| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Experimental Studies |
1 From the Department of Radiology, Feinberg School of Medicine, Northwestern University Medical School, Chicago, Ill. Received June 24, 2002; revision requested August 28; final revision received May 14, 2003; accepted June 18. Address correspondence to D.S.F. Cedars-Sinai Medical Center, S. Mark Taper Imaging Center, Rm 1258, 8700 Beverly Blvd, Los Angeles, CA 90048 (e-mail: David.Fieno@eshs.org).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
MATERIALS AND METHODS: Nine dogs and five pigs underwent cine MR imaging of the entire LV from base to apex. Manual and automatic segmentation times were recorded, and LV masses determined with each were compared with each other and with the true LV mass at autopsy. Estimated mass and true mass at autopsy were compared by calculating the correlation coefficient and the mean difference between the two for each MR sequence and segmentation method.
RESULTS: True LV mass at autopsy correlated well with masses determined with manual and automatic contours on true FISP MR images. Mean differences between true LV mass and masses determined from manual contours on true FISP and FLASH images were -0.8 g ± 2.6 and 3.7 g ± 6.8, respectively. When manually drawn end-diastolic contours were automatically propagated to end systole, mean differences were 2.0 g ± 3.6 (P = .05) and 9.1 g ± 6.5 (P < .05) for true FISP and FLASH images, respectively. For automatic contours, mean differences were 10.6 g ± 8.5 (P < .05) and 27.7 g ± 13.4 (P < .05) for true FISP and FLASH images, respectively. Mean automatic segmentation time was six times less than mean manual segmentation time.
CONCLUSION: LV mass was determined most accurately by using manual contours on true FISP images. In these animal models, fully automatic segmentation of true FISP images was performed in one-sixth of the time of manual segmentation and yielded LV masses with a mean error of approximately 5% of true LV mass.
© RSNA, 2003
Index terms: Animals • Experimental study • Heart, MR, 524.12142 • Heart, ventricles • Magnetic resonance (MR), cine study, 524.12142
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Echocardiography is routinely used to determine LV mass noninvasively (1–4). However, image quality and assumptions regarding LV geometry limit the accuracy of this method for estimation of LV mass. Cardiac magnetic resonance (MR) imaging provides a method of calculating LV mass accurately (5–12) and reproducibly (7,11–17). One limitation of using MR imaging routinely for this purpose is the postprocessing time required to compute the LV mass manually on the basis of raw image data.
To decrease the amount of labor required to process the MR images, several semiautomatic and automatic segmentation routines have been developed (10,12,18–20). The utility of an image processing tool used for calculation of LV mass is dependent on its ability to allow accurate computation of LV mass and detection of changes in LV mass over time (17). However, few studies have been conducted to compare the LV mass determined from automatic segmentation of MR images with true values determined at autopsy (10,12). In addition, the robustness of such an automatic segmentation routine has not been evaluated rigorously for modern fast MR imaging sequences, such as recent applications of steady-state free precession to cardiac imaging (21). By using these newer imaging techniques, high-quality cine MR images of the heart may be acquired with relatively short imaging times (21) and may permit accurate distinction between myocardium and blood.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
MR Imaging and LV Specimens
Animals were anesthetized with sodium brevital (Eli Lilly and Company, Indianapolis, Ind), intubated, and ventilated mechanically with 2 L of oxygen per minute with 1%–3% isoflurane (Abbott Laboratories, North Chicago, Ill). Animals were imaged in the right lateral decubitus position by using a 1.5-T whole-body MR imager (Sonata; Siemens Medical Systems, Erlangen, Germany). A 14 x 28-cm flexible surface coil was positioned directly over the heart. Electrocardiographically gated cine MR images encompassing the entire LV were acquired during repeated breath holding, which was achieved by turning off the ventilator temporarily, allowing time in between for reoxygenation and reestablishment of an end tidal CO2 of approximately 35 mm Hg.
Contiguous 5-mm short-axis true fast imaging with steady-state precession (FISP) and fast low-angle shot (FLASH) cine MR images (13–15 sections per examination) were acquired from base to apex. Typical imaging parameters were repetition time msec/echo time msec of 3.0/1.6 with a 70° flip angle for the true FISP sequence and 8.0/4.0 with a 20° flip angle for the FLASH sequence. The mean field of view was 300 x 290 mm and was adjusted depending on the size of the animal. A rectangular matrix was used to reduce acquisition time. The readout matrix was always 256, while the phase-encoding matrix depended on the rectangular field of view. Mean pixel size was 1.17 x 1.17 mm (range, 1.05 x 1.05–1.41 x 1.41 mm). No section spacing was used.
After MR imaging, the animals were sacrificed, the heart was excised, and the LV was isolated by one author (D.S.F.) as described previously (22). The mitral and aortic valves were removed, and the right ventricle was trimmed off of the interventricular septum at the anterior and inferior insertions. The heart was weighed by using a digital scale, and the amount of water displaced was measured in a graduated cylinder and recorded. The mass was then verified by multiplying the experimentally determined LV volume by the density of myocardium (1.05 g/mL).
MR Image Analysis
All image data sets were analyzed by one trained observer (C.J.F.), who was blinded to postmortem results. Commercially available software (ARGUS A1.5; Siemens Medical Systems) was used for histogram-based thresholding, active contouring, and shape matching (23). Separate algorithms optimized for true FISP and FLASH MR imaging were used to analyze true FISP and FLASH images, respectively. End diastole was defined as the first phase of images acquired after the R wave of the electrocardiographic signal, and end systole was the phase with the smallest LV blood pool area at the level of the midventricle, typically about one-third of the way into the cardiac cycle. Although determination of the myocardial borders, level of the base, and section location of the LV apex can be somewhat subjective, a standardized protocol described previously (22) was followed for determination of LV margins. The short-axis base and apex were determined by ascertaining their location on long-axis MR images (22,24). Images were cropped and zoomed to maximize the size of the LV in the viewing area. First, endo- and epicardial borders were contoured manually at end diastole and end systole from base to apex. The endocardial contours were drawn to include the papillary muscles in the LV mass and to lie on the edge of the myocardium, not within the blood pool. These manually drawn contours were used to determine the LV mass based on manual segmentation.
After manual contouring, the end diastolic contours that were drawn manually were automatically propagated to end systole to generate semiautomatic contours. The value of LV mass determined from only the generated end systolic contours was defined as the semiautomatic segmentation.
Finally, automatic end diastolic and end systolic contours were generated by using autosegmentation routines optimized for FLASH and true FISP MR sequences. After selecting the LV base and apex, the automatic end diastolic contours were computed by placing circular endo- and epicardial seed contours on a midventricular image below the level of the papillary muscles. The "autofit" function was used to fit the seed contours to the LV. These fitted endo- and epicardial contours were propagated to the other images in end diastole and end systole. It should be noted that none of the semiautomatically or automatically propagated contours were adjusted. Figures 1–3 illustrate our procedure for contouring.
|
|
|
Statistics
On the basis of the results of a previous study (12) on the comparison of LV mass determined by using automatic segmentation with LV mass determined at autopsy in dogs, we calculated the sample size (n = 12) to detect an expected difference of 5 g with the Student t test for a significance level of .05 and a power of .8. MR imaging–derived LV mass values were compared with the actual LV mass values (range, 46.8–103.9 g) by using two-tailed paired Student t tests. Results are reported as mean ± SD. A P value less than .05 was considered to indicate a significant difference. In addition, linear regression analysis was used to correlate the values of LV mass determined by using MR imaging with those determined at autopsy. Bland-Altman plots (25) were used to analyze the differences between the calculated LV masses and those determined at autopsy. While the data from the dogs and pigs were initially analyzed separately, the data were combined in the final analysis to evaluate the accuracy of the autosegmentation program over a wider range of LV mass values.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
The LV masses determined with MR imaging and the true values determined at autopsy are summarized in Table 1. The mean LV mass at autopsy was 67.8 g ± 17.5 (range, 46.8–103.9 g). The mean LV masses determined from true FISP images by using manual, semiautomatic, and automatic segmentation, respectively, were 67.0 g ± 18.7 (range, 44.6–103.1 g), 69.8 g ± 19.4 (range, 44.8–107.5 g), and 72.0 g ± 20.9 (range, 40.8–109.4 g). The mean LV masses determined from FLASH images by using manual, semiautomatic, and automatic segmentation, respectively, were 71.5 g ± 20.1 (range, 42.3–103.3 g), 76.9 g ± 20.8 (range, 49.9–112.6 g), and 88.8 g ± 27.1 (range, 56.6–148.0 g). A very strong correlation (r > 0.9) was found between LV mass determined with all MR segmentation methods and true LV mass (Fig 4). In fact, there was no statistically significant difference between the mean LV mass computed from manual contours and the true LV mass for both true FISP and FLASH MR images. However, the difference between the values of LV mass determined with MR images was statistically significant (P < .05) compared with the true masses when semiautomatic and automatic methods were used. LV mass values obtained from true FISP images were significantly less than values determined from FLASH images (P < .05), as shown in Table 1.
|
|
|
|
|
|
|
|
|
|
|
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
As in the study of Barkhausen et al (20), estimates of LV mass from automatically drawn contours on true FISP images were more accurate than those on FLASH images when compared with manually drawn contours. The automatically propagated epicardial contours of the true FISP images tended to be more accurate than those of the FLASH images. This may be a result of the greater distinction between myocardium and right ventricular blood pool seen at the LV septum and between LV free wall and surrounding tissue when using the true FISP sequence. Because contrast between myocardium and blood pool in FLASH imaging is in part dependent on blood flow, there are fluctuations in contrast throughout the cardiac cycle as a result of changes in blood flow velocity (21). In true FISP imaging, such fluctuations in contrast are not seen because in-flow effects are minimized with a short echo time (21).
Because of the influence of through-plane motion on the position of the base (24), occasionally a basal MR image will be acquired that shows an interrupted myocardial border. In our study, we found that this did not lead to large errors in estimation of the mass of the most basal section with automatic segmentation.
Despite the fact that the automatic contours were similar to the manual contours at the base and the contribution of the papillary muscles was ignored, values for total LV mass were larger than the actual mass values. As shown in Figure 7, we believe the greatest source of error is secondary to overestimation of the size of the LV myocardium near the apex. For the most apical section, especially during end systole when there is thickening of the myocardial wall, there is often a section through myocardium with no blood pool visible. Baldy et al (18) described similar difficulties with automatic segmentation at the apex. Because the greatest errors in segmentation occurred at the most apical sections, a compromise between the accuracy of manual segmentation and speed of automatic segmentation might be to automatically propagate contours to all sections except the apex and then draw the contours manually for the most apical section. In studies where the base was not completely circumferential, the automatic segmentation program would include portions of the mitral valve or right atrium in the LV, causing a small overestimation in LV mass.
In humans, the transition from left atrium to LV is not as smooth as that in dogs and pigs, which may cause further problems for the automatic segmentation program. Although the present study did not address this issue in human subjects, one option for dealing with this potential problem would be to draw contours manually at the base and the apex, and then automatically propagate contours through the rest of the LV. Also, automatically drawn endocardial contours were smaller and epicardial contours larger than manually drawn contours, which caused overestimation of myocardial volume. These errors in endo- and epicardial contours were even greater for FLASH images, an observation also noted by Moon et al (26).
In the current study, the most accurate estimates of LV mass in these animal models were obtained when the papillary muscles were included in the myocardium and excluded from the LV blood pool. When the papillary muscles were excluded from the LV mass, the values were significantly lower than the true mass measured at autopsy. While our study did not include any human volunteers, authors in previous cardiac MR imaging studies of humans have described similar findings. van der Geest et al (19) found that papillary muscles accounted for 6.5% of end diastole volume, and Yamaoka et al (16) determined that they accounted for 6.7% of LV mass. In contrast to our findings, Young et al (12) found that exclusion of the papillary muscles from the LV mass in dogs resulted in a very small difference (0.3 g) in LV mass calculated with their three-dimensional model. However, the animals in their study were examined several months after rupture of the chordae tendonae, which would be expected to cause atrophy of the papillary muscles.
Previous autopsy studies have shown that human papillary muscles are normally the same thickness as the LV free wall or septum (27,28). Furthermore, there is proportional hypertrophy of the papillary muscles in patients with LV hypertrophy (27). Therefore, we believe it would be important for the papillary muscles to be routinely included in the LV mass and excluded from the LV blood in human studies, as well.
A limitation of this study is that we only evaluated animal models, while the automatic segmentation routine used in this study was designed for humans. The structure of the papillary muscles in humans is highly variable (27,28) and is characterized by many more trabeculations, which may increase the amount of time required for placing contours manually. Depending on the spacing between the trabeculations, the automatic segmentation program may place the endocardial border at the tip or base of the trabeculations. This would be expected to cause an underestimation of LV mass or overestimation of LV volume, respectively.
When compared with FLASH imaging, true FISP cine MR imaging of the heart provides excellent distinction of all parts of the LV, including the base, apex, epicardium, and papillary muscles throughout the cardiac cycle. By using a commercially available segmentation software package, true FISP cine MR images were segmented automatically in one-sixth the time required for manual segmentation and provided estimates of LV mass that were, on average, 5% greater than the true LV mass.
Practical application: While cardiac MR imaging provides accurate estimates of LV mass, its widespread clinical use in patients with heart disease is somewhat limited because of the time-consuming process of drawing the myocardial borders manually. Automatic segmentation programs have been developed but have not been evaluated rigorously, with comparisons made between estimated LV mass and actual LV mass measured at autopsy. This study demonstrated that accurate estimates of LV mass can be obtained with either FLASH or true FISP MR imaging by using manual segmentation. When an automatic segmentation routine is used to evaluate LV mass, however, FLASH images caused overestimation of LV mass by 30%, while values obtained with true FISP images were, on average, 5.6% greater than the true LV mass.
|
FOOTNOTES |
|---|
Author contributions: Guarantor of integrity of entire study, J.P.F.; study concepts, C.J.F., D.S.F., S.M.S.; study design, D.S.F., C.J.F.; literature research, C.J.F., D.S.F., S.M.S.; experimental studies, D.S.F.; data acquisition, D.S.F., C.J.F.; data analysis/interpretation, C.J.F., D.S.F., S.M.S.; statistical analysis, C.J.F.; manuscript preparation, definition of intellectual content, and editing, all authors; manuscript revision/review, J.P.F., D.S.F.; manuscript final version approval, C.J.F.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  J. P. Finn, K. Nael, V. Deshpande, O. Ratib, and G. Laub Cardiac MR Imaging: State of the Technology. Radiology, November 1, 2006; 241(2): 338 - 354. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 N. B. Mascarenhas, R. Muthupillai, B. Cheong, M. Pereyra, and S. D. Flamm Fast 3D cine steady-state free precession imaging with sensitivity encoding for assessment of left ventricular function in a single breath-hold. Am. J. Roentgenol., November 1, 2006; 187(5): 1235 - 1239. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. J. M. van Geuns, T. Baks, E. H. B. M. Gronenschild, J.-P. M. M. Aben, P. A. Wielopolski, F. Cademartiri, and P. J. de Feyter Automatic Quantitative Left Ventricular Analysis of Cine MR Images by Using Three-dimensional Information for Contour Detection. Radiology, July 1, 2006; 240(1): 215 - 221. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 C. Rickers, N. M. Wilke, M. Jerosch-Herold, S. A. Casey, P. Panse, N. Panse, J. Weil, A. G. Zenovich, and B. J. Maron Utility of Cardiac Magnetic Resonance Imaging in the Diagnosis of Hypertrophic Cardiomyopathy Circulation, August 9, 2005; 112(6): 855 - 861. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. A. Berg, V. L. Sorrell, K. B. Kern, R. W. Hilwig, M. I. Altbach, M. M. Hayes, K. A. Bates, and G. A. Ewy Magnetic Resonance Imaging During Untreated Ventricular Fibrillation Reveals Prompt Right Ventricular Overdistention Without Left Ventricular Volume Loss Circulation, March 8, 2005; 111(9): 1136 - 1140. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 P Friberg, A Allansdotter-Johnsson, A Ambring, R Ahl, H Arheden, J Framme, A Johansson, D Holmgren, H Wahlander, and S Marild Increased left ventricular mass in obese adolescents Eur. Heart J., June 1, 2004; 25(11): 987 - 992. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
