D, the symmetry of vertebral anatomical structures is automatically evaluated for each combination of the three rotation angles by mirroring the edges of anatomical structures (i.e., image intensity gradients) over each plane and comparing them to the corresponding edges on the other side of that plane. Due to the anatomical 200 Image Anal Stereol 2015;34:199-208 Fig. 3. Computerized measurement of coronal vertebral inclination (CVI) in 2D, shown for the T9 scoliotic vertebra, is performed by evaluating the symmetry in the left and right, and in the cephalic and caudal parts of the vertebral body. characteristics and deformations of vertebral bodies (e.g., wedging), the symmetry of vertebral structures may not be perfect. However, an optimization procedure is applied to search for the planes of maximal available symmetry that define the final rotation angles. By performing measurements in 2D, CVI was automatically determined in the same oblique 2D coronal cross-sections that were used for manual measurements (Fig. 3). By performing measurements in 3D, the sagittal, coronal and axial angle of rotation were simultaneously determined in 3D images, with the coronal angle of rotation representing CVI (Fig. 4). The average time required to perform computerized measurements for one vertebra was estimated to around 2 s for measurements in 2D and around 2.5 min for measurements in 3D (performed on a standard personal computer without code optimization or parallelization and without graphics processing unit acceleration). STATISTICAL ANALYSIS For each of the 28 vertebrae, CVI was determined manually 36 times (3 observers x 2 sets x 6 manual measurements) and automatically 12 times (3 observers x 2 initializations x 2 computerized measurements). Statistical analysis was performed in terms of intra-observer variability (observer reproducibility), inter-observer variability (observer reliability) and inter-method variability (measurement agreement), described by standard deviations (SD), intraclass correlation coefficients (ICC) and mean absolute differences (MAD) of the resulting CVI angles. Paired samples t-tests were used to search for statistically significant differences in the obtained results (level of significance a = 0.05, which was subjected to the Bonferroni correction where necessary). sagittal plane of coronal plane of axial plane of maximal symmetry maximal symmetry maximal symmetry (left vs. right part) (anterior vs. posterior part) (cephalic vs. caudal part) mid-coronal cross-section mid-sagittal cross-section mid-sagittal cross-section Fig. 4. The sagittal, coronal and axial planes of symmetry are used to determine the coronal inclination (CVI), sagittal inclination (SVI) and axial rotation (AVR)for the L1 scoliotic vertebra in 3D by comparing the points (e.g., P and Q) on one side of each plane to the corresponding (mirror) points (e.g., P* and Q*) on the other side of that plane. RESULTS Fig. 5 shows the mean CVI for each vertebra according to each measurement. For the normal 201 Vrtovec T et al: Variability of for measuring coronal vertebral inclination in CT images vertebrae, the results follow the spinal curvature of a normal spine, which is approximately a straight line. For the scoliotic vertebrae, the results show a right thoracic curve (represented by negative angles) followed by a compensating left thoracolumbar curve (represented by positive angles). INTRA-OBSERVER VARIABILITY Table 1 shows the intra-observer variability for each observer and for each measurement. The average intra-observer variability for observers 1, 2 and 3 was 1.1°, 1.0° and 1.5° SD (0.988, 0.992 and 0.979 ICC), respectively, for manual measurements, and 0.8°, 0.9° and 1.1° SD (0.995, 0.993 and 0.987 ICC), respectively, for computerized measurements. The average reproducibility was therefore estimated to 1.2° SD (0.986 ICC) for manual measurements and 0.9° SD (0.992 ICC) for computerized measurements (average statistical power of 0.73 for the 95% confidence interval). No statistically significant differences in measurements were found within any observer (p > 0.05). INTER-OBSERVER VARIABILITY Table 2 shows the inter-observer variability for each pair of observers and for each measurement. The average inter-observer variability for observer pairs 1/2, 1/3 and 2/3 was 1.6°, 2.0° and 1.9° SD (0.994, 0.990 and 0.991 ICC), respectively, for manual measurements, and 1.2°, 1.3° and 1.4° SD (0.996, 0.996 and 0.996 ICC), respectively, for computerized measurements. The average reliability was therefore estimated to 1.9° SD (0.992 ICC) for manual measurements and 1.2° SD (0.996 ICC) for computerized measurements (average statistical power of 0.77 for the 95% confidence interval). No statistically significant differences in measurements were found between any observer pair (p > 0.05). INTER-METHOD VARIABILITY AND DIFFERENCE The analysis of inter-method variability (SD) and inter-method difference (MAD) is presented in Table 3 for each measurement pair (average statistical power of 0.82 for the 95% confidence interval). Statistically significant differences were found between the superior tangents and left tangents (p < 0.03) or right tangents (p < 0.04), between the inferior tangents and left tangents (p < 0.04) or right tangents (p < 0.05), between the left tangents and every other method (p < 0.04), and between the right tangents and every other method (p < 0.05). However, by applying the Bonferroni correction, statistically significant differences were found only between the left tangents and right tangents, mid-wall lines or computerized measurements in 2D and in 3D (p < 0.002), and between the right tangents and mid-wall lines or computerized measurements in 2D and in 3D (p < 0.002). DISCUSSION Several methods for assessing the degree of spinal deformities in the coronal plane were developed, such as the Ferguson method (Ferguson, 1930), Cobb method (Cobb, 1948) and centroid method (Chen et al., 2007). As the Cobb angle represents the standard method for radiographic quantification of scoliotic deformities, a number of studies examined its variability. Traditional manual measurements are performed by drawing lines on antero-posterior (AP) or postero-anterior (PA) radiographs. Such measurements are biased by the selection of the most tilted endplates, errors in drawing lines and systematic errors of inaccurate measuring devices (Capasso et al., 1992). As a result, the reported intra-observer SD between 1.5° and 8.5°, and inter-observer SD between 2.5° and 8.8° (Chen et al., 2007; Jeffries et al., 1980; Oda et al., 1982; Goldberg et al., 1988; Dutton et al., 1989; Ylikoski and Tallroth, 1990; Carman et al., 1990; Pruijs et al., 1994; Loder et al., 1995; Diab et al., 1995; Shea et al., 1998; Facanha-Filho et al., 2001; Loder et al., 2004; Wills et al., 2007; Gstoettner et al., 2007; De Carvalho et al., 2007; Tanure et al. ,2010) span across a relatively large range of values. Adam et al. (2005) evaluated the Cobb angle in CT images by extracting reformatted cross-sections, resulting in intra- and inter-observer variability of 3.4° and 2.7° SD, respectively. Computer-assisted measurements, performed by manually drawing lines on digital radiographs using a computer, improved the reproducibility and reliability of Cobb angle measurements, as studies reported SD between 1.3° and 4.6° for intra-observer, and between 1.6° and 3.2° for inter-observer variability (Jeffries et al., 1980; Dutton et al., 1989; Shea et al., 1998; Wills et al., 2007; Gstoettner etal., 2007; Tanure etal., 2010; Mok et al., 2008). Chockalingam et al. (2002) performed computer-assisted measurements by constructing the spinal midline from several points that were manually identified on the left and right vertebral body walls. The reported intra- and inter-observer variability in terms of technical error of measurement (TEM) were 0.74° and 1.22° (0.985 and 0.988 ICC), which according to the equations presented by the authors result in relatively large SD of 6.0° and 11.1° SD, respectively. On the other hand, the computer-assisted 202 Image Anal Stereol 2015;34:199-208 Table 1. Intm-observer variability for observers 1, 2 and 3, reported as .standard deviations (SD) and intraclass correlation coefficients (ICC). Measurement Vertebrae Intra-observer SD (°) 12 3 mean Intra-observer ICC 1 2 3 mean superior tangents normal scoliotic both 1.0 0.5 0.6 1.0 0.9 1.4 0.9 1.0 0.888 0.999 0.958 0.998 0.857 0.996 0.901 0.998 0.8 0.8 1.2 1.0 0.998 0.997 0.994 0.996 normal 0.7 0.7 0.9 0.8 0.958 0.959 0.914 0.944 inferior tangents scoliotic 0.9 1.1 1.5 1.2 0.998 0.998 0.996 0.997 both 0.8 0.9 1.2 1.0 0.997 0.997 0.994 0.996 normal 1.7 0.8 1.3 1.3 0.693 0.939 0.810 0.814 left tangents scoliotic 1.3 1.4 2.4 1.8 0.988 0.989 0.962 0.978 both 1.5 1.1 1.9 1.5 0.971 0.986 0.954 0.970 normal 1.0 1.2 2.0 1.5 0.951 0.929 0.892 0.924 right tangents scoliotic 1.9 1.5 2.2 1.9 0.981 0.988 0.970 0.978 both 1.5 1.4 2.1 1.7 0.979 0.981 0.956 0.972 normal 0.7 0.6 0.6 0.6 0.947 0.964 0.938 0.950 mid-endplate lines scoliotic 0.5 0.8 1.2 0.9 0.999 0.999 0.997 0.998 both 0.6 0.7 0.9 0.7 0.999 0.998 0.997 0.998 normal 1.0 0.9 1.0 1.0 0.833 0.863 0.888 0.861 mid-wall lines scoliotic 1.2 0.9 1.5 1.2 0.991 0.995 0.984 0.990 both 1.1 0.9 1.3 1.1 0.985 0.990 0.978 0.984 normal 1.1 0.9 1.6 1.2 0.865 0.921 0.750 0.845 computerized (2D) scoliotic 0.8 1.3 1.5 1.2 0.997 0.992 0.988 0.992 both 1.0 1.1 1.5 1.2 0.991 0.989 0.977 0.986 normal 0.5 0.5 0.4 0.5 0.966 0.966 0.980 0.971 computerized (3D) scoliotic 0.3 0.5 0.5 0.4 0.999 0.999 0.998 0.999 both 0.4 0.5 0.5 0.5 0.998 0.997 0.997 0.997 measurements based on the identification of vertebral body corners resulted in intra-observer SD between 1.6° and 2.3°, and inter-observer SD between 2.6° and 3.2° (Tanure et al., 2010; Cheung et al., 2002; Stokes and Aronsson, 2006). Further reduction of manual observer interaction was made possible by (semi)automated computerized measurements, which incorporate image processing and analysis techniques into the Cobb angle measurements. Allen et al. (2008) developed a method based on active shape models and reported TEM for intra- and inter-observer variability of 2.0° and (0.930 and 0.940 ICC), which correspond to relatively large SD of 8.1° and 8.4°, respectively. The measurements of Zhang et al. (2010) were based on finding the inclination of the edges obtained by the Hough transform. The authors reported intra-observer SD of 1.2° (ICC between 0.916 and 0.994) and inter- observer SD between 1.8° and 2.1° (ICC between 0.908 and 0.985). Chen et al. (2007) performed manual measurements by a different method that was based on the identification of vertebral centroids and reported intra- and inter-observer variability of 2.2° and 2.6° SD, respectively. In a recent evaluation of manual and computerized measurement of CVI in magnetic resonance (MR) images (Vrtovec et al., 2013), the mid-endplate lines proved to be the most reproducible (1.0° SD) and reliable (1.4° SD) manual measurements, while the computerized measurements in 3D yielded lower intra-observer (0.8° SD) and interobserver (1.3° SD) variability. The strongest inter-method agreement (1.2° SD and 0.4° MAD) was found among lines parallel to vertebral endplates, however, the computerized measurements in 3D were most in agreement with the mid-endplate lines (1.9° SD and 203 Vrtovec T et al: Variability of for measuring coronal vertebral inclination in CT images Table 2. Inter-observer variability for observer pairs 1/2, 1/3 and 2/3, reported as standard deviations (SD) and intraclass correlation coefficients (ICC). Measurement Vertebrae Inter-observer SD (°) Inter-observer ICC 1/2 1/3 2/3 mean 1/2 1/3 2/3 mean normal 1.3 1.3 1.4 1.3 0.912 0.947 0.876 0.912 superior tangents scoliotic 1.6 1.9 2.0 1.8 0.998 0.998 0.997 0.998 both 1.6 1.6 1.8 1.7 0.996 0.997 0.995 0.996 normal 1.0 1.6 1.4 1.4 0.980 0.914 0.938 0.944 inferior tangents scoliotic 1.3 1.5 1.6 1.5 0.999 0.998 0.999 0.999 both 1.3 1.6 1.5 1.5 0.998 0.996 0.998 0.997 normal 1.7 2.1 1.7 1.8 0.936 0.862 0.871 0.890 left tangents scoliotic 2.0 2.2 2.4 2.2 0.990 0.995 0.993 0.993 both 1.9 2.2 2.1 2.1 0.988 0.988 0.986 0.987 normal 1.9 2.9 2.7 2.5 0.949 0.895 0.923 0.922 right tangents scoliotic 2.3 3.0 2.8 2.7 0.993 0.984 0.988 0.988 both 2.1 2.9 2.7 2.6 0.989 0.972 0.979 0.980 normal 0.9 1.1 1.2 1.1 0.972 0.945 0.922 0.946 mid-endplate lines scoliotic 1.2 1.3 1.4 1.3 0.999 0.999 0.999 0.999 both 1.2 1.3 1.3 1.2 0.998 0.998 0.998 0.998 normal 1.2 1.6 1.5 1.4 0.965 0.880 0.907 0.917 mid-wall lines scoliotic 1.5 1.8 1.8 1.7 0.997 0.996 0.994 0.996 both 1.3 1.7 1.7 1.6 0.996 0.990 0.990 0.992 normal 1.5 1.6 1.8 1.6 0.935 0.962 0.916 0.938 computerized (2D) scoliotic 1.6 1.7 1.9 1.7 0.996 0.997 0.995 0.996 both 1.5 1.6 1.8 1.6 0.994 0.995 0.992 0.994 normal 0.7 0.8 0.8 0.8 0.980 0.972 0.981 0.978 computerized (3D) scoliotic 0.6 0.8 0.6 0.7 0.999 0.999 0.999 0.999 both 0.7 0.8 0.7 0.7 0.998 0.997 0.999 0.998 1.1° MAD). The results obtained in the current study are, in terms of intra- and inter-observer variability, comparable to the above mentioned findings. Although none of these studies was focused on segmental measurements, the angle between arbitrary two vertebral levels can be obtained from the segmental angles, e.g. the difference between the angles of the superior and inferior tangents at two selected vertebrae results in the classical Cobb angle measurement. If the mean measured angles for the T12 and T5 vertebral level are subtracted, (-2.4° - (-5.2°)) = 2.8° is obtained for the normal and (+37.2° — (—25.4°)) = 62.6° is obtained for the scoliotic spine, which approximately correspond to the diagnosed Cobb angles of 1° and 60°, respectively. However, to compare the variability of our measurements to the classical Cobb angle measurements, the variabilities of both superior and inferior tangents have to be considered. The resulting intra-observer variability of V1.02 +1.02 = 1.4° SD and inter-observer variability of V1.72 + 1.52 = 2.3° SD are lower than the values reported by studies that performed computer-assisted Cobb angle measurements based on the identification of vertebral body corners. This may result from the fact that in radiographs, the vertebral body corners are more difficult to identify than in CT cross-sections due to the occlusion of anatomical structures and projective nature of radiographic imaging. In comparison to MR images (Vrtovec et al., 2013), the lower variability of measurements in CT images obtained in the current study points to the fact that the edges of bone structures can be extracted more accurately from CT than from MR images. Nevertheless, most of the above mentioned studies were focused on evaluating the 204 Image Anal Stereol 2015;34:199-208 Table 3. Inter-method variability (lower-left triangle in normal text), reported as .standard deviations (SD) of measurements, and inter-method difference (upper-right triangle in bold text), reported as mean absolute differences (MAD) of measurements. CD Í3 T3 U N T3 U N (H CC O ä Cß Cß T3 'C CD 'C CD • in Ö (H OJ s ! o ö •d