4D radial coronary artery imaging within a single breath-hold: Cine angiography with phase-sensitive fat suppression (CAPS)

4D Radial Coronary Artery Imaging Within a SingleBreath-Hold: Cine Angiography With Phase-Sensitive FatSuppression (CAPS)

Jaeseok Park,1,2 Andrew C. Larson,1 Qiang Zhang,3 Orlando Simonetti,3 andDebiao Li1,2*

Coronary artery data acquisition with steady-state free preces-sion (SSFP) is typically performed in a single frame in mid-diastole with a spectrally selective pulse to suppress epicardialfat signal. Data are acquired while the signal approaches steadystate, which may lead to artifacts from the SSFP transientresponse. To avoid sensitivity to cardiac motion, an accuratetrigger delay and data acquisition window must be determined.Cine data acquisition is an alternative approach for resolvingthese limitations. However, it is challenging to use conventionalfat saturation with cine imaging because it interrupts thesteady-state condition. The purpose of this study was to de-velop a 4D coronary artery imaging technique, termed “cineangiography with phase-sensitive fat suppression” (CAPS), thatwould result in high temporal and spatial resolution simulta-neously. A 3D radial stacked k-space was acquired over theentire cardiac cycle and then interleaved with a sliding window.Sensitivity-encoded (SENSE) reconstruction with rescaling wasdeveloped to reduce streak artifact and noise. Phase-sensitiveSSFP was employed for fat suppression using phase detection.Experimental studies were performed on volunteers. The pro-posed technique provides high-resolution coronary artery im-aging for all cardiac phases, and allows multiple images atmid-diastole to be averaged, thus enhancing the signal-to-noise ratio (SNR) and vessel delineation. Magn Reson Med 54:833–840, 2005. © 2005 Wiley-Liss, Inc.

Key words: magnetic resonance imaging (MRI); coronary MRA;parallel MRI; rapid imaging; SSFP; fat suppression

It is difficult to employ coronary magnetic resonance an-giography (MRA) in a clinical setting because of respira-tory and cardiac motion. One can compensate for respira-tory motion by using either a free-breathing navigator echo(1,2) or a breath-hold (3,4) technique. Free-breathing cor-onary MRA with navigator echo has been shown to beuseful for achieving high spatial resolution, a reasonablesignal-to-noise ratio (SNR), and sufficient spatial coverage.However, residual image artifacts may result from the re-spiratory navigation if there is an inconsistent breathingpattern. Breath-hold techniques eliminate sensitivity tothe irregular breathing pattern, but spatial resolution is

limited by the short imaging time. To avoid cardiac mo-tion, data acquisition is limited to mid-diastole. This ne-cessitates the use of a short acquisition window (�100 ms)and an accurate trigger delay.

Given the small size of the coronary arteries, high spatialresolution is required to diagnose disease. To improve thespatial resolution, partially parallel acquisition (PPA)techniques (5,6) have been used to acquire a fraction of thephase-encoding lines as compared to conventional dataacquisition, using an array of simultaneously operated re-ceiver coils. One can remove foldover artifacts resultingfrom undersampling of k-space by exploiting informationregarding the spatial coil sensitivity. Alternatively, non-Cartesian sampling schemes with spiral (7) or radial (8)k-space trajectories can also be used to increase spatialresolution in a fixed imaging time, because the spatialresolution is not traded off with the number of acquiredviews.

Real-time 2D coronary MRA techniques (9,10) obviatethe need to determine an accurate trigger delay and dataacquisition window in coronary artery imaging. Thesemethods also enhance SNR by averaging multiple imagesof the same slice. However, if k-space is highly under-sampled to increase the temporal resolution, aliasing arti-facts may be introduced in the region of interest (ROI) andimpede the depiction of vessels. In addition, it is difficultto delineate tortuous vessels or vessels moving through animage plane using this 2D imaging technique.

The purpose of this study was to implement a breath-hold 3D coronary artery imaging method with both hightemporal and high spatial resolution, spanning multiplephases of the cardiac cycle. This method is termed “cineangiography with phase-sensitive fat suppression” (CAPS).

THEORY

A schematic of the CAPS technique is depicted in Fig. 1.In-plane k-space, kx-ky, is acquired using radial sampling,while kz is acquired using conventional Fourier phaseencoding. To achieve high temporal resolution, radial k-space is highly undersampled azimuthally from 0 to 180°at each cine phase. Each cine phase is composed of eithereven or odd views. The radial k-space in each cine phaseis combined using a factor of 2 interleaved view sharingwith a sliding window. The interleaved radial k-spaceundergoes the following two processes: 1) self-calibratingradial SENSE reconstruction with rescaling method to re-duce conventional streak artifact and noise from under-sampling of radial k-space, and 2) conventional density-compensated gridding reconstruction (11) to make a water

1Department of Radiology, Northwestern University, Chicago, Illinois, USA.2Department of Biomedical Engineering, Northwestern University, Chicago,Illinois, USA.3Siemens Medical Solutions, Chicago, Illinois, USA.*Correspondence to: Debiao Li, Ph.D., 448 East Ontario Street, Suite 700,Chicago, IL 60611. E-mail: [email protected] sponsor: National Institutes of Health; Grant numbers: EB002623;HL38698; Grant sponsor: Siemens Medical Solutions.Received 8 November 2004; revised 25 April 2005; Accepted 25 April 2005.DOI 10.1002/mrm.20627Published online 7 September 2005 in Wiley InterScience (www.interscience.wiley.com).

Magnetic Resonance in Medicine 54:833–840 (2005)

© 2005 Wiley-Liss, Inc. 833

mask image (composed of zero (� fat pixels) and one (�water pixels)) by exploiting the fact that water and fat areout of phase in SSFP within a specific range of TRs (2.5–7.1 ms) (12). The sensitivity-encoded (SENSE) image ismultiplied by the water mask image, yielding a final fat-suppressed image.

Self-Calibrating Radial SENSE Reconstruction Using aRescaling Method

Conventional SENSE reconstruction with a radial k-spacetrajectory (13) is reduced to a large system of linear equa-tions after noise decorrelation:

EHEx � EHm [1]

where E is the encoding matrix, EH is its Hermitian matrix,m is the acquired radial k-space, and x is the reconstructedimage. Equation [1] is solved iteratively using the conju-gate-gradient (CG) method with forward and reverse grid-ding operations between Cartesian and radial k-spaces.However, conventional SENSE is computationally compli-cated, and requires a highly accurate convolution process.In addition, rounding errors in each iteration lead to anincrease in noise

To reduce the complexity of the conventional SENSEoperation, in this new method the CG iteration loop ismodified to incorporate the rescaling method (14) (Fig. 2).The radial k-space is expanded by a scale factor of r (newmatrix size � rN � rN; N � acquired readout matrix size)by rounding a rescaled coordinate off to the nearest recti-linear grid location. To correct for the repeated mapping ofseveral points onto the same coordinate, the mean valuesare calculated. The rescaled k-space map in Fig. 2 is con-structed by setting the located coordinates to be one andall others to zero.

Coil sensitivity (Sn, n � coil index) is calculated usinglow-frequency signals and zero padding in the rescaledk-space matrix. Inverse fast Fourier transformation (IFFT)is performed on the zero-padded, low-frequency k-space ofeach coil, which results in a low-resolution image. Eachcoil image is normalized by the root sum-of-squares (SOS)magnitudes of all the coil images, and coil sensitivity mapsare generated.

In computing the EH, each rescaled coil k-space is pro-cessed by an IFFT. This generates a reconstructed image inthe central N � N region with severe aliasing artifactoutside the central region. An ROI map is constructedusing ones in the central N � N region and zeros outsidethe central region. The ROI map is multiplied to the imagepixel-by-pixel, replacing the outer region by zeros andpreserving the central image region. Each coil image ismultiplied by the complex conjugate of coil sensitivity(Sn*), and then summed coil-wise.

In computing E, a residual image is multiplied by thecoil sensitivity and followed by FFT, resulting in therescaled coil k-space. The coil k-space is updated by mul-tiplying the rescaled k-space map for the next iteration.

Once EHmr (mr � rescaled matrix) in Eq. [1] is initial-ized, the residual image (rN � rN) is multiplied by EHE.During each iteration, rounding errors resulting from therescaling process lead to noise amplification. Hence, it isnecessary to determine the number of iterations (Niter)required to achieve an optimal trade-off between imageaccuracy and noise amplification. Once the CG loop isterminated at the presumed optimal iteration number, thefinal residual image must be cropped such that only theROI (the central N � N image) is left. This is because therescaling process is equivalent to subsampling of mea-sured data, and results in a much larger FOV than the datasampling.

FIG. 1. A schematic of the CAPS technique, with 4D radial k-spaceacquisition and image reconstruction using a factor of 2 interleavedview sharing, radial SENSE with rescaled matrix, and fat suppres-sion with phase detection.

FIG. 2. Implementation of CG iterativeSENSE reconstruction using the rescalingmethod. (Note that all of the operations areperformed only in a rectilinear grid, whicheliminates the need for conventional convo-lution based gridding (13).)

834 Park et al.

The entire reconstruction process of this method is per-formed in the rescaled rectilinear grids, which eliminatesthe need for convolution gridding.

Masking Water Pixels Using Phase Detection

Conventional density-compensated gridding reconstruc-tion (11) is performed using the central oversampled re-gion of the interleaved radial k-space to avoid aliasingartifacts and obtain a low-resolution coil image. To detectimage phase, it is necessary to select a particular coilimage in which the ROI has a relatively uniform intensityand slow phase variation. The phase correction describedin Ref. 12 is then applied to the chosen coil image. Fat andwater masks are generated based on the sign of the real partof the signal.

MATERIALS AND METHODS

Data acquisition was performed in five healthy volunteerson a 1.5 T whole-body MR scanner (MAGNETOM Sonata;Siemens Medical Solutions, Erlangen, Germany) equippedwith a high-performance gradient subsystem (maximumamplitude � 40 mT/m; maximum slew rate � 200 mT/m/ms). Informed written consent was obtained from eachvolunteer before the study, and the protocol was approvedby our institutional review board. During each scan, thevolunteers were instructed to hold their breath at the endof inspiration. An eight-element rectangular surface coilarray (dimension of coil element � 11 � 12 cm2) wasplaced anterior and posterior to the subject. The anteriorcoil array was composed of two elements along the bodyaxis and three elements along the left–right direction infront of the scanner. The posterior coil array consisted oftwo elements along the body axis. The internal rectangularcoils were overlapped to null the mutual inductance be-tween neighboring coil elements. Image reconstructionwas performed on a Pentium-IV 3.0 GHz system usingMatlab (MathWorks, Natick, MA, USA).

Determining the Optimum Rescaling Factor and IterationNumber for SENSE Reconstruction

A set of right coronary artery (RCA) image data was ac-quired in one of the volunteers during presumed mid-diastole in a single breath-hold, using an ECG-triggered 3Dsegmented radial stacked SSFP sequence. The imagingparameters were as follows: TR/TE/flip angle � 4.2 ms/2.1 ms/50°, FOV � 200 � 200 mm2, number of acquiredviews/heartbeat � 48, data acquisition matrix � 192 � 192(oversampled to 384 in readout direction), number of par-titions � 6 (interpolated to 12), breath-hold duration �20 s, and slice thickness � 3 mm (interpolated to 1.5 mm).A Hamming-windowed sinc radiofrequency (RF) pulse of600 �s duration was used for slab excitation. A spectral-selective fat saturation RF pulse followed by eight dummypulses with sinusoidally varying flip angles (15) was ap-plied to establish a smooth transition of signal to steadystate before data were acquired at each heartbeat. The viewacquisitions covered an azimuthal k-space sampling rangefrom 0° to 180° (8).

The view acquisition (matrix � 192 � 192) was deci-mated by a reduction factor (R) of 2 (matrix � 96 � 192) tosimulate angular undersampling for SENSE reconstructionwith the rescaling method. Image progression in SENSEreconstruction (R � 2, r � 2) was demonstrated as Niter

increased. A reference image was generated by SENSEwith the rescaling method (r � 3, Niter � 6) using the fullyacquired data. To determine the optimum rescaling factorand iteration number at which image accuracy was tradedoff with noise amplification for R � 2, the artifact power(AP) for an ROI was calculated at r � 1, 2, and 3 with theincrease of Niter:

Artifact Power �AP� �

�j

�IjREFERENCE � � �Ij

SENSE �2

�j

�IjSENSE �2

[2]

where j is a pixel index, IREFERENCE is the reference image,and ISENSE is the SENSE image. A higher AP shows in-creased artifacts and reduced image quality. The referenceimage was compared with the SENSE images at r � 1, 2,and 3, with Niter � 6. The SENSE reconstruction time wasmeasured during a single iteration for the three rescalingfactors. The total reconstruction time for the entire 3D cinedata was also measured.

Volunteer Studies for the CAPS Technique

To investigate the feasibility of the CAPS technique, weacquired left anterior descending (LAD) data from fivevolunteers using a 3D segmented cine radial stacked SSFPsequence in a single breath-hold. The imaging parameterswere as follows: TR/TE/flip angle � 4.2 ms/2.1 ms/50°,FOV � 200 � 200 mm2, number of views/cine phase/heartbeat � 12, data acquisition matrix/cine phase � 48 �192 (oversampled to 384 in the readout direction), numberof cardiac phases � 17–20, heartbeats/partition � 4, num-ber of partitions � 6 (interpolated to 12), breath-hold du-ration � 20–24 s, and slice thickness � 3 mm (interpolatedto 1.5 mm). A Hamming-windowed sinc RF pulse of600 �s duration was used for slab excitation. No spectral-selective fat saturation pulse was applied. Acquisitionswere preceded by an �/2 RF preparation pulse (� � RF flipangle for data acquisition) and dummy pulses in the firstheartbeat to establish steady state. The view of in-planeradial k-space was undersampled azimuthally rangingfrom 0° to 180°. The in-plane k-space in each cine phasewas composed of either even or odd views in an inter-leaved fashion.

The effectiveness of the overall implementation shownin Fig. 1 was demonstrated using the fifth slice in the14th–15th cardiac phases. Four images were generated by1) the conventional gridding and root SOS reconstructionof all the coil images using the 14th phase data (matrix �48 � 192), 2) the conventional gridding and root SOSreconstruction of all the coil images using a factor of 2interleaved view sharing (matrix � 96 � 192), 3) SENSEreconstruction with the rescaling method (r � 2) using theinterleaved data (matrix � 96 � 192), and 4) SENSE re-construction with the rescaling method (r � 2) and fat

4D Coronary MRA With CAPS 835

suppression with phase detection using the interleaveddata (matrix � 96 � 192). The dynamic motion of the LADwith in-plane spatial resolution of 1.0 � 1.0 mm2 andtemporal resolution of 48 ms was demonstrated using theimages in the sixth slice of a 3D slab, as well as 3D maxi-mum intensity projection (MIP) images for comparison.Multiple MIP processed images acquired during cardiacmid-diastole (11th–16th cine phases) were employed foraveraging.

Conventional 3D radial data acquisition was comparedwith the proposed CAPS technique for the same LAD inone of the volunteers. The conventional data acquisitionmethod was implemented with a single cine phase inpresumed mid-diastole (RR interval � 840 ms, data acqui-sition window � 202 ms). The imaging parameters werethe same as those used in the previous RCA data acquisi-tion. Images were generated using SENSE with the rescal-ing method (r � 2, matrix �192 � 192), and then pro-cessed by 3D MIP. For the proposed technique, the imag-ing parameters were the same as those used in the previousCAPS data acquisition. Images were reconstructed usingSENSE with the rescaling method (r � 2) after interleavingwith a factor of 2 (matrix � 96 � 192). Multiple MIPprocessed images from the 13th–16th cine phases wereretrospectively averaged for comparison with the conven-tional single-phase data acquisition.

RESULTS

Figure 3 shows the image progression of radial SENSE(R � 2) with the rescaling method (r � 2) as the number of

iterations (Niter) is increased. High-intensity blurred im-ages are generated during the first several iterations (Fig.3a and b). Image accuracy is increased as Niter rises to 8(Fig. 3c). As Niter reaches 20 and 30 (Fig. 3d and e), severenoise is observed over the entire image.

The AP in the SENSE images (R � 2) is shown with theincrease of Niter at r � 1, 2, and 3 (Fig. 4a). The APdecreases rapidly until Niter is around 6 with all of therescaling factors, and increases slowly afterwards. Theminimum AP at r � 1 is nearly three times higher than thatwith r � 2 and 3. After Niter passes 6, the AP rises morerapidly with r � 1 than with r � 2 or 3. Correspondingly,amplified noise is observed in SENSE image with r � 1(Fig. 4c) as compared to the reference image (Fig. 4b). As ris increased, SENSE reconstruction demonstrates reducednoise and yields nearly the same quality of image betweenr � 2 and 3 (Fig. 4d and e). The SENSE reconstruction timefor a single iteration is 3.6, 10.4, and 25.2 s for r � 1, 2, and3, respectively. Considering the minimum artifact leveland reconstruction time, the rescaling factor of 2 is esti-mated to be appropriate. For the entire 3D cine data set, ittakes about 1 hr for SENSE reconstruction with r � 2 andNiter � 6.

Figure 5 demonstrates the effectiveness of the overallimplementation described in Fig. 1, using the radial k-space in the fifth slice of a 3D slab over the 14th–15thcardiac cine phases. The conventional gridding followedby the root SOS reconstruction of all the coil images withdata from one cardiac phase (the 14th cardiac phase, ma-trix � 48 � 192) results in severe noise and streak artifacts(Fig. 5a). The conventional gridding and root SOS recon-

FIG. 3. Image progression withan iteration count (Niter) in radialSENSE reconstruction (R � 2,matrix � 96 � 192) with therescaling method (r � 2). Notethat the image progresses fromlow (a and b) to high (c–e) spatialresolution during the iterations.Noise is also increased with theiterations.

836 Park et al.

struction with a factor of 2 interleaved view sharing (14th–15th cardiac phases, matrix � 96 � 192) reduces the noiseand streak artifacts (Fig. 5b). The radial SENSE with therescaling method (r � 2) generates an image that is free ofstreak artifact and amplified noise after six iterations (Fig.5c). The remaining fat around the coronary artery is nearlyremoved using phase detection with phase-sensitive SSFP(Fig. 5d).

Dynamic motion of the LAD is demonstrated over sev-eral cardiac phases from cardiac systole to mid-diastole(Fig. 6). Each image is reconstructed using a factor of 2interleaved view sharing, radial SENSE with the rescalingmethod, and fat suppression using phase detection. Figure6a shows LAD motion in the sixth slice of a 3D slab. A partof the LAD appears in the slice of the zeroth cardiac cinephase. The LAD moves out of the slice after the zerothphase, and then back into it at the ninth phase. A largesegment of the LAD remains in the slice from the 11th to13th phases, and then starts gradually moving out of itafter the 13th phase. Figure 6b shows 3D MIP images thatdescribe the motion of the LAD over the entire cardiaccycle. Using 3D MIP processing, the dynamic motion ofLAD is traced even in cardiac systole (phases 0–7). The 3DMIP LAD images remain nearly the same from the 11th tothe 16th phases, enabling multiple images to be averaged

(Fig. 6d). Compared with the enlarged image without av-eraging (Fig. 6c), the averaged image in Fig. 6d demon-strates a higher SNR as well as clearer delineation of thevessel (arrows).

Figure 7 compares conventional 3D radial data acquisi-tion using a spectrally selective, fat-saturation pre-pulse inmid-diastole with the proposed CAPS data acquisition. Inthe conventional data acquisition (Fig. 7a), remaining fatsignals are observed (arrows) due to the longitudinal signalrecovery over the large data acquisition window(�202 ms) following the fat saturation pre-pulse. Addi-tionally, image blurring results from cardiac motion due toa non-optimal ECG trigger delay. Compared to the conven-tional data acquisition, the proposed CAPS technique withretrospective averaging demonstrates superior image qual-ity resulting from enhanced fat suppression, improvedSNR, and increased vessel sharpness over the entire image.

DISCUSSION

The proposed 4D coronary MRA technique, CAPS, wassuccessfully performed and yielded artifact-free and fat-suppressed images with a factor of 2 interleaved viewsharing, self-calibrating radial SENSE reconstruction with

FIG. 4. Effect of the rescalingfactor on radial SENSE recon-struction: (a) AP vs. Niter withr � 1, 2, and 3 (R � 2), and(b–e) SENSE images using (b)reference (r � 3, R � 1), (c) r �1, R � 2, (d) r � 2, R � 2, and(e) r � 3, R � 2. Six iterationsare used for the SENSE im-ages. Note the decreasednoise with r � 2 and 3 (d ande) compared to r � 1 (c).

4D Coronary MRA With CAPS 837

FIG. 5. LAD images generated using the fifth sliceof a 3D slab in the 14th–15th cine phases: (a)conventional gridding and root SOS reconstruc-tion with the 14th phase data (matrix � 48 � 192),(b) conventional gridding and root SOS recon-struction with a factor of 2 interleaved view sharing(matrix � 96 � 192), (c) SENSE with the rescalingmethod (r � 2, Niter � 6, matrix � 96 � 192), and(d) SENSE with the rescaling method (r � 2, Niter �6, matrix � 96 � 192) with fat suppression usingphase detection.

FIG. 6. LAD dynamic motion in cardiac cine phases: (a) images in the sixth slice in a 3D slab, (b) 3D MIP processed images, (c) an enlarged3D MIP image in the 11th phase, and (d) an average of six multiple MIP processed images (11th–16th cine phases). Note that the LAD istraced for all cardiac phases in b (in contrast to a), and SNR and vessel delineation are enhanced in d compared to c.

838 Park et al.

the rescaling method, and phase-sensitive SSFP usingphase detection.

The CAPS technique offers several advantages over con-ventional coronary MRA. First, single-frame mid-diastolicdata acquisition may require an extra scan to determine anaccurate trigger delay and data acquisition window. Thecontinuous data acquisition with high temporal resolutionin CAPS eliminates concerns about the trigger delay andacquisition window without extra scans. Second, to in-crease the efficiency of fat suppression, the centric reor-dering of phase-encoding lines is typically used for dataacquisition after a frequency-selective fat saturation pre-pulse. However, signals from myocardial tissue around thecoronary artery are not sufficiently suppressed, becausethe images are not T2-weighted. To obtain better contrastbetween blood vessel and myocardial tissue, T2-preparedmagnetization must be applied before data acquisition(16). However, the CAPS steady-state data acquisition al-lowed complete T2/T1 weighting with fat suppression,eliminating the need for T2-prepared magnetization prep-aration. Third, a real-time 2D coronary MRA techniquewith cine data acquisition, which is similar to CAPS, hasbeen developed (9,10). However, since this method is sen-sitive to coronary artery motion, it is necessary to detectthe imaging frames in which the vessel is contained foraveraging. The CAPS technique makes it possible to tracecoronary artery motion over all cardiac phases using 4Ddata acquisition and 3D MIP. Multiple images acquiredduring cardiac mid-diastole can be simply averaged forSNR improvement without detecting specific imagingframes in which the coronary artery is visible.

Since the reduction factor is increased in conventionalSENSE, reconstruction speed becomes faster because theconventional gridding operation takes less time. In SENSEreconstruction using the rescaling method, the griddingprocess is removed, and the entire operation is performedin the rescaled rectilinear grids. The reconstruction speeddepends on the rescaling factor and FFT, rather than thereduction factor. Further investigations with a clinicalscanner are needed for a quantitative comparison of com-putational speed between the two reconstruction algo-rithms.

In the proposed SENSE reconstruction, initializing thematrix EHm in Eq. [1] includes the rescaling process ofmeasured radial k-space onto a large rectilinear grid withno density compensation. Removing the density compen-sation in the initializing process makes the image progressfrom a high-intensity, low-resolution image (as shown in

Fig. 3). As the CG loop is converged, spatial resolution isincreased, but noise is also amplified. The noise amplifi-cation results from accumulated rounding errors in therescaling process at each iteration. SNR can be traded offfor spatial resolution by adjusting the iteration number atwhich the CG loop stops. The iteration number is thereforeequivalent to the regularization parameter of the matrixinversion (17,18). However, initializing the matrix EHE inEq. [1] without density compensation requires a highernumber of iterations for convergence of the CG loop com-pared to conventional SENSE. As regards image accuracyand noise, SENSE reconstruction using the rescalingmethod currently requires six to 10 iterations for R � 2, asshown in Fig. 4a. Further studies are needed to enhancethe convergence rate of CG iteration using Fourier precon-ditioning methods, as discussed by Clinthorne et al. (19).

In CAPS, the conventional fat saturation pulse is notapplied before data acquisition, because it interrupts thesteady state and reduces temporal resolution. Instead, fatsuppression is performed using the multiplication of aSENSE image with a water mask image. The water maskimage is generated using phase detection of a coil imageafter the gridding reconstruction. It is important to applyphase detection after the gridding reconstruction becausethe image phase in SSFP acquisition may not be preservedduring the iterations in the SENSE. Compared to a spec-trally-selective fat-saturation pre-pulse, fat suppressionusing phase-sensitive SSFP can be more robust to a broadrange of off-resonance by increasing TR to 4.6 ms, asshown in Ref. 12. However, banding artifacts may appearand impede the depiction of vessel with the increased TRin SSFP. Additionally, signals in the phase-sensitive SSFPinterfere destructively at the boundary between the coro-nary artery and epicardial fat. This leads to a decrease insignal at the boundary due to phase cancellation, whichgives the appearance of reduced vessel lumen size. In thecurrent SSFP implementation at TR � 4.2 ms, the bandingartifacts did not appear in the ROI, and misregistration offat and water was observed only outside the ROI due tooff-resonance.

In this work we determined the length of the slidingwindow to achieve a reduction factor of 2 by interleavingthe data in two consecutive cine phases. If the slidingwindow is extended, the reduction factor of the in-planeradial k-space can be reduced, thus enhancing SNR andreducing the number of iterations in the SENSE recon-struction. However, image reconstruction can be sensitiveto cardiac motion with the extended sliding window.

FIG. 7. Comparison of the 3D MIPimages generated using (a) conven-tional 3D radial coronary data acqui-sition with a spectrally-selective fat-saturation pre-pulse (matrix � 192 �192) and SENSE reconstruction, and(b) the proposed CAPS data acqui-sition, SENSE reconstruction, andretrospective averaging over fourcine phases in mid-diastole. Notethe remaining fat signals in a, andenhanced fat suppression and SNRin b.

4D Coronary MRA With CAPS 839

Therefore, there is a limit to increasing the length of slid-ing window in breath-hold 3D cine coronary artery imag-ing. To overcome the current limitation, self-gating forrespiratory motion (20) is needed to increase temporalresolution as well as the number of view acquisitionsachievable in a cine phase.

CONCLUSIONS

The proposed 4D radial coronary MRA technique, CAPS,offers the potential to trace the dynamic motion of coro-nary arteries during the entire cardiac cycle with simulta-neous high temporal and spatial resolution. Further inves-tigations are needed to resolve the reduction of vessellumen size from phase-sensitive fat suppression, and toincrease through-plane spatial resolution.

ACKNOWLEDGMENTS

The authors thank Dr. Jordin Green and Dr. Rohan Dhar-makumar for valuable comments and editing the languagein the manuscript, and Dr. Mark Griswold for valuablediscussions on parallel imaging.

REFERENCES

1. Stuber M, Danias PG, Botnar RM, Sodickson DK, Kissinger KV, Man-ning WJ. Superiority of prone position in free-breathing 3D coronaryMRA in patients with coronary disease. J Magn Reson Imaging 2001;13:185–191.

2. Stuber M, Botnar RM, Danias PG, Sodickson DK, Kissinger KV, VanCauteren M, De Becker J, Manning WJ. Double-oblique free-breathinghigh resolution three-dimensional coronary magnetic resonance an-giography. J Am Coll Cardiol 1999;34:524–531.

3. Deshpande VS, Shea SM, Laub G, Simonetti OP, Finn JP, Li D. 3Dmagnetization-prepared true-FISP: a new technique for imaging coro-nary arteries. Magn Reson Med 2001;46:494–502.

4. Edelman RR, Manning WJ, Burstein D, Paulin S. Coronary arteries:breath-hold MR angiography. Radiology 1991;181:641–643.

5. Huber ME, Kozerke S, Pruessmann KP, Smink J, Boesiger P. Sensitivity-encoded coronary MRA at 3T. Magn Reson Med 2004;52:221–227.

6. Park J, McCarthy R, Li D. Feasibility and performance of breath-hold 3Dtrue-FISP coronary MRA using self-calibrating parallel acquisition.Magn Reson Med 2004;52:7–13.

7. Meyer CH, Hu BS, Nishimura DG, Macovski A. Fast spiral coronaryartery imaging. Magn Reson Med 1992;28:202–213.

8. Larson AC, Simonetti OP, Li D. Coronary MRA with 3D undersampledprojection reconstruction TrueFISP. Magn Reson Med 2002;48:594–601.

9. Nayak KS, Pauly JM, Yang PC, Hu BS, Meyer CH, Nishimura DG.Real-time interactive coronary MRA. Magn Reson Med 2001;46:430–435.

10. Hardy CJ, Saranathan M, Zhu Y, Darrow RD. Coronary angiography byreal-time MRI with adaptive averaging. Magn Reson Med 2000;44:940–946.

11. Rasche V, Proksa R, Sinkus R, Bornert P, Eggers H. Resampling of databetween arbitrary grids using convolution interpolation. IEEE TransMed Imaging 1999;18:385–392.

12. Hargreaves BA, Vasanawala SS, Nayak KS, Hu BS, Nishimura DG.Fat-suppressed steady-state free precession imaging using phase detec-tion. Magn Reson Med 2003;50:210–213.

13. Pruessmann KP, Weiger M, Bornert P, Boesiger P. Advances in sensi-tivity encoding with arbitrary k-space trajectories. Magn Reson Med2001;46:638–651.

14. Oesterle C, Markl M, Strecker R, Kraemer FM, Hennig J. Spiral recon-struction by regridding to a large rectilinear matrix: a practical solutionfor routine systems. J Magn Reson Imaging 1999;10:84–92.

15. Paul D, Hennig J. Comparison of different flip angle variation functionsfor improved signal behavior in SSFP sequences. In: Proceedings of the12th Annual Meeting of ISMRM, Kyoto, Japan, 2004. p 2663.

16. Brittain JH, Hu BS, Wright GA, Meyer CH, Macovski A, Nishimura DG.Coronary angiography with magnetization-prepared T2 contrast. MagnReson Med 1995;33:689–696.

17. Sodickson DK. Tailored SMASH image reconstructions for robust invivo parallel MR imaging. Magn Reson Med 2000;44:243–251.

18. Lin FH, Kwong KK, Belliveau JW, Wald LL. Parallel imaging recon-struction using automatic regularization. Magn Reson Med 2004;51:559–567.

19. Clinthorne NH, Pan TS, Chiao PC, Rogers WL, Stamos JA. Precondi-tioning methods for improved convergence rates in iterative recon-structions. IEEE Trans Med Imaging 1993;12:78.

20. Larson AC, Kellman P, Arai A, Hirsch GA, McVeigh E, Li D, SimonettiOP. Preliminary investigation of respiratory self-gating for free-breath-ing segmented cine MRI. Magn Reson Med 2005;53:159–168.

840 Park et al.