NOTE Magnetic Resonance in Medicine 74:474–481 (2015)
Reducing View-Sharing Using Compressed Sensing in Time-Resolved Contrast-Enhanced Magnetic Resonance Angiography Stanislas Rapacchi,1 Yutaka Natsuaki,2 Adam Plotnik,1 Simon Gabriel,1 Gerhard Laub,2 J. Paul Finn,1 and Peng Hu1* Purpose: To study temporal and spatial blurring artifacts from k-space view-sharing in time-resolved MR angiography (MRA) and to propose a technique for reducing these artifacts. Methods: We acquired k-space data sets using a threedimensional time-resolved MRA view-sharing sequence and retrospectively reformatted them into two reconstruction frameworks: full view-sharing via time-resolved imaging with stochastic trajectories (TWIST) and minimal k-space view-sharing and compressed sensing (CS-TWIST). The two imaging series differed in temporal footprint but not in temporal frame rate. The artifacts from view-sharing were compared qualitatively and quantitatively in nine patients in addition to a phantom experiment. Results: CS-TWIST was able to reduce the imaging temporal footprint by two- to three-fold compared with TWIST, and the overall subjective image quality of CS-TWIST was higher than that for TWIST (P < 0.05). View sharing caused a delay in the visualization of small blood vessels, and the mean transit time of the carotid artery calculated based on TWIST reconstruction was 0.6 s longer than that for CS-TWIST (P < 0.01). In thoracic MRA, the shorter temporal footprint decreased the sensitivity to physiological motion blurring, and vessel sharpness was improved by 8.8% 6 6.0% using CS-TWIST (P < 0.05). Conclusion: In time-resolved MRA, the longer temporal footprint due to view-sharing causes spatial and temporal artifacts. CS-TWIST is a promising method for reducing these C 2014 Wiley artifacts. Magn Reson Med 74:474–481, 2015. V Periodicals, Inc. Key words: time-resolved CE-MRA; MR angiography; viewsharing; compressed sensing (CS); temporal footprint; TWIST
INTRODUCTION Dynamic (time-resolved) contrast-enhanced magnetic resonance angiography (CE-MRA) is widely employed as a clinical tool in vascular imaging from head to toe. Its noninvasive nature and lack of ionizing radiation make it an appealing alternative to digital subtraction angiography and computed tomography angiography (1–5). Fur-
thermore, time-resolved CE-MRA provides simplified scanning logistics, as it circumvents some of the limitations of traditional single-phase CE-MRA, such as incorrect contrast bolus timing and venous contamination. Time-resolved CE-MRA is also used for qualitative and quantitative assessment of vascular dynamics and tissue perfusion in various applications (6,7). However, due to limitations in image acquisition speed, it is often challenging to achieve adequate temporal and spatial resolution at the same time for certain applications. To obtain a faster apparent temporal frame rate, the commonly used method of time-resolved imaging with stochastic trajectories (TWIST) (8,9) uses a view-sharing technique (10) whereby the k-space center is updated more frequently than the peripheral k-space, and subsequently the undersampled peripheral k-space is shared among successive temporal frames. Consequently, the temporal footprint of each image frame is typically 3–5 times longer than the apparent temporal frame duration, which may result in temporal blurring of rapidly changing events in smaller vessels and at vessel edges (11,12) and image artifacts associated with motion occurring throughout the long temporal footprint. View-sharing can also produce reconstruction artifacts (e.g., from parallel imaging) as it combines k-space data acquired over a long time span where signal intensity varies rapidly during contrast injection. In this study, we examined the temporal and spatial blurring effect of view-sharing in time-resolved CE-MRA using phantom and in vivo data. Furthermore, we propose a reconstruction technique that combines compressed sensing (CS), which has shown great potential to achieve significant acceleration of time-resolved CEMRA (13–15), with parallel imaging in order to shorten the temporal footprint and mitigate the associated temporal and spatial blurring effects. METHODS TWIST Acquisition Strategy
1
Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, California, USA. 2 Siemens Healthcare, Los Angeles, California, USA. *Correspondence to: Peng Hu, Ph.D., 300 UCLA Medical Plaza Suite B119, Los Angeles, CA 90095. E-mail: [email protected] Potential conflict of interest: The authors received research support from Siemens Healthcare. Received 11 March 2014; revised 23 July 2014; accepted 25 July 2014 DOI 10.1002/mrm.25414 Published online 26 August 2014 in Wiley Online Library (wileyonlinelibrary. com). C 2014 Wiley Periodicals, Inc. V
The TWIST pulse sequence is a gradient echo type sequence with 3-dimensional Cartesian k-space data acquisition (8,16). The TWIST sequence separates the ky-kz plane into two regions: an inner central region A, which typically contains 5%–30% of the k-space samples, and an outer peripheral region B. After an initial “prep” phase of 2 s to bring the magnetization into steady state, the TWIST sequence acquires full k-space data only at the beginning. For each subsequent temporal
474
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
475
FIG. 1. TWIST and CS-TWIST kspace sampling patterns. The particular sampling distribution along a spiral in ky-kz created a pseudoincoherent undersampling in the B region. The A region was regularly undersampled. Acquisition parameters: A ¼ 15%; four B trajectories; matrix ¼ 356 106. The same data were used for the two reconstruction schemes.
frame after the initial prep phase, the entire A region is sampled. The B region is further divided into Nb ¼ 2–5 trajectories. Each trajectory is a spiral-like sampling trajectory that covers the same B region, but with a two- to five-fold reduced sampling density. For each temporal frame, only one of the Nb trajectories is sampled (Fig. 1). As a result, only 1/Nb of the k-space samples in the B region is sampled between two consecutive A regions. To compensate for this, a view-sharing strategy is applied in the standard TWIST reconstruction, where a composite k-space is formed by including the A region from the current time point and all B region trajectories from 1 to Nb that are sampled at a time point that is closest to the current time (Fig. 1, bottom). The view-sharing is preferably performed forward (11), combining each A and B pair with the missing B trajectories acquired previously to provide sharp imaging of contrast bolus as it travels through the blood vessels. The resulting images have a frame rate equal to the frequency of sampling the A region, but the temporal footprint of each image is greatly lengthened due to the need for including B region trajectories that are sampled at a time up to 5–7 seconds before the current time point. Clinically, TWIST is combined with parallel imaging, where resulting undersampled k-space is sorted into A and B regions for viewsharing acquisition. To reduce the view-sharing, we evaluated an alternative approach based on the same TWIST k-space data, but the view-sharing is limited to only B trajectories that are sampled immediately before and after the A region (Fig. 1, top). This strategy results in undersampled kspace data, but the temporal footprint of each frame is significantly reduced by two- to three-fold and is equivalent to the apparent duration of each temporal frame.
Image Reconstruction Because of the reduction of view-sharing in our approach, the B region is further undersampled and its spiral-like trajectory is incoherent sampling that is suitable for CS but not directly suitable for traditional parallel imaging. The A region is regularly undersampled, which is not directly suitable for CS. In our approach, we propose combining a recently developed CS technique that uses a magnitude subtraction to enhance sparsity in CEMRA data (15) and the SPIRiT technique (17). The combined technique can deal with both the regularly undersampled A region and the incoherently undersampled B region in a standard TWIST dataset. In a time-resolved CE-MRA experiment, suppose K1 and K2 are two consecutive k-space data vectors in the temporal series of undersampled k-space data, and the CS reconstruction of images I1 and I2 corresponds to the minimization of one L2-norm of the difference between estimated and measured samples (the fidelity term) and one or several L1-norms of sparsity penalizations (the sparsity terms). Using the previously described magnitude subtraction CS algorithm (15) ( ðI1 ; I2 Þ ¼ argmin
jj U1 F ðI1 Þ K1 jj22 þ lTV ðI1 Þ þ mjjI2 j jI1 jj1
)
jj U2 F ðI2 Þ K2 jj22 þ lTV ðI2 Þ þ mjjI2 j jI1 jj1
[1] where U1 and U2 are the under-sampling masks, F is the Fourier transform, TV is the total variation of the images I1 and I2, and the last L1 norm is the magnitude subtraction penalty applied only on the magnitude images. In a TWIST data set, the A region is regularly undersampled and the coherent artifacts cannot be removed
476
Rapacchi et al.
FIG. 2. Phantom experiment. An air bubble was injected into a set of tubes. The same acquisition was reconstructed with full viewsharing (TWIST) and with reduced view-sharing (CS-TWIST). a: The maximum intensity projection of a single CS-TWIST time frame prior to injection shows the phantom setup. b: Reduced signal intensity was observed in time along a projection line perpendicular to the tubes. c: The intensity of the tube cross-sections, marked with blue and red asterisks in panel b, is plotted against time. The signal intensity from the air bubble regions using conventional TWIST reconstruction is approximately two to three times higher than the proposed CS-TWIST. Acquisition parameters: A ¼ 5%, five B trajectories; frame rate ¼ 1.25s/frame; resolution ¼ 1.3 1.3 1.3 mm3; parallel imaging ¼ 2 2. Abbreviations: CSTW, CS-TWIST; TW, TWIST.
using the CS algorithm shown in Equation 1. Therefore, a SPIRiT intercoil operator S (17) was applied in the image space as shown in Equation 2: ( ðI1 ; I2 Þ ¼ argmin
jjU1 F ðI1 Þ K1 jj22 þ jjðS IÞI1 jj22 þ lTVðI1 Þ þ mjjI2 j jI1 jj1
!)
jjU2 F ðI2 Þ K2 jj22 þ jjðS IÞI2 jj22 þ lTVðI2 Þ þ mjjI2 j jI1 jj1
[2] Here, S establishes a direct intercoil transformation for each pixel from N image to N image, where N is the number of receiving coil elements and I is the identity matrix. The minimization for Equation 2 is performed using a split Bregman (18) algorithm. The redundancy across coil elements allows for the removal of aliasing artifacts. The intercoil relationships in S are assumed to be constant over time. SPIRiT operators S are estimated on the basis of 11 9 9 kernels from the autocalibration lines that are typically acquired separately at the beginning of the sequence. To reduce the requirements of computer memory for handling such a large data set, the reconstruction was performed in a slice-wise fashion: the frequency-encoding dimension (kx) was Fouriertransformed (after zero-padding of partial Fourier data) and the reconstruction algorithm was performed separately in the ky-kz space for each slice in the x direction. Parameters for the split Bregman minimization were m ¼ 0.1, l ¼ 0.1 and the thresholds for shrinkage were data-driven using the 90% level of the histogram of each Bregman distance. These values were determined so that images reconstructed from full view-shared k-space using our CS algorithm had a noise level similar to that of the images reconstructed using parallel imaging only. The maximum number of iteration was set to 30, with a stopping criterion if the normalized error between two consecutive estimates fell below 104. For comparison, the standard TWIST reconstruction was also performed on the same data set offline. In the
standard TWIST reconstruction, view-sharing combines each A and B pair with older B trajectories such that a regularly undersampled k-space is generated. Hence, traditional parallel imaging using SPIRiT alone was employed in a projection onto convex sets algorithm, which solves Equation 3: I1 ¼ argmin jjU1 F ðSI1 Þ K1 jj22 ; subject to I1 ¼ SI1
[3]
Phantom Study A phantom experiment was designed to demonstrate the effect of view-sharing on time-resolved CE-MRA. The experiment reproduces the scenario of using TWIST to image fast flow in smaller vessels to observe temporal blurring and validate the CS reconstruction of reduced view-shared data. A set of tubes was filled with GdDTPA (Magnevist; Bayer Healthcare Pharmaceuticals, Whippany, New Jersey, USA) diluted with saline. An air bubble was introduced using a clinical power injector (Spectris; Medrad, Indianola, Pennsylvania, USA) at 0.75 mL/s, giving an average velocity of 4–5 cm/s. A TWIST acquisition was performed throughout the injection to track the movement of the air bubble on a 3T scanner (Magnetom Tim Trio; Siemens Medical Solutions, Erlangen, Germany). The acquisition parameters were as follows: size of A region ¼ 5% of k-space; five B trajectories; temporal frame rate ¼ 1.25 s/frame; spatial resolution ¼ 1.3 1.3 1.3 mm3; matrix size ¼ 230 230 46; parallel imaging (19) acceleration of 2 2 (24 24 calibration lines); partial Fourier ¼ 6/8; and echo time (TE)/pulse repetition time (TR) ¼ 1.56/2.55 ms. Both TWIST (temporal footprint Tf ¼ 6.0 s) using Equation 3 and CS-TWIST (Tf ¼ 2.2 s) reconstructions using Equation 2 were performed based on the same dataset. To compare and validate the temporal behavior of the image
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
477
FIG. 3. Single in vivo timeresolved CE-MRA data were reconstructed using backward view-sharing TWIST (1–4 [one new B and four future B]), typical forward view-sharing TWIST (4-1 [four past B and one new B]), CS-TWIST (1-1 [one new B and one past B]) and no view-sharing with SPIRiT reconstruction SPIRiT (0–1 [one new B]). Thin coronal MIP images are shown by TWIST (4-1) and CS-TWIST (1-1) (top and bottom). Enhancement curves were plotted (middle row) for a carotid artery ROI (small vessel) and a jugular vein ROI (large vessel). The signal curves for the jugular vein ROI using the four image reconstruction techniques were similar. However, the signal curves for the carotid artery using the four reconstruction techniques were different. TWIST (1–4) and TWIST (4-1) both demonstrated temporal blurring with longer time duration of contrast enhancement than the CS-TWIST (1-1) and SPIRiT (0–1). Acquisition parameters: A ¼ 8%; five B trajectories; resolution ¼ 1.3 1.3 1.3 mm3; frame rate ¼ 1.7 s/frame.
series, the signal intensity curves at a cross-section along the tubing were generated based on TWIST and CS-TWIST, respectively, and the full-width half-maximum (FWHM) was calculated as the apparent time it took for the air bubble to pass the cross-section in the TWIST and CS-TWIST image series, respectively. The ground truth was also calculated by dividing the actual length of the air bubble by the expected flow velocity estimated based on the set volumetric flow rate of the power injector and the cross-sectional area of the tubing. In Vivo Study Because our image reconstruction technique can be applied to the k-space raw data of a standard TWIST acquisition, our in vivo study was performed retrospectively based on TWIST data acquired for nine patients who underwent the standard clinical TWIST CE-MRA imaging protocol for evaluation of thoracic (N ¼ 5) or intracranial blood vessels (N ¼ 4). The study was approved by our Institutional Review Board for retro-
spective analysis. The sequence parameters for the intracranial MRA scans were as follows: TE/TR ¼ 0.8–0.9/1.9– 2.1 ms; resolution ¼ 1.3 1.3 1.3 mm3; temporal frame rate ¼1.25 s/frame; parallel imaging (19) ¼ 3 2 with separate 24 24 calibration lines; partial Fourier (8/10 in ky and kz) and view-sharing (A ¼ 8%–14% and five B trajectories). The sequence parameters for thoracic MRA scans were as follows: TE/TR ¼ 0.9–1.0/2.5–2.6 ms; 1.1 1.1 3–6 mm3; temporal frame rate ¼ 2.35 s/frame; parallel imaging (19) ¼ 2 2 with separate 24 24 calibration lines; partial Fourier (8/10 in ky and kz) and viewsharing (A ¼ 25% and four B trajectories). For each acquired dataset, two image reconstructions were performed. First, a TWIST reconstruction involving full forward view-sharing (four to five B trajectories from past time points and one A) and parallel imaging using SPIRiT as shown in Equation 3 was performed. For demonstration purposes, backward TWIST was reconstructed from one dataset (five B trajectories from future time points and one A); Second, CS-TWIST reconstruction involving reduced view-sharing (two adjacent B
478
Rapacchi et al.
FIG. 4. View-sharing impacts the accurate visualization of the bolus timing. On the left, the temporal maximum intensity projections (coronal volume rendering up 75 mm) confirm minimal compromise in spatial resolution when replacing TWIST with CS-TWIST. However, the mean transit time (MTT, coronal volume rendering up 75 mm) maps (right) exhibit a significant timing offset (in carotids: 17.1 6 3.6 s for TWIST versus 16.5 6 3.4 s for CS-TWIST; P < 0.01) between TWIST and CS-TWIST because of view-sharing. Significant differences in MTT values were observed in smaller vessels (arrows) but not in larger vessels. Acquisition parameters: A ¼ 8%; five B trajectories; time resolution ¼ 1.25 s; resolution ¼ 1.4 1.1 1.3 mm3; parallel imaging ¼ 3 2.
trajectories and one A) was performed using the combined CS-SPIRiT technique shown in Equation 2. With the incorporation of CS to the TWIST reconstruction, the temporal footprint was reduced in intracranial MRA from 6 s to 2.2 s and in thoracic MRA from 5.6 s to 3.7 s. Image Analysis An important physiological parameter that can be extracted from time-resolved CE-MRA is the mean transit time (MTT). MTT is a surrogate for cerebral blood flow employed in nuclear imaging (20–22) that adds critical information in the work-up of stroke patients. To evaluate the impact of view-sharing on MTT and the benefit of temporal footprint reduction, quantitative MTT maps were computed from the two image series as Z I ðtÞt dt MTT ¼ Zt [4] I ðtÞ dt t
where I(t) is the signal intensity at time t. Two regions of interest (ROIs) were drawn, one at the common carotid artery, and one at the jugular vein. The mean MTT values within each ROI were calculated for comparison purposes. For visualization, MTT maps were masked using a two standard deviation threshold from the temporal maximum intensity projection. MTT maps provide comparison for spatial resolution and temporal resolution between the two techniques. The elongated temporal footprint could also introduce increased physiological motion blurring. To quantify spatial blurring introduced by view-sharing in chest MRA, vessel sharpness was quantified for both recon-
structed time series after mask (first volume reconstructed) subtraction. Vessel sharpness was defined as the sum of finite differences of each volume (23). A paired t test was performed to evaluate the significance of sharpness differences. The quality of the images reconstructed from the two different methods in comparison was graded in a blinded fashion by two radiologists trained in cardiovascular MRI. From a consensus reading session, the two readers were asked to evaluate overall image series quality for the purpose of clinical diagnosis. Evaluations were calibrated and blinded to the information related to subject and reconstruction method using a four-point scale (1 ¼ poor, 2 ¼ fair, 3 ¼ good, 4 ¼ very good). From the reconstructed images, maximum intensity projection (MIP) series were presented as uncompressed DICOM to the evaluators on a computer workstation side by side. Statistical significance of MTT and quality scoring was determined using a nonparametric Wilcoxon signed-rank test. Results were considered significant at P < 0.05. RESULTS Phantom Study The phantom experiment results in Figure 2 illustrate the impact of view-sharing on the tracking of the injected air bubble. With the standard TWIST reconstruction, the signal of the tube was only slightly diminished during the passage of the air bubble. This was because of the long temporal footprint that combines data from a time window that is much wider than the time it took the air bubble to pass each point along the tube. However, the reduction of view-sharing using CS-TWIST reconstruction allowed us to observe greater signal dropout during
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
479
FIG. 5. Thin MIP (90 mm thick) of a thoracic CE-time-resolved CE-MRA experiment. The prolonged temporal footprint (Tf) due to viewsharing increased the potential of physiological motion blurring. Mixing data acquired within a shorter duration, CS-TWIST images show reduced motion blurring, which resulted in sharper enhanced blood vessels. Acquisition parameters: A ¼ 25%; four B trajectories; time resolution ¼ 2.35 s; resolution ¼ 0.9 0.9 6 mm3; parallel imaging ¼ 2 2.
the passage of the air bubble. The FWHM of the signal curves were 3.7 s and 5.6 s based on CS-TWIST and TWIST, respectively, while the ground truth for the expected time duration was 2.3 s. In Vivo Study Figure 3 demonstrates the impact of view-sharing in vivo on the carotid artery and jugular vein signal over time. Although the three approaches (backward TWIST [1–4], forward TWIST [4-1], and the proposed CS-TWIST [1-1]) yielded very similar signal intensity curves for the jugular vein, the signal for the carotid artery was dramatically different. TWIST (4-1), which mixes older k-space data with the current data, resulted in an apparent delay of the contrast wash-out. Backward TWIST (1–4), which mixes future k-space data with the current data, resulted in an early apparent wash-in. As contrast wash-in often represents clinical interest, forward view-sharing has been recommended (11). Figure 3 shows that the significant temporal blurring effect was observed primarily in the arteries (at approximately 20–23 seconds postinjection), when the contrast appeared to remain in the carotid arteries in the TWIST image, whereas it was
cleared out of the carotid arteries in the CS-TWIST images. Both TWIST (1–4) and TWIST (4-1) carotid artery curves showed longer signal enhancement, with an FWHM of 7.1 s for TWIST (1–4) and 9.8 s for TWIST (4-1) compared with 6.5 s for CS-TWIST (1-1). Figure 4 shows the difference of MTT quantification between standard TWIST and CS-TWIST time-resolved MRA. Across the patients imaged for intracranial MRA (N ¼ 4) in this study, the time values observed on the map from the CS-TWIST reconstruction had significantly lower MTT values in the carotid arteries compared with the MTT values extracted from the TWIST images series (17.1 6 3.6 s for TWIST versus 16.5 6 3.4 s for CS-TWIST; P < 0.01). However, there was no significant difference of MTT values observed in the jugular veins (20.1 6 1.3 s for TWIST versus 19.8 6 1.2 for CS-TWIST; P ¼ 0.3). Therefore, the elongated temporal footprint affects the smaller vessels and the larger vessels in different ways. The subjective image quality scoring (all patients, N ¼ 9) shows significantly improved diagnostic image quality by CS-TWIST when compared with TWIST (3.1 6 0.6 versus 2.4 6 0.3; P < 0.05). In time-resolved CEMRA of the thorax, the prolonged temporal footprint due to view-sharing resulted in spatial blurring of the vessels
480
that are affected with physiological motion (Figure 5). This blurring is reduced in the CS-TWIST reconstruction due to the much shorter temporal footprint. Based on five thoracic MRA patients in this study, vessel sharpness scores were improved by 8.8% 6 6.0 % (P < 0.01) using CS-TWIST compared with TWIST. DISCUSSION The results of our study suggest that the temporal blurring of time-resolved three-dimensional CE-MRA can be significantly improved by reducing view-sharing and shortening the temporal footprint. The reduced temporal footprint can also reduce spatial blurring effects from physiological motion. An algorithm combining a previously described compressed sensing algorithm with magnitude subtraction and the standard SPIRiT parallel imaging was proposed and validated in a small cohort of patients to reduce the temporal footprint of standard TWIST CE-MRA and its associated spatial and temporal blurring effects. We have shown that our CS-TWIST algorithm improves delineation of blood vessels and evaluation of the hemodynamics of contrast arrival for clinical time-resolved CE-MRA. With a detailed comparison between the proposed CS-TWIST algorithm and standard TWIST, we were able to demonstrate the impact of view-sharing on the CE-MRA images. The phantom experiment reproduced the scenario of imaging fast-traveling boluses that occur in relatively small structures. Because these small structures are mainly represented by the high spatial frequency region of the k-space, which is heavily shared over a long time, the visualization of small structures (e.g., a small blood vessel) is particularly compromised in standard TWIST reconstruction. To put it simply, view-sharing behaves similarly to a convolution of high-frequency information with a window that is larger than the frame rate. In clinical practice, view-sharing artifacts can be difficult to notice, because they depend on multiple parameters such as the size of the blood vessel, spatial resolution, flow velocity, and viewsharing parameters. Our in vivo results from clinical examinations (acquired as part of routine clinical care) show image artifacts related to view-sharing that are similar to our phantom results. Although it is well known that view-sharing causes temporal blurring, its effect on time-resolved CE-MRA has not received sufficient attention in the past. Furthermore, the possibility of missing a small vessel with fast blood flow (Fig. 3) and spatial blurring (Fig. 4) should also be considered when performing TWIST-type CE-MRA. Although CSTWIST uses fewer k-space data than standard TWIST reconstruction, it was still able to provide higher image quality, because reducing the temporal and spatial blurring caused by view-sharing was able to more than compensate any image quality degradation in the CS-TWIST algorithm. It is known that constrained reconstructions can cause loss of small image features. Therefore, it is possible that the signal loss of the carotid arteries based on both CS-TWIST and SPIRiT reconstructions (shown in Figure 3 at TT ¼ 23.3 s) could be due to an image recon-
Rapacchi et al.
struction bias, reduction in temporal footprint, or a combination of both factors. Future studies in larger patient cohorts are needed to separate the contributions to signal loss caused by these factors. CS reconstructions are also known to cause image blurring if certain sparsifying transforms (e.g., total variation) are used. This is evident in Figure 4, where the CS-TWIST image is slightly blurred compared with that for TWIST. However, in Figure 5, CS-TWIST resulted in sharper images than TWIST because of the presence of physiological motion. In this case, the motion blurring caused by long temporal footprint of the TWIST data was much greater than the slight blurring caused by the CS reconstruction algorithm. Finally, the impact of view-sharing on signal intensity could increase errors in quantification of hemodynamic parameters. Blurring from view-sharing can be both spatial and temporal and is nonuniform, as temporal biases depend on the image resolution and the size of vessels. For measurement of MTT, a major hemodynamic biomarker assessed with positron emission tomography (20), MRI may compete favorably as an ionizing radiation-free alternative if spatio-temporal sampling is appropriate. Currently, view-sharing stands as a limitation to the accuracy of MTT measured using timeresolved CE-MRA. Our study shows that the error in measuring MTT or other parameters derived from a perfusion model (6) may not be as simple as a bulk shift in the time-intensity curve, as the time shift in MTT also depends on the size and flow of each vessel relative to the imaging resolution. In conclusion, we have defined and evaluated several limitations of view-sharing in time-resolved CE-MRA and have implemented an image reconstruction method to mitigate them. Understanding the implications of view-sharing may be crucial for appropriate use of timeresolved MRA in complex vascular beds. CS is a potentially powerful tool to improve visualization of rapidly changing signal in small blood vessels. REFERENCES 1. Prince M, Meaney J. Expanding role of MR angiography in clinical practice. Eur Radiol 2006;16(suppl):B3–B8. 2. Hartung MP, Grist TM, Franc¸ois CJ. Magnetic resonance angiography: current status and future directions. J Cardiovasc Magn Reson 2011; 13:19. 3. Willinek WA, Hadizadeh DR, von Falkenhausen M, Urbach H, Hoogeveen R, Schild HH, Gieseke J. 4D Time-resolved MR angiography with keyhole (4D-TRAK): more than 60 times accelerated MRA using a combination of CENTRA, keyhole, and SENSE at 3.0T. J Magn Reson Imaging 2008;27:1455–1460. 4. Fenchel M, Saleh R, Dinh H, Lee MH, Nael K, Krishnam M, Ruehm SG, Miller S, Child J, Finn JP. Juvenile and adult congenital heart disease: time-resolved 3D contrast-enhanced MR angiography. Radiology 2007;244:399–410. 5. Krishnam MS, Tomasian A, Lohan DG, Tran L, Finn JP, Ruehm SG. Low-dose, time-resolved, contrast-enhanced 3D MR angiography in cardiac and vascular diseases: correlation to high spatial resolution 3D contrast-enhanced MRA. Clin Radiol 2008;63:744–755. 6. Wright KL, Seiberlich N, Jesberger JA, Nakamoto DA, Muzic RF, Griswold MA, Gulani V. Simultaneous magnetic resonance angiography and perfusion (MRAP) measurement: initial application in lower extremity skeletal muscle. J Magn Reson Imaging 2013;38:1237–1244. 7. Wang K, Schiebler ML, Francois CJ, Del Rio AM, Cornejo MD, Bell LC, Korosec FR, Brittain JH, Holmes JH, Nagle SK. Pulmonary
Reducing View-Sharing Using CS in Time-Resolved CE-MRA perfusion MRI using interleaved variable density sampling and highly constrained cartesian reconstruction (HYCR). J Magn Reson Imaging 2013;38:751–756. 8. Lim RP, Shapiro M, Wang EY, et al. 3D time-resolved MR angiography (MRA) of the carotid arteries with time-resolved imaging with stochastic trajectories: comparison with 3D contrast-enhanced boluschase MRA and 3D time-of-flight MRA. Am J Neuroradiol 2008;29: 1847–1854. 9. Nael K, Krishnam M, Ruehm SG, Michaely HJ, Laub G, Finn JP. Time-resolved MR angiography in the evaluation of central thoracic venous occlusive disease. Am J Roentgenol 2009;192:1731–1738. 10. Carroll TJ, Korosec FR, Swan JS, Grist TM, Frayne R, Mistretta CA. Method for rapidly determining and reconstructing the peak arterial frame from a time-resolved CE-MRA exam. Magn. Reson. Med. 2000; 44:817–820. 11. Mostardi PM, Haider CR, Rossman PJ, Borisch EA, Riederer SJ. Controlled experimental study depicting moving objects in view-shared time-resolved 3D MRA. Magn Reson Med 2009;62:85–95. 12. Johnson CP, Polley TW, Glockner JF, Young PM, Riederer SJ. Buildup of image quality in view-shared time-resolved 3D CE-MRA. Magn Reson Med 2013;70:348–357. 13. Trzasko JD, Haider CR, Borisch EA, Campeau NG, Glockner JF, Riederer SJ, Manduca A. Sparse-CAPR: highly accelerated 4D CEMRA with parallel imaging and nonconvex compressive sensing. Magn Reson Med 2011;66:1019–1032. 14. Lee GR, Seiberlich N, Sunshine JL, Carroll TJ, Griswold MA. Rapid timeresolved magnetic resonance angiography via a multiecho radial trajectory and GraDeS reconstruction. Magn Reson Med 2013;69:346–359.
481 15. Rapacchi S, Han F, Natsuaki Y, Kroeker R, Plotnik A, Lehman E, Sayre J, Laub G, Finn JP, Hu P. High spatial and temporal resolution dynamic contrast-enhanced magnetic resonance angiography using compressed sensing with magnitude image subtraction. Magn Reson Med 2014;71:1771–1783. 16. Song T, Laine AF, Chen Q, Rusinek H, Bokacheva L, Lim RP, Laub G, Kroeker R, Lee VS. Optimal k-space sampling for dynamic contrast-enhanced MRI with an application to MR renography. Magn Reson Med 2009;61:1242–1248. 17. Lustig M, Pauly JM. SPIRiT: iterative self-consistent parallel imaging reconstruction from arbitrary k-space. Magn Reson Med 2010;64:457–471. 18. Goldstein T, Osher S. The split Bregman method for L1-regularized problems. SIAM J Imaging Sci 2009;2:323. 19. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J, Kiefer B, Haase A. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 2002;47:1202–1210. 20. Sobesky J. Refining the mismatch concept in acute stroke: lessons learned from PET and MRI. J Cereb Blood Flow Metab 2012;32:1416– 1425. 21. Carrera E, Jones PS, Iglesias S, Guadagno JV, Warburton EA, Fryer TD, Aigbirhio FI, Baron J-C. The vascular mean transit time: a surrogate for the penumbra flow threshold? J Cereb Blood Flow Metab 2011;31:1027–1035. 22. Carrera E, Jones PS, Alawneh JA, et al. Predicting infarction within the diffusion-weighted imaging lesion: does the mean transit time have added value? Stroke 2011;42:1602–1607. 23. Deriche R. Fast algorithms for low-level vision. IEEE Trans Pattern Anal Mach Intell 1990;12:78–87.
Reducing View-Sharing Using Compressed Sensing in Time-Resolved Contrast-Enhanced Magnetic Resonance Angiography Stanislas Rapacchi,1 Yutaka Natsuaki,2 Adam Plotnik,1 Simon Gabriel,1 Gerhard Laub,2 J. Paul Finn,1 and Peng Hu1* Purpose: To study temporal and spatial blurring artifacts from k-space view-sharing in time-resolved MR angiography (MRA) and to propose a technique for reducing these artifacts. Methods: We acquired k-space data sets using a threedimensional time-resolved MRA view-sharing sequence and retrospectively reformatted them into two reconstruction frameworks: full view-sharing via time-resolved imaging with stochastic trajectories (TWIST) and minimal k-space view-sharing and compressed sensing (CS-TWIST). The two imaging series differed in temporal footprint but not in temporal frame rate. The artifacts from view-sharing were compared qualitatively and quantitatively in nine patients in addition to a phantom experiment. Results: CS-TWIST was able to reduce the imaging temporal footprint by two- to three-fold compared with TWIST, and the overall subjective image quality of CS-TWIST was higher than that for TWIST (P < 0.05). View sharing caused a delay in the visualization of small blood vessels, and the mean transit time of the carotid artery calculated based on TWIST reconstruction was 0.6 s longer than that for CS-TWIST (P < 0.01). In thoracic MRA, the shorter temporal footprint decreased the sensitivity to physiological motion blurring, and vessel sharpness was improved by 8.8% 6 6.0% using CS-TWIST (P < 0.05). Conclusion: In time-resolved MRA, the longer temporal footprint due to view-sharing causes spatial and temporal artifacts. CS-TWIST is a promising method for reducing these C 2014 Wiley artifacts. Magn Reson Med 74:474–481, 2015. V Periodicals, Inc. Key words: time-resolved CE-MRA; MR angiography; viewsharing; compressed sensing (CS); temporal footprint; TWIST
INTRODUCTION Dynamic (time-resolved) contrast-enhanced magnetic resonance angiography (CE-MRA) is widely employed as a clinical tool in vascular imaging from head to toe. Its noninvasive nature and lack of ionizing radiation make it an appealing alternative to digital subtraction angiography and computed tomography angiography (1–5). Fur-
thermore, time-resolved CE-MRA provides simplified scanning logistics, as it circumvents some of the limitations of traditional single-phase CE-MRA, such as incorrect contrast bolus timing and venous contamination. Time-resolved CE-MRA is also used for qualitative and quantitative assessment of vascular dynamics and tissue perfusion in various applications (6,7). However, due to limitations in image acquisition speed, it is often challenging to achieve adequate temporal and spatial resolution at the same time for certain applications. To obtain a faster apparent temporal frame rate, the commonly used method of time-resolved imaging with stochastic trajectories (TWIST) (8,9) uses a view-sharing technique (10) whereby the k-space center is updated more frequently than the peripheral k-space, and subsequently the undersampled peripheral k-space is shared among successive temporal frames. Consequently, the temporal footprint of each image frame is typically 3–5 times longer than the apparent temporal frame duration, which may result in temporal blurring of rapidly changing events in smaller vessels and at vessel edges (11,12) and image artifacts associated with motion occurring throughout the long temporal footprint. View-sharing can also produce reconstruction artifacts (e.g., from parallel imaging) as it combines k-space data acquired over a long time span where signal intensity varies rapidly during contrast injection. In this study, we examined the temporal and spatial blurring effect of view-sharing in time-resolved CE-MRA using phantom and in vivo data. Furthermore, we propose a reconstruction technique that combines compressed sensing (CS), which has shown great potential to achieve significant acceleration of time-resolved CEMRA (13–15), with parallel imaging in order to shorten the temporal footprint and mitigate the associated temporal and spatial blurring effects. METHODS TWIST Acquisition Strategy
1
Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, California, USA. 2 Siemens Healthcare, Los Angeles, California, USA. *Correspondence to: Peng Hu, Ph.D., 300 UCLA Medical Plaza Suite B119, Los Angeles, CA 90095. E-mail: [email protected] Potential conflict of interest: The authors received research support from Siemens Healthcare. Received 11 March 2014; revised 23 July 2014; accepted 25 July 2014 DOI 10.1002/mrm.25414 Published online 26 August 2014 in Wiley Online Library (wileyonlinelibrary. com). C 2014 Wiley Periodicals, Inc. V
The TWIST pulse sequence is a gradient echo type sequence with 3-dimensional Cartesian k-space data acquisition (8,16). The TWIST sequence separates the ky-kz plane into two regions: an inner central region A, which typically contains 5%–30% of the k-space samples, and an outer peripheral region B. After an initial “prep” phase of 2 s to bring the magnetization into steady state, the TWIST sequence acquires full k-space data only at the beginning. For each subsequent temporal
474
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
475
FIG. 1. TWIST and CS-TWIST kspace sampling patterns. The particular sampling distribution along a spiral in ky-kz created a pseudoincoherent undersampling in the B region. The A region was regularly undersampled. Acquisition parameters: A ¼ 15%; four B trajectories; matrix ¼ 356 106. The same data were used for the two reconstruction schemes.
frame after the initial prep phase, the entire A region is sampled. The B region is further divided into Nb ¼ 2–5 trajectories. Each trajectory is a spiral-like sampling trajectory that covers the same B region, but with a two- to five-fold reduced sampling density. For each temporal frame, only one of the Nb trajectories is sampled (Fig. 1). As a result, only 1/Nb of the k-space samples in the B region is sampled between two consecutive A regions. To compensate for this, a view-sharing strategy is applied in the standard TWIST reconstruction, where a composite k-space is formed by including the A region from the current time point and all B region trajectories from 1 to Nb that are sampled at a time point that is closest to the current time (Fig. 1, bottom). The view-sharing is preferably performed forward (11), combining each A and B pair with the missing B trajectories acquired previously to provide sharp imaging of contrast bolus as it travels through the blood vessels. The resulting images have a frame rate equal to the frequency of sampling the A region, but the temporal footprint of each image is greatly lengthened due to the need for including B region trajectories that are sampled at a time up to 5–7 seconds before the current time point. Clinically, TWIST is combined with parallel imaging, where resulting undersampled k-space is sorted into A and B regions for viewsharing acquisition. To reduce the view-sharing, we evaluated an alternative approach based on the same TWIST k-space data, but the view-sharing is limited to only B trajectories that are sampled immediately before and after the A region (Fig. 1, top). This strategy results in undersampled kspace data, but the temporal footprint of each frame is significantly reduced by two- to three-fold and is equivalent to the apparent duration of each temporal frame.
Image Reconstruction Because of the reduction of view-sharing in our approach, the B region is further undersampled and its spiral-like trajectory is incoherent sampling that is suitable for CS but not directly suitable for traditional parallel imaging. The A region is regularly undersampled, which is not directly suitable for CS. In our approach, we propose combining a recently developed CS technique that uses a magnitude subtraction to enhance sparsity in CEMRA data (15) and the SPIRiT technique (17). The combined technique can deal with both the regularly undersampled A region and the incoherently undersampled B region in a standard TWIST dataset. In a time-resolved CE-MRA experiment, suppose K1 and K2 are two consecutive k-space data vectors in the temporal series of undersampled k-space data, and the CS reconstruction of images I1 and I2 corresponds to the minimization of one L2-norm of the difference between estimated and measured samples (the fidelity term) and one or several L1-norms of sparsity penalizations (the sparsity terms). Using the previously described magnitude subtraction CS algorithm (15) ( ðI1 ; I2 Þ ¼ argmin
jj U1 F ðI1 Þ K1 jj22 þ lTV ðI1 Þ þ mjjI2 j jI1 jj1
)
jj U2 F ðI2 Þ K2 jj22 þ lTV ðI2 Þ þ mjjI2 j jI1 jj1
[1] where U1 and U2 are the under-sampling masks, F is the Fourier transform, TV is the total variation of the images I1 and I2, and the last L1 norm is the magnitude subtraction penalty applied only on the magnitude images. In a TWIST data set, the A region is regularly undersampled and the coherent artifacts cannot be removed
476
Rapacchi et al.
FIG. 2. Phantom experiment. An air bubble was injected into a set of tubes. The same acquisition was reconstructed with full viewsharing (TWIST) and with reduced view-sharing (CS-TWIST). a: The maximum intensity projection of a single CS-TWIST time frame prior to injection shows the phantom setup. b: Reduced signal intensity was observed in time along a projection line perpendicular to the tubes. c: The intensity of the tube cross-sections, marked with blue and red asterisks in panel b, is plotted against time. The signal intensity from the air bubble regions using conventional TWIST reconstruction is approximately two to three times higher than the proposed CS-TWIST. Acquisition parameters: A ¼ 5%, five B trajectories; frame rate ¼ 1.25s/frame; resolution ¼ 1.3 1.3 1.3 mm3; parallel imaging ¼ 2 2. Abbreviations: CSTW, CS-TWIST; TW, TWIST.
using the CS algorithm shown in Equation 1. Therefore, a SPIRiT intercoil operator S (17) was applied in the image space as shown in Equation 2: ( ðI1 ; I2 Þ ¼ argmin
jjU1 F ðI1 Þ K1 jj22 þ jjðS IÞI1 jj22 þ lTVðI1 Þ þ mjjI2 j jI1 jj1
!)
jjU2 F ðI2 Þ K2 jj22 þ jjðS IÞI2 jj22 þ lTVðI2 Þ þ mjjI2 j jI1 jj1
[2] Here, S establishes a direct intercoil transformation for each pixel from N image to N image, where N is the number of receiving coil elements and I is the identity matrix. The minimization for Equation 2 is performed using a split Bregman (18) algorithm. The redundancy across coil elements allows for the removal of aliasing artifacts. The intercoil relationships in S are assumed to be constant over time. SPIRiT operators S are estimated on the basis of 11 9 9 kernels from the autocalibration lines that are typically acquired separately at the beginning of the sequence. To reduce the requirements of computer memory for handling such a large data set, the reconstruction was performed in a slice-wise fashion: the frequency-encoding dimension (kx) was Fouriertransformed (after zero-padding of partial Fourier data) and the reconstruction algorithm was performed separately in the ky-kz space for each slice in the x direction. Parameters for the split Bregman minimization were m ¼ 0.1, l ¼ 0.1 and the thresholds for shrinkage were data-driven using the 90% level of the histogram of each Bregman distance. These values were determined so that images reconstructed from full view-shared k-space using our CS algorithm had a noise level similar to that of the images reconstructed using parallel imaging only. The maximum number of iteration was set to 30, with a stopping criterion if the normalized error between two consecutive estimates fell below 104. For comparison, the standard TWIST reconstruction was also performed on the same data set offline. In the
standard TWIST reconstruction, view-sharing combines each A and B pair with older B trajectories such that a regularly undersampled k-space is generated. Hence, traditional parallel imaging using SPIRiT alone was employed in a projection onto convex sets algorithm, which solves Equation 3: I1 ¼ argmin jjU1 F ðSI1 Þ K1 jj22 ; subject to I1 ¼ SI1
[3]
Phantom Study A phantom experiment was designed to demonstrate the effect of view-sharing on time-resolved CE-MRA. The experiment reproduces the scenario of using TWIST to image fast flow in smaller vessels to observe temporal blurring and validate the CS reconstruction of reduced view-shared data. A set of tubes was filled with GdDTPA (Magnevist; Bayer Healthcare Pharmaceuticals, Whippany, New Jersey, USA) diluted with saline. An air bubble was introduced using a clinical power injector (Spectris; Medrad, Indianola, Pennsylvania, USA) at 0.75 mL/s, giving an average velocity of 4–5 cm/s. A TWIST acquisition was performed throughout the injection to track the movement of the air bubble on a 3T scanner (Magnetom Tim Trio; Siemens Medical Solutions, Erlangen, Germany). The acquisition parameters were as follows: size of A region ¼ 5% of k-space; five B trajectories; temporal frame rate ¼ 1.25 s/frame; spatial resolution ¼ 1.3 1.3 1.3 mm3; matrix size ¼ 230 230 46; parallel imaging (19) acceleration of 2 2 (24 24 calibration lines); partial Fourier ¼ 6/8; and echo time (TE)/pulse repetition time (TR) ¼ 1.56/2.55 ms. Both TWIST (temporal footprint Tf ¼ 6.0 s) using Equation 3 and CS-TWIST (Tf ¼ 2.2 s) reconstructions using Equation 2 were performed based on the same dataset. To compare and validate the temporal behavior of the image
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
477
FIG. 3. Single in vivo timeresolved CE-MRA data were reconstructed using backward view-sharing TWIST (1–4 [one new B and four future B]), typical forward view-sharing TWIST (4-1 [four past B and one new B]), CS-TWIST (1-1 [one new B and one past B]) and no view-sharing with SPIRiT reconstruction SPIRiT (0–1 [one new B]). Thin coronal MIP images are shown by TWIST (4-1) and CS-TWIST (1-1) (top and bottom). Enhancement curves were plotted (middle row) for a carotid artery ROI (small vessel) and a jugular vein ROI (large vessel). The signal curves for the jugular vein ROI using the four image reconstruction techniques were similar. However, the signal curves for the carotid artery using the four reconstruction techniques were different. TWIST (1–4) and TWIST (4-1) both demonstrated temporal blurring with longer time duration of contrast enhancement than the CS-TWIST (1-1) and SPIRiT (0–1). Acquisition parameters: A ¼ 8%; five B trajectories; resolution ¼ 1.3 1.3 1.3 mm3; frame rate ¼ 1.7 s/frame.
series, the signal intensity curves at a cross-section along the tubing were generated based on TWIST and CS-TWIST, respectively, and the full-width half-maximum (FWHM) was calculated as the apparent time it took for the air bubble to pass the cross-section in the TWIST and CS-TWIST image series, respectively. The ground truth was also calculated by dividing the actual length of the air bubble by the expected flow velocity estimated based on the set volumetric flow rate of the power injector and the cross-sectional area of the tubing. In Vivo Study Because our image reconstruction technique can be applied to the k-space raw data of a standard TWIST acquisition, our in vivo study was performed retrospectively based on TWIST data acquired for nine patients who underwent the standard clinical TWIST CE-MRA imaging protocol for evaluation of thoracic (N ¼ 5) or intracranial blood vessels (N ¼ 4). The study was approved by our Institutional Review Board for retro-
spective analysis. The sequence parameters for the intracranial MRA scans were as follows: TE/TR ¼ 0.8–0.9/1.9– 2.1 ms; resolution ¼ 1.3 1.3 1.3 mm3; temporal frame rate ¼1.25 s/frame; parallel imaging (19) ¼ 3 2 with separate 24 24 calibration lines; partial Fourier (8/10 in ky and kz) and view-sharing (A ¼ 8%–14% and five B trajectories). The sequence parameters for thoracic MRA scans were as follows: TE/TR ¼ 0.9–1.0/2.5–2.6 ms; 1.1 1.1 3–6 mm3; temporal frame rate ¼ 2.35 s/frame; parallel imaging (19) ¼ 2 2 with separate 24 24 calibration lines; partial Fourier (8/10 in ky and kz) and viewsharing (A ¼ 25% and four B trajectories). For each acquired dataset, two image reconstructions were performed. First, a TWIST reconstruction involving full forward view-sharing (four to five B trajectories from past time points and one A) and parallel imaging using SPIRiT as shown in Equation 3 was performed. For demonstration purposes, backward TWIST was reconstructed from one dataset (five B trajectories from future time points and one A); Second, CS-TWIST reconstruction involving reduced view-sharing (two adjacent B
478
Rapacchi et al.
FIG. 4. View-sharing impacts the accurate visualization of the bolus timing. On the left, the temporal maximum intensity projections (coronal volume rendering up 75 mm) confirm minimal compromise in spatial resolution when replacing TWIST with CS-TWIST. However, the mean transit time (MTT, coronal volume rendering up 75 mm) maps (right) exhibit a significant timing offset (in carotids: 17.1 6 3.6 s for TWIST versus 16.5 6 3.4 s for CS-TWIST; P < 0.01) between TWIST and CS-TWIST because of view-sharing. Significant differences in MTT values were observed in smaller vessels (arrows) but not in larger vessels. Acquisition parameters: A ¼ 8%; five B trajectories; time resolution ¼ 1.25 s; resolution ¼ 1.4 1.1 1.3 mm3; parallel imaging ¼ 3 2.
trajectories and one A) was performed using the combined CS-SPIRiT technique shown in Equation 2. With the incorporation of CS to the TWIST reconstruction, the temporal footprint was reduced in intracranial MRA from 6 s to 2.2 s and in thoracic MRA from 5.6 s to 3.7 s. Image Analysis An important physiological parameter that can be extracted from time-resolved CE-MRA is the mean transit time (MTT). MTT is a surrogate for cerebral blood flow employed in nuclear imaging (20–22) that adds critical information in the work-up of stroke patients. To evaluate the impact of view-sharing on MTT and the benefit of temporal footprint reduction, quantitative MTT maps were computed from the two image series as Z I ðtÞt dt MTT ¼ Zt [4] I ðtÞ dt t
where I(t) is the signal intensity at time t. Two regions of interest (ROIs) were drawn, one at the common carotid artery, and one at the jugular vein. The mean MTT values within each ROI were calculated for comparison purposes. For visualization, MTT maps were masked using a two standard deviation threshold from the temporal maximum intensity projection. MTT maps provide comparison for spatial resolution and temporal resolution between the two techniques. The elongated temporal footprint could also introduce increased physiological motion blurring. To quantify spatial blurring introduced by view-sharing in chest MRA, vessel sharpness was quantified for both recon-
structed time series after mask (first volume reconstructed) subtraction. Vessel sharpness was defined as the sum of finite differences of each volume (23). A paired t test was performed to evaluate the significance of sharpness differences. The quality of the images reconstructed from the two different methods in comparison was graded in a blinded fashion by two radiologists trained in cardiovascular MRI. From a consensus reading session, the two readers were asked to evaluate overall image series quality for the purpose of clinical diagnosis. Evaluations were calibrated and blinded to the information related to subject and reconstruction method using a four-point scale (1 ¼ poor, 2 ¼ fair, 3 ¼ good, 4 ¼ very good). From the reconstructed images, maximum intensity projection (MIP) series were presented as uncompressed DICOM to the evaluators on a computer workstation side by side. Statistical significance of MTT and quality scoring was determined using a nonparametric Wilcoxon signed-rank test. Results were considered significant at P < 0.05. RESULTS Phantom Study The phantom experiment results in Figure 2 illustrate the impact of view-sharing on the tracking of the injected air bubble. With the standard TWIST reconstruction, the signal of the tube was only slightly diminished during the passage of the air bubble. This was because of the long temporal footprint that combines data from a time window that is much wider than the time it took the air bubble to pass each point along the tube. However, the reduction of view-sharing using CS-TWIST reconstruction allowed us to observe greater signal dropout during
Reducing View-Sharing Using CS in Time-Resolved CE-MRA
479
FIG. 5. Thin MIP (90 mm thick) of a thoracic CE-time-resolved CE-MRA experiment. The prolonged temporal footprint (Tf) due to viewsharing increased the potential of physiological motion blurring. Mixing data acquired within a shorter duration, CS-TWIST images show reduced motion blurring, which resulted in sharper enhanced blood vessels. Acquisition parameters: A ¼ 25%; four B trajectories; time resolution ¼ 2.35 s; resolution ¼ 0.9 0.9 6 mm3; parallel imaging ¼ 2 2.
the passage of the air bubble. The FWHM of the signal curves were 3.7 s and 5.6 s based on CS-TWIST and TWIST, respectively, while the ground truth for the expected time duration was 2.3 s. In Vivo Study Figure 3 demonstrates the impact of view-sharing in vivo on the carotid artery and jugular vein signal over time. Although the three approaches (backward TWIST [1–4], forward TWIST [4-1], and the proposed CS-TWIST [1-1]) yielded very similar signal intensity curves for the jugular vein, the signal for the carotid artery was dramatically different. TWIST (4-1), which mixes older k-space data with the current data, resulted in an apparent delay of the contrast wash-out. Backward TWIST (1–4), which mixes future k-space data with the current data, resulted in an early apparent wash-in. As contrast wash-in often represents clinical interest, forward view-sharing has been recommended (11). Figure 3 shows that the significant temporal blurring effect was observed primarily in the arteries (at approximately 20–23 seconds postinjection), when the contrast appeared to remain in the carotid arteries in the TWIST image, whereas it was
cleared out of the carotid arteries in the CS-TWIST images. Both TWIST (1–4) and TWIST (4-1) carotid artery curves showed longer signal enhancement, with an FWHM of 7.1 s for TWIST (1–4) and 9.8 s for TWIST (4-1) compared with 6.5 s for CS-TWIST (1-1). Figure 4 shows the difference of MTT quantification between standard TWIST and CS-TWIST time-resolved MRA. Across the patients imaged for intracranial MRA (N ¼ 4) in this study, the time values observed on the map from the CS-TWIST reconstruction had significantly lower MTT values in the carotid arteries compared with the MTT values extracted from the TWIST images series (17.1 6 3.6 s for TWIST versus 16.5 6 3.4 s for CS-TWIST; P < 0.01). However, there was no significant difference of MTT values observed in the jugular veins (20.1 6 1.3 s for TWIST versus 19.8 6 1.2 for CS-TWIST; P ¼ 0.3). Therefore, the elongated temporal footprint affects the smaller vessels and the larger vessels in different ways. The subjective image quality scoring (all patients, N ¼ 9) shows significantly improved diagnostic image quality by CS-TWIST when compared with TWIST (3.1 6 0.6 versus 2.4 6 0.3; P < 0.05). In time-resolved CEMRA of the thorax, the prolonged temporal footprint due to view-sharing resulted in spatial blurring of the vessels
480
that are affected with physiological motion (Figure 5). This blurring is reduced in the CS-TWIST reconstruction due to the much shorter temporal footprint. Based on five thoracic MRA patients in this study, vessel sharpness scores were improved by 8.8% 6 6.0 % (P < 0.01) using CS-TWIST compared with TWIST. DISCUSSION The results of our study suggest that the temporal blurring of time-resolved three-dimensional CE-MRA can be significantly improved by reducing view-sharing and shortening the temporal footprint. The reduced temporal footprint can also reduce spatial blurring effects from physiological motion. An algorithm combining a previously described compressed sensing algorithm with magnitude subtraction and the standard SPIRiT parallel imaging was proposed and validated in a small cohort of patients to reduce the temporal footprint of standard TWIST CE-MRA and its associated spatial and temporal blurring effects. We have shown that our CS-TWIST algorithm improves delineation of blood vessels and evaluation of the hemodynamics of contrast arrival for clinical time-resolved CE-MRA. With a detailed comparison between the proposed CS-TWIST algorithm and standard TWIST, we were able to demonstrate the impact of view-sharing on the CE-MRA images. The phantom experiment reproduced the scenario of imaging fast-traveling boluses that occur in relatively small structures. Because these small structures are mainly represented by the high spatial frequency region of the k-space, which is heavily shared over a long time, the visualization of small structures (e.g., a small blood vessel) is particularly compromised in standard TWIST reconstruction. To put it simply, view-sharing behaves similarly to a convolution of high-frequency information with a window that is larger than the frame rate. In clinical practice, view-sharing artifacts can be difficult to notice, because they depend on multiple parameters such as the size of the blood vessel, spatial resolution, flow velocity, and viewsharing parameters. Our in vivo results from clinical examinations (acquired as part of routine clinical care) show image artifacts related to view-sharing that are similar to our phantom results. Although it is well known that view-sharing causes temporal blurring, its effect on time-resolved CE-MRA has not received sufficient attention in the past. Furthermore, the possibility of missing a small vessel with fast blood flow (Fig. 3) and spatial blurring (Fig. 4) should also be considered when performing TWIST-type CE-MRA. Although CSTWIST uses fewer k-space data than standard TWIST reconstruction, it was still able to provide higher image quality, because reducing the temporal and spatial blurring caused by view-sharing was able to more than compensate any image quality degradation in the CS-TWIST algorithm. It is known that constrained reconstructions can cause loss of small image features. Therefore, it is possible that the signal loss of the carotid arteries based on both CS-TWIST and SPIRiT reconstructions (shown in Figure 3 at TT ¼ 23.3 s) could be due to an image recon-
Rapacchi et al.
struction bias, reduction in temporal footprint, or a combination of both factors. Future studies in larger patient cohorts are needed to separate the contributions to signal loss caused by these factors. CS reconstructions are also known to cause image blurring if certain sparsifying transforms (e.g., total variation) are used. This is evident in Figure 4, where the CS-TWIST image is slightly blurred compared with that for TWIST. However, in Figure 5, CS-TWIST resulted in sharper images than TWIST because of the presence of physiological motion. In this case, the motion blurring caused by long temporal footprint of the TWIST data was much greater than the slight blurring caused by the CS reconstruction algorithm. Finally, the impact of view-sharing on signal intensity could increase errors in quantification of hemodynamic parameters. Blurring from view-sharing can be both spatial and temporal and is nonuniform, as temporal biases depend on the image resolution and the size of vessels. For measurement of MTT, a major hemodynamic biomarker assessed with positron emission tomography (20), MRI may compete favorably as an ionizing radiation-free alternative if spatio-temporal sampling is appropriate. Currently, view-sharing stands as a limitation to the accuracy of MTT measured using timeresolved CE-MRA. Our study shows that the error in measuring MTT or other parameters derived from a perfusion model (6) may not be as simple as a bulk shift in the time-intensity curve, as the time shift in MTT also depends on the size and flow of each vessel relative to the imaging resolution. In conclusion, we have defined and evaluated several limitations of view-sharing in time-resolved CE-MRA and have implemented an image reconstruction method to mitigate them. Understanding the implications of view-sharing may be crucial for appropriate use of timeresolved MRA in complex vascular beds. CS is a potentially powerful tool to improve visualization of rapidly changing signal in small blood vessels. REFERENCES 1. Prince M, Meaney J. Expanding role of MR angiography in clinical practice. Eur Radiol 2006;16(suppl):B3–B8. 2. Hartung MP, Grist TM, Franc¸ois CJ. Magnetic resonance angiography: current status and future directions. J Cardiovasc Magn Reson 2011; 13:19. 3. Willinek WA, Hadizadeh DR, von Falkenhausen M, Urbach H, Hoogeveen R, Schild HH, Gieseke J. 4D Time-resolved MR angiography with keyhole (4D-TRAK): more than 60 times accelerated MRA using a combination of CENTRA, keyhole, and SENSE at 3.0T. J Magn Reson Imaging 2008;27:1455–1460. 4. Fenchel M, Saleh R, Dinh H, Lee MH, Nael K, Krishnam M, Ruehm SG, Miller S, Child J, Finn JP. Juvenile and adult congenital heart disease: time-resolved 3D contrast-enhanced MR angiography. Radiology 2007;244:399–410. 5. Krishnam MS, Tomasian A, Lohan DG, Tran L, Finn JP, Ruehm SG. Low-dose, time-resolved, contrast-enhanced 3D MR angiography in cardiac and vascular diseases: correlation to high spatial resolution 3D contrast-enhanced MRA. Clin Radiol 2008;63:744–755. 6. Wright KL, Seiberlich N, Jesberger JA, Nakamoto DA, Muzic RF, Griswold MA, Gulani V. Simultaneous magnetic resonance angiography and perfusion (MRAP) measurement: initial application in lower extremity skeletal muscle. J Magn Reson Imaging 2013;38:1237–1244. 7. Wang K, Schiebler ML, Francois CJ, Del Rio AM, Cornejo MD, Bell LC, Korosec FR, Brittain JH, Holmes JH, Nagle SK. Pulmonary
Reducing View-Sharing Using CS in Time-Resolved CE-MRA perfusion MRI using interleaved variable density sampling and highly constrained cartesian reconstruction (HYCR). J Magn Reson Imaging 2013;38:751–756. 8. Lim RP, Shapiro M, Wang EY, et al. 3D time-resolved MR angiography (MRA) of the carotid arteries with time-resolved imaging with stochastic trajectories: comparison with 3D contrast-enhanced boluschase MRA and 3D time-of-flight MRA. Am J Neuroradiol 2008;29: 1847–1854. 9. Nael K, Krishnam M, Ruehm SG, Michaely HJ, Laub G, Finn JP. Time-resolved MR angiography in the evaluation of central thoracic venous occlusive disease. Am J Roentgenol 2009;192:1731–1738. 10. Carroll TJ, Korosec FR, Swan JS, Grist TM, Frayne R, Mistretta CA. Method for rapidly determining and reconstructing the peak arterial frame from a time-resolved CE-MRA exam. Magn. Reson. Med. 2000; 44:817–820. 11. Mostardi PM, Haider CR, Rossman PJ, Borisch EA, Riederer SJ. Controlled experimental study depicting moving objects in view-shared time-resolved 3D MRA. Magn Reson Med 2009;62:85–95. 12. Johnson CP, Polley TW, Glockner JF, Young PM, Riederer SJ. Buildup of image quality in view-shared time-resolved 3D CE-MRA. Magn Reson Med 2013;70:348–357. 13. Trzasko JD, Haider CR, Borisch EA, Campeau NG, Glockner JF, Riederer SJ, Manduca A. Sparse-CAPR: highly accelerated 4D CEMRA with parallel imaging and nonconvex compressive sensing. Magn Reson Med 2011;66:1019–1032. 14. Lee GR, Seiberlich N, Sunshine JL, Carroll TJ, Griswold MA. Rapid timeresolved magnetic resonance angiography via a multiecho radial trajectory and GraDeS reconstruction. Magn Reson Med 2013;69:346–359.
481 15. Rapacchi S, Han F, Natsuaki Y, Kroeker R, Plotnik A, Lehman E, Sayre J, Laub G, Finn JP, Hu P. High spatial and temporal resolution dynamic contrast-enhanced magnetic resonance angiography using compressed sensing with magnitude image subtraction. Magn Reson Med 2014;71:1771–1783. 16. Song T, Laine AF, Chen Q, Rusinek H, Bokacheva L, Lim RP, Laub G, Kroeker R, Lee VS. Optimal k-space sampling for dynamic contrast-enhanced MRI with an application to MR renography. Magn Reson Med 2009;61:1242–1248. 17. Lustig M, Pauly JM. SPIRiT: iterative self-consistent parallel imaging reconstruction from arbitrary k-space. Magn Reson Med 2010;64:457–471. 18. Goldstein T, Osher S. The split Bregman method for L1-regularized problems. SIAM J Imaging Sci 2009;2:323. 19. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J, Kiefer B, Haase A. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 2002;47:1202–1210. 20. Sobesky J. Refining the mismatch concept in acute stroke: lessons learned from PET and MRI. J Cereb Blood Flow Metab 2012;32:1416– 1425. 21. Carrera E, Jones PS, Iglesias S, Guadagno JV, Warburton EA, Fryer TD, Aigbirhio FI, Baron J-C. The vascular mean transit time: a surrogate for the penumbra flow threshold? J Cereb Blood Flow Metab 2011;31:1027–1035. 22. Carrera E, Jones PS, Alawneh JA, et al. Predicting infarction within the diffusion-weighted imaging lesion: does the mean transit time have added value? Stroke 2011;42:1602–1607. 23. Deriche R. Fast algorithms for low-level vision. IEEE Trans Pattern Anal Mach Intell 1990;12:78–87.
