Published online before print October 1, 2001, 10.1148/radiol.2212010177
(Radiology. 2001;221:537-542.)
© RSNA, 2001
Automated Observer-independent Acquisition of Cardiac Short-Axis MR Images: A Pilot Study1
Boudewijn P. F. Lelieveldt, PhD,
Rob J. van der Geest, MSc,
Hildo J. Lamb, PhD,
Hein W. M. Kayser, MD and
Johan H. C. Reiber, PhD
1 From the Division of Image Processing (B.P.F.L., R.J.v.d.G., J.H.C.R.), Department of Radiology (H.J.L.), Leiden University Medical Center, Bldg 1 C2-S, Albinusdreef 2, PO Box 9600, 2300 RC Leiden, the Netherlands; and Interuniversity Cardiology Institute of the Netherlands, Utrecht, the Netherlands (H.W.M.K., H.J.C.R.). Received December 22, 2000; revision requested February 5, 2001; revision received March 29; accepted May 14. Address correspondence to J.H.C.R. (e-mail: j.h.c.reiber@lumc.nl).
 |
ABSTRACT
|
|---|
The authors compared an automated observer-independent acquisition planning method for short-axis multisection multiphase cardiac magnetic resonance imaging studies with conventional manual image planning. Systematic and random differences and reproducibility of left ventricular function measurements and image geometry were evaluated in five healthy adult volunteers and 20 patient studies. Results with the automated planning method were as accurate and reproducible as those with the manual planning method.
Index terms: Heart, function, 524.91 • Heart, MR, 524.121416 • Heart, ventricles, 524.121416, 524.91 • Magnetic resonance (MR), cine study, 524.121416 • Magnetic resonance (MR), volume measurement, 524.121416
|
INTRODUCTION
|
|---|
Cardiac magnetic resonance (MR) imaging is an excellent tool to assess global and regional left ventricular (LV) function, and it is now regarded as a reference standard for analysis of ejection fraction (EF) and ventricular volumes (1). Multisection multiphase short-axis cardiac MR images are most suitable to assess LV function without any assumptions about LV geometry (2–11), but the planning procedure for these images is time-consuming and requires substantial insight into cardiac anatomy. A number of intermediate long-axis acquisitions is often necessary to accurately determine the optimal orientation of the short-axis volume, and many radiologists and technicians find it difficult to plan the short-axis images in a time-efficient and reproducible manner.
Recently, image analysis tools were introduced to help automated planning of the short-axis image and reduce radiologist interaction to a minimum. This approach is based on the matching of a digital atlas of the thorax to scout or localizer images in a particular patient, and technical details are described elsewhere (12–13). This procedure can be applied to calculate the two parameters required to plan a short-axis volume: the position and the orientation of the LV long axis. The purpose of our study was to assess whether the accuracy and reproducibility of the automated image planning method are equal to those achieved with conventional manual image planning.
|
Materials and Methods
|
|---|
Automated and Manual Image Planning Protocols
In previous work (12,13), we described a three-dimensional deformable atlas matching method for the thorax, with which a thoracic MR image can be automatically segmented into the lungs, heart, and cardiac ventricles. Use of the model matching procedure results in a set of geometric transformations that define a one-to-one mapping between the modeled thoracic anatomy and that of a particular subject. Anatomic and pathologic variations in organ shapes and dimensions are described in terms of variations in size, position, and orientation of the different thoracic organs with respect to each other. When the model is aligned with the image, the approximate locations of the boundaries of the heart, cardiac ventricles, and right and left lungs are known. Boundary localization with the modeling and matching procedure has been validated in healthy subjects (12) and patients (13); the method was robust with respect to noise, image artifacts, and initial model position.
By matching this thorax model to scout images, the geometric transformations that map the model to the scout images can be applied to estimate the position and orientation of the LV long axis. On the basis of these estimates, the geometry of the short-axis image volume is automatically defined directly from the scout images. The three-dimensional angulation of the LV long axis, which is defined in this study as the geometric principal axis of the left ventricle, is used as the orientation of the image. Figure 1 is an example of a matching result and the corresponding automatically defined short-axis image geometry.
View larger version (120K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1. Example of a deformable thorax model after automatically matching it to a thoracic scout image. Top row: Three orthogonal sections from a thorax scout image consist of 27 sections. The dotted boundary strips represent the location of the boundaries of the left lung, right lung, diaphragm, and epicardial surfaces. Bottom row: The image geometry for this subject was automatically derived from the deformed model.
|
|
Accuracy and reproducibility with the automated image planning procedure were compared with those achieved with conventional manual image planning. A standardized manual MR planning protocol to determine LV function (11) was adopted as a manually defined independent reference. According to this protocol, patient studies minimally consist of the following four images per patient:
1. Cardiac gated scout image (nine coronal, nine sagittal, and nine transverse images) acquired with a multistack protocol during free breathing (turbo field-echo pulse sequence: repetition time of 8 msec, echo time of 4 msec, flip angle of 20°, section thickness of 10 mm, section gap of 1 mm, field of view of 450 x 450 mm).
2. Cardiac gated vertical long-axis image, which was planned manually on a transverse section in the scout images. The image was acquired in a plane parallel to the imager bore axis and the LV long axis, through the LV apex (fast field-echo echo-planar pulse sequence: repetition time equal to the R-R interval, echo time of 11 msec, prospective triggering, end expiration, section thickness of 8 mm, reduced field of view of 400 x 237 mm).
3. Four-chamber image, which was planned on the vertical long-axis image. The image was acquired in a plane through the apex that was oriented perpendicular to the vertical long-axis image and parallel to the LV long-axis image, as seen in the vertical long-axis image (fast field-echo echo-planar pulse sequence: repetition time equal to the R-R interval, echo time of 11 msec, prospective triggering, end expiration, section thickness of 8 mm, reduced field of view of 400 x 237 mm).
4. Multisection multiphase gated short-axis cardiac MR image, which was acquired perpendicular to the four-chamber image through the apex and perpendicular to the LV long axis (fast field-echo echo-planar pulse sequence: repetition time equal to the R-R interval, echo time of 11 msec, prospective gating, end expiration, section thickness of 10 mm, section gap of 1 mm, reduced field of view of 400 x 237 mm).
Experimental Setup
Two experimental studies, both approved by our institutional review board, were performed. Manual and automated short-axis image planning were compared on the basis of two criteria: (a) prospectively, the accuracy and reproducibility of clinical parameters, such as LV EF, end-diastolic (ED) volume, end-systolic (ES) volume, LV wall mass, and ED and ES LV wall thicknesses; and (b) retrospectively, image geometry.
All MR studies were performed with a standard 1.5-T MR imaging system (Gyroscan NT; Philips, Best, the Netherlands).
Clinical parameters.—To compare the accuracy and reproducibility of clinically relevant parameters calculated with the automated planning protocol and with the manual planning protocol, five healthy adult volunteers (three men and two women; age range, 22–38 years; mean age, 29 years) without history of cardiovascular disease underwent manually and automatically planned short-axis MR imaging. Our institutional review board approved the study. Informed consent was obtained from all subjects. To investigate the reproducibility, the same procedure was repeated after the subjects were removed from the imager bore and subsequently repositioned. The number of sections in the automatically defined short-axis grid was fixed at 13 to ensure complete coverage of the LV cavity in the image volume. Computations were performed with a standard workstation (Ultrasparc 10; Sun Microsystems, Mountain View, Calif).
Owing to the developmental stage of the automated image planning procedure, the automatically computed imaging parameters were manually transferred to the imager console. No subjective user interaction was allowed during automated image planning. The time required for automated image planning was 5–7 minutes, including acquisition of the scout views (2 minutes) and calculation of the acquisition parameters for the imaging volume (3–5 minutes). On all image pairs acquired (2 x 5 manual images; 2 x 5 automatic images; total, 20 short-axis images), the contours of the LV endo- and epicardium were manually drawn by an independent expert (R.J.v.d.G.) by using cardiac MR postprocessing software (MRI-MASS; Medis Medical Imaging Systems, Leiden, the Netherlands). Contours were outlined on the basis of the conventions for endo- and epicardial contour drawing (10,14), that is, papillary muscles and epicardial fat were excluded from the contours. LV mass, ED volume, ES volume, and EF were calculated. In addition, the mean LV wall thickness was calculated in end diastole and end systole by averaging results in several contiguous sections of the left ventricle. The most apical and the most basal sections were excluded from these wall thickness analyses because, in some cases, no endocardium was present in systole in these outer sections as a result of LV longitudinal contraction.
Image geometry.—To compare the accuracy of image geometry obtained with the automated planning method with that achieved in routine clinical practice, 20 patient examinations were collected in a retrospective study. Our institutional review board approved this component of our study and did not require patient informed consent. No gender or age selection was performed. Patients were randomly selected. Their diagnoses included various pathologic cardiac conditions: myocardial infarction, n = 14; cardiomyopathy, n = 3; congenital anomaly, n = 1; arrhythmia, n = 1; and LV aneurysm, n = 1. All examinations were acquired with the manual planning protocol. Automated image planning was applied retrospectively to the 20 scout views to determine the position of the center of gravity of the image volume and the orientation of the imaging planes.
Comparison between manual and automated image geometry.—To compare the geometry of the automatically and manually planned imaging volumes, the orientation of the geometric LV long axis at end diastole in the manually imaged data sets was selected as the reference standard for the orientation of the LV long axis. To calculate this orientation, the contours of the endocardium were manually drawn by an expert (R.J.v.d.G.) on all short-axis cardiac MR images at end diastole. The orientation of the reference standard LV long axis was calculated by fitting a straight line through the contour centroids of the ED contours by means of a least squares distance criterion. Figure 2 depicts an example of a reference standard LV long axis.
View larger version (75K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2. Example of a reference LV long axis (line). The long-axis position and orientation were calculated from LV contours manually drawn on a short-axis image, which was acquired in the same examination. (a) LV long axis projected on a vertical long-axis image. (b) LV long axis projected on a four-chamber image.
|
|
To define the best accuracy that can be achieved and to calculate an allowed error margin for the reference long axis, the variation in the LV orientation during the cardiac cycle (C) was calculated. This was estimated by assessing the angle 
C between the ED and ES geometric LV long axes as determined from the contours on the short-axis images in end diastole and end systole, respectively.
The variability introduced in the manual (M) image planning process was assessed by calculating the angle M between the reference standard LV long axis and the orientation (normal vector) of the manually planned short-axis images (Fig 3). This angle expresses the error introduced by the radiologist that was accumulated during all the stages of the manual planning procedure, and it provides an indicator of the accuracy with which the LV long axis is determined in routine clinical practice with use of a standard clinical MR system.
View larger version (13K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3. To quantify the errors in both the manual and automatic image planning procedures, the angle between the vector n (perpendicular to the image plane) and the geometric LV long axis v was calculated.
|
|
To evaluate the automatically (A) estimated versus the reference standard LV long-axis orientations, the angle A was calculated. The angle A represents the differences in orientation between the automatically defined and reference standard imaging volumes.
Statistical Analysis
Values of the functional parameters EF, ED and ES volumes, and ED and ES wall thicknesses were obtained for the automated and manual image planning procedures. A two-tailed paired-samples t test was used to compare these measurements. The level of significance was a P value of .05.
To evaluate the image geometry, the difference between M and A was investigated with a two-tailed paired-samples t test. The difference in SDs in M and A was tested with a one-tailed F test. The level of significance was a P value of .05.
|
Results
|
|---|
Clinical Parameters
An example of manually and automatically planned short-axis MR images is given in Figure 4, which illustrates their high level of agreement. The mean paired differences in EF, ED and ES volumes, ED LV wall mass, ED and mean ES wall thicknesses are given in Table 1. The paired differences in these parameters were not significant.
View larger version (104K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4. Comparison between (top) manually and (bottom) automatically planned short-axis images. Although the section locations differ slightly, the appearances of the left and right ventricles are similar on both image sets. LV = left ventricle, RV = right ventricle.
|
|
View this table:
[in this window]
[in a new window]
|
TABLE 1. Comparison of the Accuracy of Six LV Function Parameters as Measured in 10 Manually or Automatically Planned Image Pairs
|
|
Table 2 lists the mean paired differences in the clinical parameters for repeated studies. None of the differences were significant.
Image Geometry
Results of the quantitative comparison between the automatically and manually planned image sets in the 20 examinations are given in Table 3.
View this table:
[in this window]
[in a new window]
|
TABLE 3. Offset Angles for Automatic and Manual Procedures and between ED and ES LV Long Axes Caused by Cardiac Contraction
|
|
Angular variations in the orientation of the reference LV long axis during the cardiac cycle, angle C, was a mean 6.1° (SD, 5.0°; range, 0.2°–22.2°). Angle M was a mean 9.7° (SD, 5.8°; range, 2.2°–28.9°), while the angular offset angle A was a mean 12.2° (SD, 6.8°; range, 1.9°–25.2°). The difference between M and A was not significant in a two-tailed paired samples t test (P = .23). Figure 5 shows two examples of automatically defined image grids to illustrate the geometric accuracy of automated image planning.
View larger version (71K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 5. Two examples of automatically planned imaging grids. Left: Imaging grids were depicted on three orthogonal sections from the scout data for two patients (top, bottom). Right: To visually verify the correctness of the planned image volumes, the imaging grids were projected on the four-chamber and vertical long-axis views, respectively, which were acquired with the manual image planning protocol. These images were not used in the model matching procedure.
|
|
|
Discussion
|
|---|
The comparison between global LV function parameters calculated from automatically and manually planned images shows a good correspondence in EF, ES and ED volumes, and LV wall mass. Good correspondence was found in two regional parameters, mean ED and ES wall thicknesses. In the paired differences between manually and automatically planned images, no statistically significant difference was present for all these parameters. The SDs in the paired differences of these parameters (Table 1) were well within clinically acceptable margins, and they are in the range of commonly reported inter- and intraobserver variabilities that result from manual contour drawing (10).
Regarding the reproducibility of clinical parameters assessed with both methods, no statistically significant difference between repeated study measurement pairs was present. Again, variations were well within the range of inter- and intra-observer variabilities associated with manual contour drawing.
The manually introduced angle between the short-axis image orientation and the reference LV long axis (M) together with the change in direction of the LV long axis during the cardiac cycle (C) resulted in differences from the reference standard orientation in an order of magnitude between 2° and 29°. This illustrates that short-axis views as planned manually in routine clinical practice may deviate substantially from the optimal image perpendicular to the geometric LV long axis. The error in the automated planning method, A, however, varied in a slightly smaller range (between 2° and 25°) as compared with the error in the manual planning procedure. The mean angular offset in the automatically planned short-axis volumes was 12.2°, which is comparable to the inaccuracies in the manual image planning procedure (mean, 9.7°). This difference was not statistically significant (P = .23).
The automated method also demonstrated an SD comparable to that with the manual method (6.8° vs 5.8°) in the angular offset from the reference standard LV long axis. This difference was not significant (one-tailed F test, P = .49), and it is in the same order of magnitude as the SD in C, 5.0°, which represents the variation in orientation of the LV long axis during the cardiac cycle.
On the basis of these findings, we conclude that the automated and manual image planning procedures are equally accurate and reproducible for the assessment of LV function in terms of EF, ED LV mass, and ED and ES volumes. This conclusion is supported by the fact that the measurements in ED and ES wall thicknesses and the orientation of the LV long axis did not differ significantly between the methods.
The automated image planning method was intended for application in a supervised manner. In this study, we eliminated any user interaction to investigate the accuracy and reliability of the method. The automated planning procedure performed with accuracy and reproducibility comparable to those with manual image planning in routine clinical practice. However, we recommend implementation of user interaction by means of visualization of the calculated image volume on the scout image sets and by allowing the radiologist to intervene. On the basis of the results from this study, we expect that such user interaction was not required in 90% of the cases.
The number of sections in the short-axis image volume was fixed in the present study. As a result, a few redundant sections that did not contain the left ventricle were included in some cases. In practice, this parameter can be adjusted manually if required. Future investigations should focus on the automated assessment of this and a number of other imaging parameters required for different cardiac function protocols, such as the navigator position for non–breath-hold protocols and the location of the aortic arch for the acquisition of velocity-encoded aortic-flow images.
The automated image planning method was developed to work with current standard MR imaging hardware. A recent advance in MR acquisition technology was the development of real-time MR imagers, which allow interactive image planning of the short-axis imaging plane. This may substantially reduce the time required to plan the short-axis image planes. However, the main advantage of automatic image planning over real-time planning is that it is not subjective and does not require specialized expert knowledge to determine the optimal short-axis views.
The time required for the current automated image planning procedure is 5–7 minutes. In general, this is only slightly shorter than the time required for the manual image planning procedure. However, we foresee two possibilities that will speed up the automated process. First, the rapid evolution of computer hardware will enable us to further shorten the automated matching procedure. Second, we performed the initial experiments by applying the automated model matching procedure to only a subset of the scout image volume. In this way, the planning procedure could be shortened by a factor of 2–4 with comparable accuracy. These two developments may potentially reduce the automated planning time to 3 minutes, including scout imaging. Comparative studies to evaluate the time efficiency of the manual and automated short-axis planning protocols are needed.
In conclusion, this study investigated the feasibility of an observer-independent image planning method for short-axis cardiac MR images; results with the automated planning procedure showed accuracy and reproducibility comparable to those with the manual planning used in routine clinical practice. Currently, the time for the automated planning protocol is slightly shorter than that for manual planning, mainly because no additional long-axis support images are required. The main advantage of the automated method during the manual planning procedure, however, is that it provided an observer-independent uniform planning approach that required no expert knowledge. The method yielded images with quality comparable to that with manual planning. No statistically significant differences were found between the methods in terms of accuracy and reproducibility of measurements of EF, ED LV and ES LV volumes, ED LV wall mass, and ED and ES wall thicknesses.
|
FOOTNOTES
|
|---|
Abbreviations: EF = ejection fraction,
ED = end diastolic,
ES = end systolic,
LV = left ventricular
Author contributions: Guarantor of integrity of entire study, J.H.C.R.; study concepts and design, B.P.F.L.; literature research, B.P.F.L.; clinical and experimental studies, R.J.v.d.G.; data acquisition, H.J.L., H.W.M.K.; data analysis/interpretation, H.J.L.; statistical analysis, R.J.v.d.G.; manuscript preparation and definition of intellectual content, R.J.v.d.G., B.P.F.L.; manuscript editing, J.H.C.R.; manuscript revision/review, B.P.F.L.; manuscript final version approval, J.H.C.R.
|
REFERENCES
|
|---|
-
Yang PC, Kerr AB, Liu AC, et al. New real-time interactive cardiac magnetic resonance imaging system complements echocardiography. J Am Coll Cardiol 1998; 32:2049-2056.
-
Shapiro EP, Rogers WJ, Beyar R. Determination of left ventricular mass by magnetic resonance imaging in hearts deformed by acute infarction. Circulation 1989; 47:706-711.
-
Haag UJ, Maier SE, Jakob M, et al. Left ventricular wall thickness measurements by magnetic resonance: a validation study. Int J Cardiac Imaging 1991; 7:31-41.
-
Azhari H, Sideman S, Weiss JL, et al. Three-dimensional mapping of acute ischemic regions using MRI: wall thickening versus motion analysis. Am J Physiol 1990; 259:H1492-H1503.
-
van Rugge FP, van der Wall EE, Spanjersberg SJ, et al. Magnetic resonance imaging during dobutamine stress for detection of coronary artery disease: quantitative wall motion analysis using a modification of the centerline method. Circulation 1994; 90:127-138.
-
van der Geest RJ, de Roos A, van der Wall EE, Reiber JHC. Quantitative analysis of cardiovascular MR images. Int J Cardiac Imaging 1997; 13:247-258.
-
van der Geest RJ, Buller VGM, Jansen E, et al. Comparison between manual and automated analysis of left ventricular volume parameters from short axis MR images. J Comput Assist Tomogr 1997; 21:756-765.
-
Lamb HJ, Singleton RR, van der Geest RJ, Pohost GM, de Roos A. MR imaging of regional cardiac function: low-pass filtering of wall thickness curves. Magn Reson Med 1995; 34:498-502.
-
Holman ER, Vliegen HW, van der Geest RJ, et al. Quantitative analysis of regional left ventricular function after myocardial infarction in the pig assessed with cine magnetic resonance imaging. Magn Reson Med 1995; 34:161-169.
-
Pattynama PMT, Lamb HJ, van der Velde EA, van der Wall EE, de Roos A. Left ventricular measurements with cine and spin-echo MR imaging: a study of reproducibility with variance component analysis. Radiology 1993; 187:261-268.
-
Lamb HJ, Doornbos J, van der Velde EA, Kruit MC, Reiber JHC, de Roos A. Echo-planar MRI of the heart on a standard system: validation of measurement of left ventricular function and mass. J Comput Assist Tomogr 1996; 20:942-949.
-
Lelieveldt BPF, van der Geest RJ, Rezaee MR, Bosch JG, Reiber JHC. Anatomical model matching with fuzzy implicit surfaces for segmentation of thoracic volume scans. IEEE Trans Med Imaging 1999; 18:218-230.
-
Lelieveldt BPF, Sonka M, Bolinger L, et al. Anatomical modeling with fuzzy implicit surface templates: application to automated localization of the heart and lungs in thoracic MR volumes. Comput Vis Image Und 2000; 80:1-20.
-
Holman ER, Buller VGM, de Roos A, et al. Detection and quantification of dysfunctional myocardium by magnetic resonance imaging: a new three-dimensional method for quantitative wall thickening analysis. Circulation 1997; 95:924-931.
TOP
ABSTRACT
INTRODUCTION
