Quantification of Myocardial Perfusion Using Free-Breathing MRI and Prospective Slice Tracking

Henrik Pedersen,1,2,* Sebastian Kelle,3 Steffen Ringgaard,1 Bernhard Schnackenburg,3

Eike Nagel,5 Kay Nehrke,4 and Won Yong Kim1

Quantification of myocardial perfusion using first-pass mag-netic resonance imaging (MRI) is hampered by respiratory mo-tion of the heart. Prospective slice tracking (PST) potentiallyovercomes this problem, and may provide an attractive alter-native or supplement to current breath-hold techniques. Thisstudy demonstrates the feasibility of patient-adapted 3D PSTon a 3.0 Tesla MR system. Eight patients underwent free-breathing studies of myocardial perfusion, simultaneously col-lecting data with and without PST. On average, PST reducedresidual in-plane motion by a factor of 2, compared to thenoncorrected images, resulting in a fourfold improvement ofperfusion measurements. In addition, a comparison of perfu-sion measurements performed with and without PST showedthat through-plane motion can contaminate measurements ofmyocardial perfusion. However, the quality of the navigatorechoes on this field strength constituted a major source oferror and needs further improvement to increase the accu-racy and robustness of the method. Magn Reson Med 61:734 –738, 2009. © 2008 Wiley-Liss, Inc.

Key words: myocardial perfusion, respiratory motion, prospec-tive slice tracking, navigator echoes

Quantification of myocardial perfusion using first-passmagnetic resonance imaging (MRI) is often hampered byrespiratory motion of the heart, thus limiting the currentclinical potential of the method. Breath-holding representsthe simplest and most widely used strategy to compensaterespiratory motion (1), but in cases where the patient is notable to properly hold his/her breath, or when there is aspatial misalignment between images acquired at rest andstress, additional realignment of the images is required.The process of aligning the images can to some extent beautomated (2–5), but reliable and accurate motion detec-tion remains difficult before and during bolus arrival. Fur-thermore, misalignments along the through-plane (TP) di-rection cannot be corrected retrospectively with current2D multislice techniques.

A fundamentally different approach, which shares thesimplicity of breath-holding and potentially overcomesthe limitations of retrospective correction, is prospectiveslice tracking (PST) (6,7). In its simplest form, PST shiftsthe imaging slice according to the respiratory motion of theheart, thereby reducing slice misalignments in both thein-plane (IP) and TP direction. PST may permit free-breathing data acquisition for the benefit of the patient, butcan also be combined with breath-hold imaging to dealwith slice misalignments between the rest and stress ex-aminations or to compensate partial or insufficient breath-holding.

Prospective slice tracking is widely used in coronary MRangiography (CMRA) to increase the size of the respiratorygating window or improve overall image quality (8–10), butthe method has not previously been used in myocardialperfusion MRI. There are two important differences betweenapplying PST in CMRA and in myocardial perfusion imag-ing. First, respiratory gating is not possible in myocardialperfusion MRI as it decreases the temporal resolution. Thisimplies that the underlying motion model must work ade-quately over the entire diaphragmatic motion range. Second,image acquisition in CMRA is restricted to mid-diastole,whereas perfusion images are typically acquired throughoutthe entire cardiac cycle. Therefore, to deal with respiratorydisplacements occurring within each cardiac cycle, repeatedmeasurements of the diaphragmatic positions are requiredwithin each heart beat.

The purpose of this study was to develop a method ofprospective slice tracking feasible for clinical MRI studiesof myocardial perfusion and to determine the accuracy ofperfusion measurements using free-breathing MRI andPST. In addition, we demonstrate for the first time theinfluence of TP motion on myocardial perfusion measure-ments. To facilitate respiratory motion correction over theentire diaphragmatic motion range and to take into ac-count the well-known subject variability (11,12), we useda patient-calibrated motion model of 3D translation incor-porating modeling of temporal correlations (13,14). Thecalibration was performed automatically using an addi-tional low-resolution prescan depicting the respiratorymotion of the heart. The image data of the prescan alsoallowed us to simulate the expected performance of alter-native motion models. For comparison, we simulated theperformance of the patient-calibrated model used in theperfusion studies and the conventional Wang model usedin CMRA (12), which includes correction of 1D translationalong the feet-to-head direction using a fixed empiricalcorrection factor of 0.6 for all patients.

1MR Research Centre, Aarhus University Hospital Skejby, Aarhus, Denmark;2Functional Imaging Unit, Glostrup Hospital, Copenhagen, Denmark; 3GermanHeart Institute, Berlin, Germany; 4Philips Research Laboratories, Hamburg,Germany; 5King’s College London, Division of Imaging Sciences, London, UK.Grant sponsor: Philips Medical Systems; Grant sponsor: University of AarhusGraduate School of Health Sciences; Grant sponsor: Danish Agency forScience, Technology and Innovation.*Correspondence to: Henrik Pedersen, Functional Imaging Unit, GlostrupHospital University of Copenhagen, Ndr. Ringvej 57, 2600 Glostrup, Copen-hagen, Denmark. E-mail: klaverhenrik@gmail.comReceived April 18, 2008; revised September 4, 2008; accepted October 9,2008.DOI 10.1002/mrm.21880Published online 18 December 2008 in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 61:734–738 (2009)

© 2008 Wiley-Liss, Inc. 734

METHODS

Rest and stress myocardial perfusion studies were per-formed in a total of eight patients with suspected coronaryartery disease undergoing clinical routine MRI. Informedconsent was obtained before imaging. All patients wereexamined on a 3.0 Tesla MR scanner (Gyroscan Achieva,Philips Healthcare, The Netherlands) equipped with a six-element cardiac receive coil. The prescan required for themodel calibration was performed initially, and subse-quently free-breathing perfusion studies were conductedwith and without PST, starting with the stress examina-tion.

Model Calibration and Simulations

For the model calibration we first aligned a 3D sagittalimaging stack and a shim volume over the heart and po-sitioned a navigator beam on the right hemidiaphragm.The respiratory motion of the heart was then imaged for atotal of 30 cardiac cycles using an ECG-triggered, fast gra-dient echo sequence (flip angle 25°, pulse repetition time[TR] 2.53 ms, field of view [FOV] 420 � 370 � 100 mm3,matrix 112 � 112 � 14, slice thickness 8 mm, SENSE � 2,T2 preparation pulse). The motion model was calibrated byregistering and correlating the motion of the heart to thenavigator positions using linear regression, as described byManke et al. (13). For the registration, the size and positionof the heart was automatically adapted from the manuallydefined shim volume (14). To take into account hystereticeffects, a 3D translational motion model related to twotemporally separated navigators was used. The two navi-gators were separated in time by 150 ms, which allowedmodeling of hysteretic effects (see Fig. 1). All necessarycalculations to complete the calibration were imple-mented on-line on the scanner, and the entire calibrationprocedure was completed in less than 2 min.

As in Ref. (13), we used the 3D images of the calibrationscan to simulate the expected performance of differentmotion models. Using a 3D affine registration, the root-mean-square (RMS) error of the mean displacement of thevoxels in each short-axis slice was calculated for the threemain orientations of the patient coordinate system: ante-rior-to-posterior (AP), left-to-right (RL), and feet-to-head(FH). Also, the RMS error was calculated for the IP and TPdirections of each short-axis slice. The RMS errors were

calculated for the following motion models: no correction,the Wang model (12), and the patient-calibrated model of3D translation.

Perfusion Studies

The calibrated motion model was used for performing PSTin free-breathing myocardial perfusion studies. Contrastagent was administered intravenously using a bolus of0.025 mmol/kg of gadopentetate dimeglumine (Magnevist;Schering, Berlin, Germany) at an injection rate of 4 mL/sfollowed by a flush of 20 mL of saline solution. For thestress examination, adenosine was used with an injectionrate of 140 �g/kg/min. The first pass of the contrast agentwas imaged in three short-axis slices using an ECG-trig-gered, spoiled gradient echo sequence (flip angle 18°, TR2.53 ms, FOV 420 � 370 mm2, matrix 144 � 136, slicethickness 8 mm, SENSE � 2, nonselective 90° saturationrecovery [SR] prepulse). The surface coil sensitivities werecorrected using the body coil reference map used for theSENSE reconstruction. The navigator echoes were evenlydistributed 150 ms apart, as shown in Figure 1a, in order todeal with respiratory motion occurring within each car-diac cycle. The timing of the imaging sequence was de-signed such that, for any of the three slices, the two im-mediately preceding navigators were used for motion pre-diction (as required by the underlying motion model). Toacquire images with and without PST under the samephysiologic and injection conditions, motion correctionwas performed in an interleaved fashion, such that PSTwas performed only in the even-numbered cardiac cycles(Fig. 1a). Thus, two data sets were generated for each slicein each experiment. In our experience, the resultant de-crease in temporal resolution (i.e., one frame every secondheart beat) is of insignificant importance compared to thegross influence of respiratory motion.

The IP displacement of the left ventricle was registeredin all frames using a normalized crosscorrelation (NCC)registration method (15). Possible outliers were registeredmanually. We used the RMS error to describe the level ofresidual IP motion in each of the two data sets. In addition,the result of the NCC registration was used to generate a setof motion-free images, which constituted the reference forthe subsequent perfusion measurements.

For each slice, myocardial perfusion was calculated forsix equiangular segments that remained fixed across all

Insp.

NL[mm]

NP[mm]-20

-100

10

-15-10-505100

5

10

15

FH

tra

nsla

tion

[mm

]

Exp.

ECG

MRI

Resp.curve

SR prepulse

Navigator

Image acq.

Odd cardiac cycle

Slice trackingNo correction

1 3 1 31 2 3 1 2 3

150 ms

Even cardiac cycle

NP NL

Insp.

NL[mm]

NP[mm]-20

-100

10

-15-10-505100

5

10

15

FH

tra

nsla

tion

[mm

]

Exp.

ECG

MRI

Resp.curve

SR prepulse

Navigator

Image acq.

Odd cardiac cycle

Slice trackingNo correction

1 3 1 31 2 3 1 2 3

150 ms

Even cardiac cycle

NP NL

a bFIG. 1. Timing of the MRI sequence used for myocardial perfusion imaging (a). Three slices are acquired in each cardiac cycle, each ofwhich is preceded by a saturation recovery (SR) prepulse. A navigator echo is used to measure the diaphragmatic position immediatelybefore imaging and 150 ms before imaging. This strategy captures hysteretic effects, as illustrated for the cardiac translation along thefeet-to-head (FH) direction shown in (b).

Myocardial Perfusion MRI Using Prospective Slice Tracking 735

frames of the perfusion series. Each myocardial signalintensity curve smyo(t) was fitted to a mono-exponentialperfusion model:

smyo�t� � k1sAIF�t� � exp� � �t � td�/k2� [1]

where sAIF(t) is the arterial input function derived from acircular region of interest (ROI) with a diameter of approx-imately 10 mm located in the center of the left ventricularcavity, k1 represents perfusion, k2 is a time constant, and Rdenotes convolution. The baseline signal was subtractedfrom all signal curves prior to fitting Eq. [1], and the timedelay (td) between the arterial input and the myocardialsignal curve was calculated automatically, as proposed byCheong et al. (16).

For each myocardial segment, we used as error measurethe absolute difference of the perfusion between the rawdata and the reference data (i.e., that registered with theNCC method), normalized with the perfusion of the refer-ence data. These errors are given in percentage. The refer-ence perfusion was calculated separately for PST and non-PST. In addition, segment perfusion was measured in thetwo data sets (i.e., those acquired with and without PST)after additional retrospective registration of residual IPtranslation using the NCC method. Provided that the IPmotion is sufficiently corrected in both data sets, any sys-tematic difference between them should reflect the influ-ence of TP motion.

RESULTS

The model calibration and subsequent application of PSTin perfusion studies was successfully applied in all pa-tients, demonstrating the feasibility of the method. Apartfrom the relatively poor quality of the navigator signal insome patients, as discussed in more detail below, we didnot experience any technical problems.

Simulations

Figure 2a shows the simulated results of the calibrationbased on all patients and short-axis slices (n � 24). Withan RMS error of 3.1 � 1.0 mm (mean � SD), the noncor-rected data confirmed that the primary orientation of themotion was along the FH direction, whereas the motion inthe RL (0.7 � 0.3 mm) and AP (0.6 � 0.3 mm) directionswas on average smaller by a factor of 4 to 5. The RMS erroralong the TP direction of the short-axis plane was 1.8 �0.8 mm and 2.4 � 0.8 mm along the IP direction.

The Wang model reduced the motion along the FH di-rection (1.2 � 0.5 mm) by a factor of more than 2, whichhalved the RMS error along the IP (1.0 � 0.4 mm) and TP(0.8 � 0.5 mm) directions. The patient-adapted model of3D translation resulted in a reduction of the RMS errorsexceeding those of the Wang model. The largest reductionwas observed in the IP direction (0.5 � 0.3 mm) with afactor of 4, whereas the FH (0.8 � 0.4 mm) and TH (0.6 �0.4 mm) directions showed a decrease of about a factor ofthree. Paired t-tests showed that these differences relativeto noncorrected data, as well as the differences betweenWang and patient-adapted 3D translation, were statisti-cally significant (P � 0.05). The latter demonstrates thebenefit of the complexity of the calibration procedure.

Perfusion Studies

Figure 3 demonstrates the steps involved in analyzing theperfusion images. In the example shown in the figure, PSTreduced the RMS error of the residual IP motion from3.8 mm to 1.54 mm (Fig. 3b). This resulted in a reductionof the perfusion error from 50% in the noncorrected im-

0,00

0,50

1,00

1,50

2,00

2,50

3,00

3,50

4,00

4,50

RM

S [

mm

]

AP RL FH IPTP

Simulations

a

IP IP

Rest Stress

No corr.

Wang

3D

b

0,00

0,50

1,00

1,50

2,00

2,50

3,00

3,50

4,00

4,50

RM

S [

mm

]

AP RL FH IPTP

Simulations

IP IP

Rest Stress

No corr.

Wang

3D

a b

FIG. 2. Mean RMS errors of all patients indicating the residualrespiratory motion of the simulated data (a) and the perfusion data(b). For the simulations the RMS errors were calculated for thefollowing directions: anterior-to-posterior (AP), right-to-left (RL),feet-to-head (FH), in-plane (IP), and through-plane (TP). The RMSerrors of the perfusion studies could only be calculated for the IPcomponent.

FIG. 3. Data analysis of a representative perfusion data set. Thesegmentation of the myocardium is shown in (a), and the registeredIP displacements of the left ventricle for each image frame areplotted in (b) with and without prospective slice tracking. The orig-inal signal curve of the marked myocardial segment is depicted in (c)and (d) with and without PST (solid black lines). Also shown are thereference curves obtained after additional retrospective registration(dashed black lines).

736 Pedersen et al.

ages (Fig. 3c) to 10% in the prospectively corrected images(Fig. 3d).

The mean RMS errors of the perfusion studies for the IPcomponent are shown in Figure 2b based on all patientsand short-axis slices (n � 24). Without correction, themean RMS error was 2.8 � 1.0 mm at rest and 2.8 �0.7 mm during stress. With PST, these errors were signif-icantly reduced both during rest (1.3 � 0.4 mm, P � 0.05)and stress (1.5 � 0.5 mm, P � 0.05). The improvement ofthe PST correction during rest compared to stress wasstatistically significant, suggesting that the increased heartrate and changes in respiratory motion pattern render PSTmore difficult during stress. Although there was a differ-ence in the performance of PST for rest and stress, we didnot observe any noticeable difference with respect to theperfusion errors. Considering all myocardial segments ofthe rest and stress examinations (n � 288), the mean per-fusion error was reduced from 79 � 254% without PST to20% � 22% with PST (P � 0.05). After additional retro-spective NCC registration of IP translation, the mean abso-lute difference in segment perfusion between PST andnon-PST was 45 � 75%, which deviated significantly fromzero (P � 0.05, t-test).

Figure 4 shows examples of pixel-based perfusion mapsobtained with and without PST. The IP motion of bothdata sets was sufficiently corrected prior to perfusion anal-ysis, such that systematic differences between the twomaps should be due to TP motion. The right-most mapshows the absolute difference in percent between the twoperfusion maps. The arrow indicates a subendocardial re-gion with a systematic difference of about 60%, showingthat TP motion can contaminate perfusion measurements.

DISCUSSION

We have presented our initial experience with prospectiveslice tracking for respiratory motion compensation in clinicalMRI studies of myocardial perfusion in patients. In an at-tempt to optimize the correction, we used a previously de-veloped online tool to determine the patient-specific rela-tionship between the motion of the heart and navigatorslocated on the right diaphragm. This allowed adapting themethod to each individual patient, but also we were able tosimulate the expected performance of different motion mod-els. In this work, we presented experimental verification ofPST using a dedicated dual data acquisition scheme collect-ing corrected and noncorrected perfusion data in the samescan. This permitted us not only to compare the degree ofresidual IP motion of the two data sets, but also to comparethe accuracy of perfusion measurements on a segment-by-segment level (using the manually corrected images as refer-ence). Furthermore, the dual data acquisition scheme al-lowed us to demonstrate the influence of TP motion onpixel-based perfusion measurements.

IP Motion

The results of the perfusion measurements suggest thatwhen the IP motion is reduced by a factor of 2, the asso-ciated error of the perfusion measurements is decreased bya factor of 4. This asymmetric relationship may be attrib-uted to the nonlinearity of Eq. [1]. However, the standarddeviation of the perfusion error with PST was above 20%,suggesting that some myocardial segments may requirefurther correction for clinical use.

Compared with the results of the simulations, patient-adapted PST performed worse in practice than expected.To a certain extent this difference may be explained bymiscalibration of the underlying motion model, resultingin poor correction in the subsequent perfusion study.However, because the results of our simulations agree withthose reported by Manke et al. (13), a more likely expla-nation is that the navigator signal was corrupted, as ob-served in some patients (see Fig. 5), thereby reducing theaccuracy of the correction. The origin of the poor navigatorquality is most likely a combination of the field inhomo-geneities on 3 Tesla, which are known to corrupt thedesired excitation profile of the navigator (17), and the lowsignal-to-noise ratio induced by the repeatedly applied SRpulses. Potential ways to overcome this problems includeincreasing the diameter of the navigator beam or using aspatially selective SR pulse that covers the heart withoutspatially interfering with the navigator (18). Further stud-ies are required to determine the optimum solution.

TP Motion

There is currently no way to retrospectively measure andcorrect TP motion in 2D multislice techniques due to rela-tively large interslice gaps and slice thicknesses of 8–10 mm.However, our simulations suggest that the TP component(with respect to the short-axis plane) is so large that it must betaken into account. With an associated mean RMS error of1.75 mm, the peak-to-peak amplitude, which is usually fourtimes larger than the RMS error, approaches the slice thick-ness. At this point the theory of tracer kinetic modeling is

FIG. 4. Pixel-based perfusion maps obtained after additional retro-spective in-plane registration of data obtained with and withoutprospective slice tracking (a–b). The relative difference between thetwo maps, which is likely due to through-plane motion, is shown in(c). The difference approaches 60% in the subendocardial regionmarked with an arrow.

Myocardial Perfusion MRI Using Prospective Slice Tracking 737

potentially invalid, depending on what other tissue isbrought into the slice by TP motion. The simulations showthat PST can significantly reduce TP motion. As expected,the patient-adapted model of 3D translation should yield thebest performance, but even the 1D Wang model should re-duce the TP component by a factor of 2. More advancedmodels, which could take into account nonrigid motion (13),such as stretching and compression of the heart, were notconsidered in the present study.

In perfusion measurements, we demonstrated a significantdifference between segments evaluated with and withoutPST after additional IP motion correction. This differencemay in part be due to TP motion, which is supported by thepixel-based perfusion maps of Figure 4, showing a subendo-cardial perfusion difference, although the IP correction wassufficient. This suggests that indeed TP motion can contam-inate measurements of myocardial perfusion.

CONCLUSIONS

Prospective slice tracking forms an attractive alternative ora valuable supplement to breath-hold myocardial perfu-sion MRI used in current clinical practice. The feasibilityof patient-adapted PST has been demonstrated on a 3.0Tesla MR system in a clinical routine setup. In terms ofaccuracy, the quality of the navigator signal on this fieldstrength constitutes the primary source of error and needsfurther improvement to increase the robustness of themethod. Nevertheless, the accuracy of the current imple-

mentation was sufficient to yield a twofold reduction ofresidual IP motion and an associated fourfold reduction ofthe average perfusion error, and to show for the first timethat TP motion does indeed contaminate measurements ofmyocardial perfusion. Whether prospective slice trackingcan be improved to enable free-breathing studies in allpatients, or whether the approach is best combined withbreath-hold imaging to compensate insufficient breath-holding, remains to be investigated.

REFERENCES1. Gebker R, Schwitter J, Fleck E, Nagel E. How we perform myocardial

perfusion with cardiovascular magnetic resonance. J Cardiovasc MagnReson 2007;9:539–547.

2. Bidaut LM, Vallee JP. Automated registration of dynamic MR imagesfor the quantification of myocardial perfusion. J Magn Reson Imaging2001;13:648–655.

3. Dornier C, Ivancevic MK, Thevenaz P, Vallee JP. Improvement in thequantification of myocardial perfusion using an automatic spline-basedregistration algorithm. J Magn Reson Imaging 2003;18:160–168.

4. Gupta SN, Solaiyappan M, Beache GM, Arai AE, Foo TK. Fast methodfor correcting image misregistration due to organ motion in time-seriesMRI data. Magn Reson Med 2003;49:506–514.

5. Adluru G, Dibella EV, Schabel MC. Model-based registration for dy-namic cardiac perfusion MRI. J Magn Reson Imaging 2006;24:1062–1070.

6. Danias PG, McConnell MV, Khasgiwala VC, Chuang ML, Edelman RR,Manning WJ. Prospective navigator correction of image position forcoronary MR angiography. Radiology 1997;203:733–736.

7. McConnell MV, Khasgiwala VC, Savord BJ, Chen MH, Chuang ML,Edelman RR, Manning WJ. Prospective adaptive navigator correctionfor breath-hold MR coronary angiography. Magn Reson Med 1997;37:148–152.

8. Stuber M, Botnar RM, Danias PG, Kissinger KV, Manning WJ. Submil-limeter three-dimensional coronary MR angiography with real-timenavigator correction: comparison of navigator locations. Radiology1999;212:579–587.

9. Jahnke C, Paetsch I, Nehrke K, Schnackenburg B, Gebker R, Fleck E,Nagel E. Rapid and complete coronary arterial tree visualization withmagnetic resonance imaging: feasibility and diagnostic performance.Eur Heart J 2005;26:2313–2319.

10. Nagel E, Bornstedt A, Schnackenburg B, Hug J, Oswald H, Fleck E.Optimization of realtime adaptive navigator correction for 3D magneticresonance coronary angiography. Magn Reson Med 1999;42:408–411.

11. Keegan J, Gatehouse P, Yang GZ, Firmin D. Coronary artery motionwith the respiratory cycle during breath-holding and free-breathing:implications for slice-followed coronary artery imaging. Magn ResonMed 2002;47:476–481.

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

13. Manke D, Nehrke K, Bornert P. Novel prospective respiratory motioncorrection approach for free-breathing coronary MR angiography usinga patient-adapted affine motion model. Magn Reson Med 2003;50:122–131.

14. Nehrke K, Bornert P. Prospective correction of affine motion for arbi-trary MR sequences on a clinical scanner. Magn Reson Med 2005;54:1130–1138.

15. Netsch T, Rosch P, van Muiswinkel A, Weese J. Towards real-timemulti-modality 3D medical image registration. Int Conf Comp VisionICCV 2001;501–508.

16. Cheong LH, Koh TS, Hou Z. An automatic approach for estimatingbolus arrival time in dynamic contrast MRI using piecewise continuousregression models. Phys Med Biol 2003;48:N83–N88.

17. Nehrke K, Bornert P, Groen J, Smink J, Bock JC. On the performanceand accuracy of 2D navigator pulses. Magn Reson Imaging 1999;17:1173–1181.

18. Pedersen H, Ringgaard S, Kim WY. A patient-adapted navigator track-ing approach for prospective 3D respiratory motion correction in free-breathing myocardial perfusion imaging. In: Proc 13th Annual MeetingISMRM, Miami; 2005. p 512.

FIG. 5. Comparison of navigator signals with good quality (a) andpoor quality (b). Each navigator consists of a time series of verticalimage profiles aligned through the right diaphragm. If the lung–liverinterface is clearly depicted the diaphragmatic position (indicated bythe dotted white line) is accurately measured. Clearly, this interfacewas absent in (b), leading to poor measurements of the diaphrag-matic motion and, hence, inaccurate motion correction.

738 Pedersen et al.