Original Research

Patient Motion Correction for Multiplanar, Multi-Breath-Hold Cardiac Cine MR Imaging

Piotr J. Slomka, PhD,1–3* David Fieno, PhD, MD,1,2 Amit Ramesh, MSc,1

Vaibhav Goyal, MSc,1,2 Hidetaka Nishina, MD,1 Louise E.J. Thompson, MBChB,1

Rola Saouaf, MD,1 Daniel S. Berman, MD,1–3 and Guido Germano, PhD1–3

Purpose: To correct for spatial misregistration of multi-breath-hold short-axis (SA), two-chamber (2CH), and four-chamber (4CH) cine cardiac MR (CMR) images caused byrespiratory and patient motion.

Materials and Methods: Twenty CMR studies from consec-utive patients with separate breath-hold 2CH, 4CH, and SA20-phase cine images were considered. We automaticallyregistered the 2CH, 4CH, and SA images in three dimen-sions by minimizing the cost function derived from planeintersections for all cine phases. The automatic alignmentwas compared with manual alignment by two observers.

Results: The processing time for the proposed method was�20 seconds, compared to 14–24 minutes for the manualcorrection. The initial plane displacement identified by theobservers was 2.8 � 1.8 mm (maximum � 14 mm). Adisplacement of �5 mm was identified in 15 of 20 studies.The registration accuracy (defined as the difference be-tween the automatic parameters and those obtained byvisual registration) was 1.0 � 0.9 mm, 1.1 � 1.0 mm, 1.1 �1.2 mm, and 2.0 � 1.8 mm for 2CH-4CH alignment and SAalignment in the mid, basal, and apical regions, respec-tively. The algorithm variability was higher in the apex(2.0 � 1.9 mm) than in the mid (1.4 � 1.4 mm) or basal(1.2 � 1.2 mm) regions (ANOVA, P � 0.05).

Conclusion: An automated preprocessing algorithm canreduce spatial misregistration between multiple CMR im-ages acquired at different breath-holds and plane orienta-tions.

Key Words: cardiac MR; image registration; motion correc-tion; heart; respiratory motionJ. Magn. Reson. Imaging 2007;25:965–973.© 2007 Wiley-Liss, Inc.

CINE CARDIAC MAGNETIC RESONANCE (CMR) imag-ing is an established method for evaluating cardiacfunction (1–3). One limitation of this technique, how-ever, is the need to perform multi-breath-hold acquisi-tions in the course of a study in which separate short-axis (SA) and long-axis (LA) 2D images are obtained atvarious locations of the ventricle (4). In principle, thesemultiple 2D planes can define the 3D spatial model ofthe heart because the relative 3D slice coordinates arerecorded in the image headers. However, spatial mis-registration may subsequently occur due to the varyingposition of a patient’s diaphragm during consecutivebreath-holds, bulk patient motion, or both. Althoughthe existence of such spatial misregistration has beenacknowledged (4), the magnitude of the differences hasnot been reported, and multiplanar realignment of mul-tiple breath-hold acquisitions is not routinely per-formed at present. Specialized acquisition protocols,such as navigator techniques (5) and single-breath-hold (4D) cine imaging methods (6), have been proposedto address this problem; however, multi-breath-holdacquisitions are commonly performed because theyprovide optimal image quality.

In this study we present a software approach for cor-recting data obtained with a standard-multi-breathhold protocol. We describe the design, implementation,and initial validation of an automatic computer algo-rithm that allows correction of interbreath motion byregistration of multiple SA and LA CMR views into com-mon 3D spatial coordinates. This algorithm is com-pletely automated and can be added as a postprocess-ing step in a standard CMR study. Composite 3D cinedisplays can be created, allowing simultaneous 3D vi-sualization of correctly registered LA and SA views. Thealgorithm was compared with the manual correction bytwo independent observers, and the amount of misreg-istration present in the original images was quantified.The algorithm was also tested in single-breath-holdstudies with simulated motion.

MATERIALS AND METHODS

Patients

Studies from 20 consecutive patients (10 males and 10females) undergoing routine CMR examination at our

1Department of Imaging, Cedars-Sinai Medical Center, Los Angeles,California, USA.2Department of Medicine, Cedars-Sinai Medical Center, Los Angeles,California, USA.3Department of Medicine, David Geffen School of Medicine, Universityof California–Los Angeles, Los Angeles, USA.*Address reprint requests to: P.S., AIM Program/Dept. of Imaging,Cedars-Sinai Medical Center, #A047, 8700 Beverly Blvd., Los Angeles,CA 90048. E-mail: Piotr.Slomka@cshs.orgReceived April 6, 2006; Accepted December 18, 2006.DOI 10.1002/jmri.20909Published online in Wiley InterScience (www.interscience.wiley.com).

JOURNAL OF MAGNETIC RESONANCE IMAGING 25:965–973 (2007)

© 2007 Wiley-Liss, Inc. 965

institution were selected retrospectively for this analy-sis. No selection of patients was attempted with regardto the amount of the breathing motion. The reported leftventricular (LV) ejection fraction in these patients var-ied from 25% to 83% and end-diastolic (ED) ventricularvolumes varied from 76 to 272 mL. All of the patientsconsented to the use of their data for the retrospectiveCMR research in accordance with the policies set forthby the Institutional Review Board of Cedars-Sinai Med-ical Center.

MRI Acquisition Protocol

SA, two-chamber (2CH), and four-chamber (4CH) cineimages were acquired on a clinical 1.5T MRI scanner(Siemens Sonata, Erlangen, Germany) as part of theclinical CMR exam. The number of SA images variedfrom 6 to 11 depending on the size of the heart. Follow-ing scouts of the SA and LA of the heart, cine imageswere acquired using a retrospectively ECG-gated se-quence in 2CH and 4CH views as well as SA viewscovering the LV, as previously described (7). Each viewwas acquired in a single breath-hold. The imaging pa-rameters were as follows: field of view (FOV) � 350 mm,matrix � 256 � 192 to 256 � 256 depending on patientsize (pixel size � 1.37–1.82 mm), slice thickness � 8mm, slice gap � 2 mm, TE/TR � 1.6/3.1 msec, pixelbandwidth � 930 kHz, flip angle � 60°, and segmentsize � 5–9 lines depending on heart rate. Each cine viewhad 20 phases. After imaging was completed, DICOMimage data sets were copied to a PC workstation forfurther processing.

Alignment Algorithm

An overview of the registration process is shown in Fig.1. The algorithm registered the 2CH with 4CH views

and subsequently registered all SA views to the prereg-istered 2CH-4CH ensemble. Gaussian-smoothed (10-mmkernel) image intensity profiles were extracted from 2CHand 4CH images. These 2CH and 4CH intensity profiles(I2Ch, I4Ch) were defined as the voxel samples in the direc-tion of the 2CH-4CH plane intersection, and were ex-tracted independently for each cine phase and combinedinto one N-pixels-long profile, where N is the product ofthe number of cine phases and the length of the individualintersection. A cost function, C1, based on the combina-tion of correlation coefficients and sums of absolute voxeldifferences was devised to define the similarity measureas follows:

C1 � a � f�I2Ch,I4Ch� � b � g�I2Ch,I4Ch�, (1)

where a and b are empirically determined constants(a � 0.5, b � 2.0), and f is a profile difference functiondefined as

f�I2Ch,I4Ch� �1N �

j � 1

N

���Ij2Ch� � ��Ij

4Ch��, (2)

where

��Ij� � � 0.5(Ij � 1 Ij 1) 1 � j � NI2 I1 j � 1

IN � IN 1 j � N, (3)

and g(I2Ch, I4Ch) is the correlation coefficient betweenthe two intensity profiles defined as

g�I2Ch,I4Ch� �

�j

�Ij2Ch � 2Ch� � �Ij

4Ch � 4Ch�

�j

�Ij2Ch � 2Ch)2�

j

�Ij4Ch � 4Ch)2

, (4)

Figure 1. Principle of the plane registra-tion algorithm. Left: Initially, 2CH and4CH image intensity profiles extractedfrom the intersection of the 2CH and 4CHplanes (white dotted line) are coregistered,establishing x and y shift parameters forthe 2CH plane. All cine phases (from ED toES phase) are combined in the registra-tion to maximize the correlating informa-tion. Right: After 2CH-4CH registration, asimilar process is performed for each SAplane. For a given SA plane, SA intensityprofiles from all phases are coregisteredwith 2CH and 4CH profiles from allphases, resulting in independent x, yshifts for each SA plane. Correspondingintensity profiles are extracted from theSA-2CH and SA-4CH intersections (whitedotted lines) Additionally, tilting of SAplanes in two directions is implemented asanisotropic scaling of the SA plane in the xand y directions. [Color figure can beviewed in the online issue, which is avail-able at www.interscience.wiley.com.]

966 Slomka et al.

where 2Ch and 4Ch are the means of the respective2CH and 4CH profiles.

The values of the constants a and b, which define therespective weights for the sum of absolute differencesand the correlation coefficient, were determined by op-timizing the agreement of the automatic alignment withthe visual alignment in a subset consisting of the firstfive studies selected from the group of 20 patients.These values were modified in increments of 0.5, andthe (a, b) pair with the smallest maximum differencebetween the visual and automated alignment was se-lected. We used a combination of two cost functions (fand g) since the correlation function (g) pertains to therelative changes of intensities, such as edges and gra-dients, and the direct subtraction of intensities (f) re-lates to the absolute level of intensities, which shouldbe similar for the same tissue and imaging sequence. Inpractice, this combination yielded better results(smaller maximum differences between the automaticand visual correction) compared to those obtainedwhen we used one of these cost measures alone (i.e.,when one of the weights (a, b) was set to 0) during ourinitial testing.

Ideally, in perfect registration, intersecting 2CH and4CH profiles should result in the maximum correlationand minimum sum of the absolute differences. Subse-quently a simplex optimization (8) found the optimal2CH-4CH alignment by adjusting the relative positions(x, y shifts) of the 2CH plane with respect to the 4CHplane. The algorithm terminated when the decrease inthe cost function was �0.01% in each simplex cycle.Due to a lack of adequate spatial information in the2CH-4CH intersection, the relative rotation of theplanes was not considered.

Following the 2CH-4CH registration we aligned eachSA plane to the 2CH-4CH 3D ensemble using an anal-ogous principle. However, during this part of the com-putation, two profile pairs were obtained for each SAslice: a SA-2CH pair at the SA-2CH intersection, and aSA-4CH pair at the SA-4CH intersection. We extractedthese profiles by changing the relative 3D SA planeposition with respect to the LA planes. Shifts for eachSA slice were computed independently. The cost func-tion C2 was defined as follows:

C2 � a � �f�ISA,I2Ch� � f�ISA,I4Ch��

� b � �g�ISA,I2Ch� � g�ISA,I4Ch��, (5)

where ISA is the SA intensity profile for a slice.To allow a 3D rotation of the SA plane with respect to

the LA views, we implemented anisotropic stretching ofthe SA profiles in the x and y directions (the geometricalequivalent of plane tilt). The cost function for each SAplane depended on four parameters (x, y shifts and x, yangles). In-plane rotation of SA slices was disabled dueto the lack of characteristic image features needed forsuch registration on the 2CH and 4CH planes corre-sponding to the SA plane. As in the 2CH/4CH registra-tion, the simplex algorithm found the optimal positionand rotation (stretch) independently for each SA slice.All cine phases were considered in the calculation of thecost function. In total, 40 profiles were used to optimize

the 3D position for each SA slice (240–440 profiles perpatient, depending on the number of slices acquired).For each patient, 3D SA alignment was performed withand without the plane rotation enabled.

Manual Slice Alignment

To verify the automatic alignment, we performed anindependent manual alignment for the 2CH and 4CHviews, as well as for each SA slice, with the software tooldeveloped for this purpose. We implemented the man-ual alignment tool in C�� using the OpenGL libraries.This tool allowed the user to align interactively the 2CHand 4CH views and subsequently align the SA views in3D with the 2CH-4CH ensemble in the views. We per-formed the alignment in 1-mm increments using aslider with subvoxel translations and bilinear interpo-lation. Each adjustment could be interactively viewedin multiple orientations, which were displayed simulta-neously in 3D mode (as shown in the Results section,Figs. 2 and 3). Each SA slice (N � 6–11 for each patient)was adjusted independently. Simultaneous cinematicdisplay in multiple orientations was allowed during theinteractive position adjustment.

Because of the complexity of the 3D SA manual align-ment with rotation enabled, the users adjusted onlyin-plane x and y shifts. Two independent observers(D.F. and H.N., both physicians with experience in clin-ical CMR imaging) were presented with original sets ofSA, 2CH, and 4CH images, and aligned images with theuse of the software tool described above. The users wereblinded to the results obtained by the automated algo-rithm. User-defined alignment parameters were subse-quently compared with the computer-generated re-sults. Since the observers could not identify obvious tiltmisalignments, and the automatic tilt adjustmentswere small, alignment with SA tilt disabled was consid-ered for further estimation of variability.

In addition to the comparisons with manual align-ment, we performed an objective validation of the auto-matic alignment algorithm by simulating motion (mis-aligning studies) in single-breath-hold studies (9) thatwere available for a subset of patients. Among the 20patients considered previously, three patients with sin-gle-breath-hold acquisitions with no visually observ-able registration errors were considered for this evalu-ation. The single-breath-hold acquisitions consisted offive slices (three SA, one 2CH, and one 4CH) per patient(9). Since in these patients there was no observableshift, the initial misalignments were considered to bezero in both the x and y directions. Series of knowndisplacements were then artificially introduced in se-quence, and automatic correction was applied to cor-rect for these shifts. The 2CH plane was shifted usingall combinations of 0-6 mm range for x and y in 2-mmincrements, for each of the three patients, resulting in atotal of 147 (7 � 7 � 3) data sets. Similarly, each of thethree SA slices was shifted using all combinations of0–12-mm shifts for x and y (2-mm increments in bothdirection), resulting in a total of 507 (13 � 13 � 3) datasets per slice. The choice of range for these shifts wasbased on the observed actual shifts in the manual com-parison case.

3D Alignment of Cine Cardiac MR 967

Statistical Analysis

All continuous variables were expressed as the mean �SD. Registration variability was defined as the differ-ence between the observers or between the observersand the automatic program in determining a given pa-rameter. Registration accuracy was defined as the dif-ference between the average of x and y of shifts asdetermined by the two observers (gold standard) andthe automatic finding. We compared the registrationvariability and accuracy using analysis of variance(ANOVA) with Bonferroni correction and analysis ofcontrasts (comparison of means). The x and y parame-ters were compared separately without Bonferroni cor-rection. All statistical tests were two-tailed, and a valueof P � 0.05 was considered significant. The ANOVA wascalculated with the use of the Analyze-It statisticalpackage (version 1.71; Analyze-it Software Ltd., Leeds,UK).

RESULTS

We registered the image data sets using the automaticand manual techniques for all patients. The automaticprogram was completely unsupervised. The averageprocessing time was 2 seconds for the 2CH-4CH align-ment, and 2 to 3 seconds for the alignment of each SAplane (approximately 20 seconds per case on a 3.0-GHzPentium Intel Xeon computer). The time required for

the manual alignment was approximately two minutesper slice (14–24 minutes per case).

Figure 2 shows a data set before and after registrationof the 2CH and 4CH planes (step 1 of the automaticregistration, as outlined in Fig. 1). Prior to registrationboth ED and end-systolic (ES) images show a noticeablemisalignment at the apex between 2CH and 4CH views,and following registration this error is reduced in bothphases. In Fig. 3 we show a data set before and after SAregistration (step 2 of the automatic registration, asoutlined in Fig. 1). Both ED and ES phases are shown.There is a clear misalignment of the original SA viewwith the 2CH and 4CH views, which is eliminated by theautomatic correction.

Figure 4 shows a 3D surface of the LV model fitted tothe collection of the SA planes. The 3D model was ob-tained twice (once from the original planes and oncefrom the coregistered planes). A visual inspection re-veals evident spatial inhomogeneities in both endo- andepicardial surfaces in the model derived from the orig-inal images, which are not present in the model derivedfrom the coregistered images.

With a significant time investment, each observer wasable to manually obtain misalignment data for each ofthe slices. Figure 5a shows the average displacementsbetween the original planes acquired at differentbreath-holds, as identified by the two observers in all20 cases. SA images were grouped into three regions(basal, mid, and apical), with a similar number of slicesin each region. Note that the total number of slicesvaried between 6 and 11, and therefore the number ofslices in each group varied between two and four. Mo-tion was identified in all patients in all regions by at

Figure 2. Example of the 2CH-4CH plane alignment. Imagesare shown before (top) and after (bottom) automated 3D align-ment using an orthogonal 3D display with 2CH (green outline)and 4CH (blue outline) planes.

Figure 3. Example of the SA to 2CH-4CH plane alignment.Images before (top) and after (bottom) automated 3D align-ment are shown using an orthogonal 3D display with SA (redoutline), 2CH (green outline), and 4CH (blue outline) planes.

968 Slomka et al.

least one of the observers. At least one of the observersidentified motion of �5 mm in 15 studies (75%) and �7mm in six studies (30%). All patients required a correc-tion of �3 mm on at least one of the image planes. Onthe 2CH images, 11 of 20 patients required some cor-rection as determined by the first observer, and 12 of 20required some correction as determined by the secondobserver. Overall, 34% of all considered image planesneeded some manual correction as determined by bothobservers. The maximum misregistration identified bythe visual observers was 7 mm for the CH2/CH4 im-ages and 14 mm for the SA images. The average planeidentified displacement was 2.8 � 1.8 mm (range �0–14 mm). No visually obvious tilts of the SA planeswere identified, and the observers did not attempt toadjust the tilts of the SA planes. The mean difference inadjustments between two observers was 0.2 � 2.2 mm(P � NS) for all corrections.

In Fig. 5b and c we compare the automated alignmentwith the manual alignment performed by the observers.Differences between the algorithm and the average ofthe two observers are shown for 2CH-4CH alignmentand for each SA slice group (apical, mid, basal), similarto Fig. 5a. The results of the automated registrationwith and without 3D rotation enabled for the SA planewere compared with visual readings. The overall accu-racy of the algorithm for the 2CH-4CH registration was1.0 � 0.9 mm. For the SA registration the registrationaccuracy was 2.0 � 1.8, 1.1 � 1.0, and 1.1 � 1.2 mm inapical, midventricular, basal regions, respectively. Theaccuracy differed between the x and y directions only inthe basal plane (P � 0.05). The accuracy of the auto-matic registration was lower in the apical region than inthe other regions (ANOVA, P � 0.05) with or withoutrotation enabled. The largest SA plane tilt found by thealgorithm was 5°. It should be noted that it is not en-tirely correct to compare the manual shifts performedwithout angle correction to the automated shifts withrotation enabled. However, since the rotation angles

found by the algorithm were less than 5°, they have anegligible effect on x, y shifts in the SA plane. We per-formed this comparison to test whether the automaticshift adjustment would remain reliable while the angleadjustment was enabled.

Figure 6 presents the results of an accuracy test per-formed on single-breath-hold studies with a range ofsimulated misalignments. In general these errors arecomparable to the results obtained in multi-breath-hold studies (Fig. 5).

In Fig. 7 we compare the interobserver and algorithm-observer variabilities. The differences between the twoobservers, and between each observer and the auto-matic program are presented for the 2CH-4CH align-ment and for each SA slice group (apical, mid, andbasal). The interuser variability and the algorithm-uservariabilities were similar and ranged from 0.9 mm to1.7 mm. These variabilities were comparable to thepixel size (1.4–1.8 mm), with the exception of the apicalregion, where the average variability (2.0 � 1.9 mm) washigher compared to the mid (1.4 � 1.4 mm) and basal(1.2 � 1.2 mm) regions (ANOVA, P � 0.05). This findingis consistent with the accuracy results (Figs. 5 and 6).The variability of the algorithm (algorithm vs. humanobservers) and interobserver variability were generallycomparable (ANOVA, P � NS), with one exception of thex parameter in the apical region, where the algorithmvs. observer 2 variability was higher than the interob-server variability (2.5 mm vs. 1.8 mm, ANOVA, P �0.05).

DISCUSSION

Clinical cine CMR acquisitions routinely employ onebreath-hold per slice, necessitating multiple breath-holds for each patient data set. The data from thepresent study suggest that 1) errors in spatial align-ment are common, 2) these errors result in artifacts ofthe 3D heart model when reconstructed in 3D imagespace, and 3) these errors can be corrected by the au-tomated registration algorithm presented herein.

The registration algorithm is completely unsuper-vised and involves a short computational time (�20seconds); therefore, it can be easily implemented as astandard postacquisition step without complicating theclinical protocols. Misalignments were present in all ofthe studied patients (on at least one slice) and in at least34% of all image slices. Furthermore, the accuracy ofthe algorithm (1.0–1.5 mm) was smaller than the aver-age displacements identified by the users (2.5–3.5 mm),which may warrant its application on every data set.The accuracy of the algorithm was lower in the apicalregion (2 mm) due to reduced image quality (Fig. 6);however, this accuracy is still comparable to the pixelsize (1.4–1.8 mm). The accuracy of the 2CH-4CH reg-istration was not significantly different from the accu-racy of SA plane registration, even though less informa-tion was available (i.e., one reference view for 2CH-4CHregistration).

In general, the algorithm-observer variability wascomparable to the interobserver variability and was on

Figure 4. LV 3D model fitted to the collection of 2CH/4CH/SAimages before (top) and after (bottom) automated alignment.Endocardial surfaces (ENDO) are shown on the left. Endocar-dial surfaces combined with epicardial surfaces (EPI) areshown on the right. [Color figure can be viewed in the onlineissue, which is available at www.interscience.wiley.com.]

3D Alignment of Cine Cardiac MR 969

the order of 1.0–1.5 mm. However, some deteriorationof the variability in the apical region was evident. Therewas also slightly higher (albeit not statistically signifi-cant) interobserver variability in the apical region,which reflects the general difficulty of determining thecorrect alignment parameters in this region due to sub-optimal image quality (10).

The extent of misregistration of breath-hold posi-tions between acquired slices due to respiratory mo-tion has been quantified, analyzed, and modeled inprevious studies (11–13). It was concluded that thetranslational correction required is significantly

higher than the corrections required for other 3Dlinear transformations (i.e., rotation, scaling, andshear) (12). However, correction involving a combina-tion of other transformations may be necessary tofurther reduce the errors (12). Furthermore, respira-tory motion is usually combined with bulk patientmotion, which is predominantly translational. Al-though the automatic registration algorithm we havedesigned applies translations and/or rotations only,it performs this on a slice-by-slice basis. Therefore,the algorithm has some ability to correct for morecomplex motions. It can, for example, approximate

Figure 5. Initial misalignment and accuracyof the automated algorithm (defined as the dif-ference between the visual and computeralignments). a: Average of absolute plane dis-placements in the x and y directions as iden-tified by two observers in all cases between2CH and 4CH slices, and between SA slicesand LA slices in apical, mid, and basal areas ofthe LV. b: Accuracy of alignment without SAplane rotation. c: Accuracy of alignment withSA plane rotation enabled in the x and y direc-tions. *ANOVA: (a) apical vs. mid and basal; (b)apical vs. 2CH, mid, and basal; P � 0.05.

970 Slomka et al.

rotation of the LA by SA shifts since it is only appliedto one slice at a time.

The proposed algorithm has certain limitations. Therelative registration of 2CH and 4CH planes does notallow for rotation, due to the lack of sufficient imagedata in these orientations. One could remedy this byincreasing the number of 2CH or 4CH views. Rotationabout the intersection of the LA views (and hence cor-rection for cardiac torsion, etc.) is not applied. Again,one could remedy this by increasing the number of 2CHor 4CH views. Furthermore, we retain the equidistantseparation of the slices and do not allow for the trans-lational SA motion in the direction of the LA of theheart. This is done for practical reasons, since the SAslice position along the LA of the heart is less welldefined (compared to the 2CH and 4CH views) than theposition in the other two directions due to the lack ofanatomical detail. Our algorithm by no means ad-dresses all possible types of misregistration; rather, itfocuses on the most obvious errors to deliver a robustsolution for practical purposes. We only addressed themisregistration errors that can be detected by visualinspection. It may be possible to design a more compre-hensive registration algorithm that can address otherforms of errors.

By bringing 2D planes acquired in multiple orienta-tions into common 3D coordinates, the algorithm al-lows 3D analysis and visualization of CMR data. 3Dvisualization can be achieved in the form of multiplanardisplays (Figs. 2 and 3) or perhaps more usefully asextracted 3D surfaces (Fig. 4). Although 3D visualiza-tion and processing have been widely used in nuclearcardiology (14), to date they are not routinely used inCMR. The 3D CMR surface display (such as that shownin Fig. 4) potentially could be used to provide an overall3D review of function, and enable parametric mappingof additional quantitative information (e.g., perfusionand delayed enhancement) similar to nuclear cardiol-ogy methods. The multiplanar display (Figs. 2 and 3)may allow an integrated evaluation of motion and thick-ening with the use of all available images (SA, 2CH, and4CH) simultaneously in regions that are not coveredadequately by one orientation, such as the apex orbase. 3D displays could also aid in the evaluation ofcomplex congenital abnormalities (15) by allowing thespatial features of the whole heart to be examined si-multaneously.

The traditional software techniques (which are avail-able on most commercial workstations) used to quan-tify function from CMR images derive the 2D contours

Figure 6. Accuracy of computer align-ment tested by simulated misalignmentin single-breath-hold studies for the2CH and SA planes in apical, mid, andbasal regions.

Figure 7. Variability of the alignmentbetween the automatic program (Auto)and each observer (Obs1 and Obs2).Variabilities in the x (plain) and y (dot-ted) directions are shown separately.*ANOVA: apical vs. mid, basal, P � 0.05.§ANOVA: Auto vs. Obs1 and Obs2, P �0.05.

3D Alignment of Cine Cardiac MR 971

from each SA slice separately, and subsequently applySimpson’s integration to calculate the 3D parameters(16). The quantification parameters obtained by suchtechniques are probably not significantly affected bythe SA slice misregistration (17). Nevertheless, these 2Dtechniques do not allow for the precise 3D definition ofthe apex and mitral valve. However, exact definitions ofthe apex and mitral-valve planes are available on the LAviews, and the definition of the mitral valve plane inparticular could have a substantial impact on the over-all result (18–21). Integrated and coregistered LA viewsobtained by a technique such as that proposed in thisstudy could be used in conjunction with the SA imagesto derive the 3D model (22). Recent quantificationschemes that promise improved automation and accu-racy rely on 3D modeling of CMR data (20,22,23). Theseschemes will require a spatially aligned 3D representa-tion of CMR images as input. Further validation of thepractical value of our technique can be performed inconjunction with these new quantitative techniques.

Previous studies applied registration algorithms toalign dynamic CMR perfusion images (24–28). How-ever, in such applications the same slice in one cardiacphase is imaged during multiple breath-holds, andtherefore sequential images can be directly registered toeach other by standard techniques. In our application,the 2D cine images to be registered are acquired atdifferent positions and orientations. Lotjonen et al (21)recently developed a multi-breath-hold cardiac motion-correction algorithm to correct for misalignment be-tween SA and LA images; however, they utilized multi-ple 4CH and SA images and a different algorithm basedon volume transformation and normalized mutual in-formation for a subset of cine phases. In their validationby simulated motion, they found a level of accuracysimilar to our results (compared to the visual “goldstandard,” 1 mm), and a similar deterioration of accu-racy in the apex. Software solutions for registering mul-tiple slices to common 3D coordinates have also beenproposed for functional MRI (fMRI) applications (29).

Self-gating and/or navigator-echo acquisition tech-niques have been developed to account for temporalvariations in heart position in a single scan (coronaryimaging) (30–35). However, the technique proposedherein corrects temporally separated studies acquiredin different orientations into one 3D spatially correcteddata set. In a related development, Larson et al (5,36)proposed a self-gating technique that was aimed ateliminating ECG gating, but potentially could be ap-plied to interslice motion correction if an image-pro-cessing technique were available to separate respira-tory and cardiac motion. Since self-gating requires alonger overall imaging time, patient motion could be ofeven greater concern. It may be possible to combine asoftware technique similar to that described in ourstudy with the self-gating approach.

Ideally, a 3D acquisition of all cardiac slices in onebreath-hold would be the ultimate solution to patientmotion. To that end, several techniques have been de-veloped for single-breath-hold, cine volumetric imagingof the heart by CMR (6,37–39). These fast acquisitiontechniques could obviate the need for software correc-tion as described in this work. However, currently the

quality of such imaging remains inferior to thatachieved by standard techniques, and the multi-breath-hold technique remains an established clinicalstandard for the evaluation of cardiac function. Fur-thermore, the heart can move even during a singlebreath-hold (11), and therefore spatial registrationtechniques may still be applicable.

The presented method uses only plane-intersectionprofiles, which contain small subsets of image data, forthe registration. However, to offset this disadvantage,all cine phases are considered, which increases thenumber of correlating samples. Another limitation isthe lack of a gold standard for the tilt adjustment of theSA plane. We have found it difficult and unreliable toadjust these parameters manually, and consequentlywe could not reliably judge the possible improvementby tilt alignment. Nevertheless, since the obtained angleadjustments were small (�5°) and the users did notidentify obvious tilt errors, one could probably disablethe rotational adjustment for practical clinical applica-tions. However, this study demonstrates the technicalfeasibility of performing an out-of-plane 2D SA plane-alignment correction if the need arises.

In conclusion, plane misregistration in multiplanar,multi-breath-hold cine CMR images due to patient mo-tion can be accurately and rapidly corrected by meansof a completely automated software algorithm, result-ing in improved 3D spatial representation of cine car-diac images.

REFERENCES

1. Sakuma H, Fujita N, Foo T, et al. Evaluation of left ventricularvolume and mass with breath-hold cine MR imaging. Radiology1993;188:377–380.

2. Pattynama PM, De Roos A, Van der Wall EE, Van Voorthuisen AE.Evaluation of cardiac function with magnetic resonance imaging.Am Heart J 1994;128:595–607.

3. Ioannidis JPA, Trikalinos TA, Danias PG. Electrocardiogram-gatedsingle-photon emission computed tomography versus cardiac mag-netic resonance imaging for the assessment of left ventricular vol-umes and ejection fraction: a meta-analysis. J Am Coll Cardiol2002;39:2059–2068.

4. Pettigrew RI, Oshinski JN, Chatzimavroudis G, Dixon WT. MRItechniques for cardiovascular imaging. J Magn Reson Imaging1999;10:590–601.

5. Larson AC, White RD, Laub G, McVeigh ER, Li D, Simonetti OP.Self-gated cardiac cine MRI. Magn Reson Med 2004;51:93–102.

6. Kaji S, Yang PC, Kerr AB, et al. Rapid evaluation of left ventricularvolume and mass without breath-holding using real-time interac-tive cardiac magnetic resonance imaging system. J Am Coll Cardiol2001;38:527–533.

7. Simonetti OP, Kim RJ, Fieno DS, et al. An improved MR imagingtechnique for the visualization of myocardial infarction. Radiology2001;218:215–223.

8. Nelder JA, Mead R. A simplex method for function minimization.Comput J 1965:7:308–313.

9. Fieno DS, Thomson LE, Slomka PJ, et al. Rapid assessment of leftventricular segmental wall motion, ejection fraction, and volumeswith single breath-hold, multi-slice TrueFISP MR imaging. J Car-diovasc Magn Reson 2006;8:435–444.

10. Francois CJ, Fieno DS, Shors SM, Finn JP. Left ventricular mass:manual and automatic segmentation of true FISP and FLASH cineMR images in dogs and pigs. Radiology 2004;230:389–395.

11. Holland AE, Goldfarb JW, Edelman RR. Diaphragmatic and cardiacmotion during suspended breathing: preliminary experience and im-plications for breath-hold MR imaging. Radiology 1998;209:483–489.

972 Slomka et al.

12. Manke D, Nehrke K, Bornert P, Rosch P, Dossel O. Respiratorymotion in coronary magnetic resonance angiography: a compari-son of different motion models. J Magn Reson Imaging 2002;15:661–671.

13. Shechter G, Ozturk C, Resar JR, McVeigh ER. Respiratory motionof the heart from free breathing coronary angiograms. IEEE TransMed Imaging 2004;23:1046–1056.

14. Germano G, Kiat H, Kavanagh PB, et al. Automatic quantificationof ejection fraction from gated myocardial perfusion SPECT. J NuclMed 1995;36:2138–2147.

15. Razavi RS, Hill DL, Muthurangu V, et al. Three-dimensional mag-netic resonance imaging of congenital cardiac anomalies. CardiolYoung 2003;13:461–465.

16. van der Geest RJ, Lelieveldt BP, Reiber JH. Quantification of globaland regional ventricular function in cardiac magnetic resonanceimaging. Top Magn Reson Imaging 2000;11:348–358.

17. Danilouchkine MG, van der Geest RJ, Westenberg JJ, LelieveldtBP, Reiber JH. Influence of positional and angular variation ofautomatically planned short-axis stacks on quantification of leftventricular dimensions and function with cardiovascular magneticresonance. J Magn Reson Imaging 2005;22:754–764.

18. Pattynama PM, Doornbos J, Hermans J, van der Wall EE, de RoosA. Magnetic resonance evaluation of regional left ventricular func-tion. Effect of through-plane motion. Invest Radiol 1992;27:681–685.

19. Marcus JT, Gotte MJ, DeWaal LK, et al. The influence of through-plane motion on left ventricular volumes measured by magneticresonance imaging: implications for image acquisition and analy-sis. J Cardiovasc Magn Reson 1999;1:1–6.

20. Swingen CM, Seethamraju RT, Jerosch-Herold M. Feedback-as-sisted three-dimensional reconstruction of the left ventricle withMRI. J Magn Reson Imaging 2003;17:528–537.

21. Lotjonen J, Pollari M, Kivisto S, Lauerma K. Correction of motionartifacts from cardiac cine magnetic resonance images. Acad Radiol2005;12:1273–1284.

22. Young AA, Cowan BR, Thrupp SF, Hedley WJ, Dell’Italia LJ. Leftventricular mass and volume: fast calculation with guide-pointmodeling on MR images. Radiology 2000;216:597–602.

23. Mitchell SC, Bosch JG, Lelieveldt BP, van der Geest RJ, Reiber JH,Sonka M. 3-D active appearance models: segmentation of cardiacMR and ultrasound images. IEEE Trans Med Imaging 2002;21:1167–1178.

24. Delzescaux T, Frouin F, De Cesare A, et al. Adaptive and self-evaluating registration method for myocardial perfusion assess-ment. MAGMA 2001;13:28–39.

25. Bidaut LM, Vallee JP. Automated registration of dynamic MR im-ages for the quantification of myocardial perfusion. J Magn ResonImaging 2001;13:648–655.

26. Gallippi CM, Kramer CM, Hu YL, Vido DA, Reichek N, Rogers WJ.Fully automated registration and warping of contrast-enhancedfirst-pass perfusion images. J Cardiovasc Magn Reson 2002;4:459–469.

27. Comte A, Lalande A, Aho S, Walker PM, Brunotte F. Realignment ofmyocardial first-pass MR perfusion images using an automaticdetection of the heart–lung interface. Magn Reson Imaging 2004;22:1001–1009.

28. Stegmann MB, Olafsdottir H, Larsson HB. Unsupervised motion-compensation of multi-slice cardiac perfusion MRI. Med ImageAnal 2005:9:394–410.

29. Kim B, Boes JL, Bland PH, Chenevert TL, Meyer CR. Motion cor-rection in fMRI via registration of individual slices into an anatom-ical volume. Magn Reson Med 1999;41:964–972.

30. Danias PG, McConnell MV, Khasgiwala VC, Chuang ML, EdelmanRR, Manning WJ. Prospective navigator correction of image posi-tion for coronary MR angiography. Radiology 1997;203:733–736.

31. Wang Y, Grimm RC, Felmlee JP, Riederer SJ, Ehman RL. Algo-rithms for extracting motion information from navigator echoes.Magn Reson Med 1996;36:117–123.

32. Wang Y, Riederer SJ, Ehman RL. Respiratory motion of the heart:kinematics and the implications for the spatial resolution in coro-nary imaging. Magn Reson Med 1995;33:713–719.

33. Nguyen TD, Nuval A, Mulukutla S, Wang Y. Direct monitoring ofcoronary artery motion with cardiac fat navigator echoes. MagnReson Med 2003;50:235–241.

34. Wang Y, Rossman P, Grimm R, Riederer S, Ehman R. Navigator-echo-based real-time respiratory gating and triggering for reduc-tion of respiration effects in three-dimensional coronary MR an-giography. Radiology 1996;198:55–60.

35. Shea SM, Kroeker RM, Deshpande V, et al. Coronary artery imag-ing: 3D segmented k-space data acquisition with multiple breath-holds and real-time slab following. J Magn Reson Imaging 2001;13:301–307.

36. Larson AC, Kellman P, Arai A, et al. Preliminary investigation ofrespiratory self-gating for free-breathing segmented cine MRI.Magn Reson Med 2005;53:159–168.

37. Lee VS, Resnick D, Bundy JM, Simonetti OP, Lee P, Weinreb JC.Cardiac function: MR evaluation in one breath hold with real-timetrue fast imaging with steady-state precession. Radiology 2002;222:835–842.

38. Kozerke S, Tsao J, Razavi R, Boesiger P. Accelerating cardiac cine3D imaging using k-t BLAST. Magn Reson Med 2004;52:19–26.

39. Narayan G, Nayak K, Pauly J, Hu B. Single-breathhold, four-di-mensional, quantitative assessment of LV and RV function usingtriggered, real-time, steady-state free precession MRI in heart fail-ure patients. J Magn Reson Imaging 2005;22:59–66.

3D Alignment of Cine Cardiac MR 973