Contrast-Kinetics-Resolved Whole-Heart Coronary MRA Using 3DPR

Abstract

Slow contrast infusion was recently proposed for contrast-enhanced whole-heart coronary MR angiography (MRA). Current protocols use Cartesian k-space sampling with empirical acquisition delays, potentially resulting in suboptimal coronary artery delineation and image artifacts if there is a timing error. This study aimed to investigate the feasibility of using time-resolved three-dimensional projection reconstruction (3DPR) for whole-heart coronary MRA. With this method, data acquisition was started simultaneously with contrast injection. Sequential time frames were reconstructed by employing a sliding window scheme with temporal tornado filtering. Additionally, a self-timing method was developed to monitor contrast enhancement during a scan and automatically determine the peak enhancement time around which optimal temporal frames were reconstructed. Our preliminary results on 6 healthy volunteers showed that by using time-resolved 3DPR, the contrast kinetics of the coronary artery system throughout a scan could be retrospectively resolved and assessed. In addition, the blood signal dynamics predicted using self-timing was closely correlated to the true dynamics in time-resolved reconstruction. This approach is useful for optimizing delineation of each coronary artery and minimizing image artifacts for contrast-enhanced whole-heart MRA.

INTRODUCTION

Contrast-enhanced coronary MR angiography (MRA) has been proven to be a promising technique for high-quality anatomic imaging of coronary arteries (1–5). Bright blood visualization of coronary arteries can be obtained with injection of T1-shortening contrast agents. Conventionally, the contrast agent is administrated at antecubital fossa in ~20 sec and a targeted coronary artery segment is scanned subsequently within a breathhold or using respiratory gating during the first passage of the contrast medium (1–5). Due to the time needed for the contrast agent to transit from the administration site to the coronary artery system, data acquisition must start with a proper delay after contrast injection to synchronize with coronary artery enhancement. Usage of an improper acquisition delay can lead to suboptimal signal-to-noise ratio (SNR) and image artifacts. However, the transit time of the contrast agent is strongly subject-dependent and therefore needs to be estimated by conducting an additional test bolus exam on each subject (1,2,5).

Recently, Bi et al proposed a slow contrast injection scheme for whole-heart coronary MRA (6). Double-dose Gd-BOPTA was slowly infused in ~2 min to achieve prolonged blood enhancement duration, enabling high spatial-resolution imaging of the entire coronary artery tree within a single measurement. An empirical acquisition delay of 25 (6) or 60 (7) seconds was used. However, during the relatively long scan time needed for whole-heart imaging, blood signal in coronary arteries inevitably exhibits considerable variations due to filling and flushing-out of the contrast agent. Theoretically, to optimize SNR, central k-space lines should be acquired during maximal coronary artery enhancement. Nonetheless, contrast kinetics can vary significantly with different injection rates and doses among subjects. Conventional test bolus with fast contrast injection produces enhancement timing inconsistent with that in coronary imaging, while test bolus performed with slow infusion results in undistinguishable blood enhancement. As a result, it is difficult to accurately estimate the temporal delay of maximum coronary artery enhancement using conventional test bolus in such a scenario. Furthermore, for the conventional method (6,7), the actual acquisition time of central k-space lines is always subject to variations of the subject's breathing pattern and heart rate during a scan and therefore can not be reliably scheduled in advance.

Another challenge for the conventional method is myocardial signal suppression. For volume-targeted coronary MRA with fast contrast injection, there is limited signal enhancement in myocardium during the first contrast passage, which can be effectively suppressed using inversion recovery (IR) pulses with a proper IR delay time (4,5). However, during a prolonged scan for whole-heart coronary MRA, greater myocardial signal enhancement is expected. The concomitant myocardium enhancement can reduce blood-myocardium contrast and obscure coronary artery visualization. This further complicates the optimization of the data acquisition scheme.

Time-resolved imaging techniques have been developed for contrast-enhanced MRA of blood vessels other than coronary arteries (8–10). Time-resolved imaging can accommodate various contrast arrival times by collecting sequential time frames and therefore are more robust than test-bolus approaches (11,12). However, due to the repetitive sampling of Cartesian k-space required for time-resolved imaging, these techniques have intrinsic limitations in achievable spatial resolution and coverage and thus can not be fitted for high-definition whole-heart coronary MRA. Alternatively, Barger et al proposed a time-resolved imaging technique based on 3D Projection Reconstruction (3DPR) and verified its feasibility in contrast-enhanced MRA of pulmonary and abdominal vasculature (13). This new approach requires only one measurement for high-resolution and broad-coverage imaging. Sequential time frames can be reconstructed using temporal tornado filtering and those frames during the arterial phase without venous overlapping can be selected retrospectively.

The objective of this study was to investigate the feasibility of contrast-kinetics-resolved whole-heart coronary MRA using 3DPR. The original acquisition scheme and tornado filter in (13) were modified for the new application with ECG triggering, respiratory gating and slow contrast infusion. Additionally, a self-timing method was developed to enable automatic selection of optimal time frames for image reconstruction.

MATERIALS AND METHODS

Data Acquisition

For this study, a new sequence was developed by implementing 3DPR in a conventional segmented spoiled-gradient-echo sequence with electrocardiogram (ECG) triggering and diaphragmatic navigator (NAV) gating. Nonselective IR pulses were applied prior to each segmented data acquisition to suppress background signals. Similar to (25), a 3D radial trajectory providing nearly isotropic k-space coverage is calculated based on the total number of k-space projections (N_P). These radial projections are divided into N_i interleaves, where each interleaf is comprised of N_s (N_s=N_P/N_i) projections and provides a nearly even angular coverage in 3D k-space. For instance, the first interleaf consists of projection 1, N_i+1, 2·N_i+1, … (N_s-1)·N_i+1 and the second interleaf consists of projection 2, N_i+2, 2·N_i+2, … (N_s-1)·N_i +2. As demonstrated in Fig. 1, 3D radial k-space was filled in a time-interleaved fashion for time-resolved imaging (13). In each heartbeat, a pre-described interleaf of k-space projections was collected. Data acquisition started at the same time as contrast injection. After the first acquisition, a part of k-space projections were resampled starting from the first projection interleaf to approximately cover the entire contrast kinetics in coronary arteries.

Fig. 1.

The upper part demonstrates the data acquisition scheme. Acquisition of 3D k-space projections was performed in a time-interleaved order and was repeated within 9 minutes. Each subset of 3D radial lines represents an interleaf of k-space projections collected in a heartbeat. The bold dashed line indicates completion of the first acquisition, after which data acquisition restarted from the first k-space interleaf. The lower part demonstrates contrast-kinetics-resolved reconstruction. Each dotted row represents a collected k-space projection. Only those samples covered in the grey area of the tornado filter were used for image reconstruction. The two double-headed horizontal arrows indicate the outer k-space window (~6 min) and center k-space aperture (~20 sec) of the tornado filter. P_OW and P_CA were slid for reconstruction of different time frames.

Fig. 3.

An axial slice at root RCA (white arrow) reconstructed without tornado filtering (a) and using the optimal tornado filter (b). Notice the enhanced blood signal intensity and slightly increased streaking artifact level in b). c) shows the reconstruction from the same data as b) using the iterative artifact-suppressing algorithm. Notice the reduced streaking artifact level in c). d) shows the vessel profile at a line crossing RCA measured in a) (dashed gray), b) (dashed black) and c) (solid black). The dotted line in a) indicates the position for the profile evaluation.

Human studies were conducted on 6 healthy volunteers using a 1.5T clinical MRI scanner (MAGNETOM Avanto, Siemens Medical Solutions, Erlangen, Germany). In accordance with our institutional review board, informed written consent was obtained from each volunteer before a study. At the beginning of each study, the heart and diaphragm were localized using a low-resolution scouting scan with standard views. A cine scouting scan was then performed at a 4-chamber-view orientation, from which cardiac motion was assessed and the optimal acquisition window in a cardiac cycle was determined. Next, coronary MRA was performed using the 3DPR sequence described above with the NAV beam placed at the right hemi-diaphragmatic dome. An empirical correction factor of 0.6 was used for slice tracking (14). A thick slab covering the entire heart was scanned using the following parameters: 350×350×350 mm³ FOV, 320 readout points; 1.1×1.1×1.1 mm³ isotropic spatial resolution (interpolated into 0.55×0.55×0.55 mm³); 42 projections/heartbeat; 20° flip angle; TR/TE = 3.3/1.78 ms; TI = 120ms. A k-space dataset consisting of 8400 projections was collected in ~6 min and was partially recollected in 9 min (Fig. 1). This corresponds to an undersampling factor of 19 relative to the Nyquist requirement for radial sampling. Gd-BOPTA (0.2 mmol/kg) was slowly infused at 0.3ml/sec, chased by 20 ml saline flush at the same rate (6).

Image Reconstruction

The reconstruction method used in this study was modified from the method described in (13) and is graphically demonstrated in Fig. 1. Both sliding window and tornado filtering were performed along time in data accepted by NAV gating for contrast-kinetics-resolved imaging. Density compensation functions were calculated based on the k-space locations of the samples selected for reconstruction using an iterative procedure as described in (16).Undersampling artifacts were suppressed using an iterative algorithm based on the angular distribution of streaking artifacts in projection reconstruction (15).

Although the optimal acquisition delay was undetermined before a scan, data acquisition covered the entire contrast kinetics with k-space projections partially resampled. Therefore, temporal sliding window could be used to select different sets of projections with different positions of outer k-space window (P_OW) corresponding to different acquisition delay times.

Meanwhile, temporal tornado filtering could be used to reconstruct time-resolved images during a scan. Instead of performing filtering in the entire k-space, this study applied tornado filtering only in the inner half k-space. As shown in Fig. 1, all samples in the outer half k-space of the selected k-space dataset were used for image reconstruction. However, in the inner half k-space, only those samples included in a tornado-shaped temporal filter were accepted. The position of the narrow center k-space aperture (P_CA) was slid in the temporal direction for reconstruction of different time frames. For 3DPR, signal intensity and tissue contrast are primarily determined by the acquisition time of center k-space samples or P_CA, while streaking artifacts are caused by undersampling in outer k-space. Therefore, this modified filter is capable of effectively resolving contrast kinetics without significantly elevating the artifact level, as compared to the original tornado filter (13).

In this study, we chose a temporal width of 200 and 10 consecutive NAV-accepted heartbeats for the outer k-space window and center k-space aperture. The former corresponds to acquisition a full k-space dataset of 8400 projections. The current temporal width for center k-space aperture was determined empirically, providing an effective temporal resolution of ~20 seconds (assuming a heart rate of 60 beats/min and a NAV acceptance of 50%). This temporal resolution is presumably sufficient for resolving the slow contrast change with the current injection scheme. Using the self-timing method described in the next section, we selected start times for the outer and inner windows, to coincide with the maximal signal enhancement. For simplicity, P_OW and P_CA will be used to represent the start times of the outer k-space window and the center k-space aperture relative to contrast injection, respectively.

Self-Timing

The modified tornado filter has two variables, P_OW and P_CA. Considering the large number of possible combinations of these two positions, reconstructing and reviewing all time frames would be formidably time-consuming. To address this problem, a self-timing method was developed for automatic design of the optimal tornado filter.

Similar to the method proposed by Peters et al for contrast enhancement estimation in hybrid 3DPR (17), the mechanism of our self-timing method was based on properties of radial k-space sampling and slow contrast infusion. First, for 3DPR, k-space center was repeatedly measured in each projection. The magnitude of the center k-space sample reflects instantaneous signal integration within the entire FOV. Because signals from background tissues (thoracic trunk, abdominal organs, et al) are substantially suppressed by IR and are relative stationary in contrast-enhanced studies, change of this center k-space magnitude is primarily induced by dynamics of blood signal in the heart. Second, with slow contrast infusion, cardiac blood enhancement sustains for a prolonged duration lasting for several minutes and exhibits slow variations. On the other hand, the delay between enhancement of right and left cardiac chambers due to pulmonary circulation is on the order of several seconds (18) and thus is relatively negligible. Therefore, the dynamics of cardiac blood signal can be approximated as overall signal increase or decrease in the whole heart, including the four cardiac chambers and the coronary artery system. Based on the above two properties, we assumed that the change of center k-space magnitude is synchronized with cardiac blood enhancement and thereby the enhancement of coronary arteries. Accordingly, integration of repetitively sampled center k-space magnitude over a temporal window is representative of the overall enhancement of coronary arteries obtained within the integration window.

Self-timing consisting of the following 3 steps was performed to determine the optimal values for P_OW and P_CA.

Step 1

Calculating the self-timing signal: Signal variation over heartbeats during a scan could be estimated by summing up all center k-space magnitude measured in each heartbeat. As shown in the example in Fig. 2.a, the resulting curve depicts signal variation induced by contrast injection but is still contaminated by high-frequency artifacts. These high-frequency components were primarily caused by breathing and cardiac motion fluctuations, which are of significantly higher frequencies than slow contrast kinetics, and therefore can be removed by low-pass filtering (pass band: 0~0.015 Hz). The smoothed signal after low-pass filtering represents the dynamics of the cardiac blood signal during the scan and those samples in NAV-accepted heartbeats, defined as the self-timing signal, were used in the following two steps.

Fig. 2.

Self-timing procedure. a) shows the original (thin gray) and low-pass-filtered (solid bold) signal variation calculated from repetitive measurement of center k-space. b) shows the self-timing signal extracted from the smoothened signal variation in a). c) shows integration over outer k-space window with different P_OW’s. d) shows integration over center k-space aperture with different P_CA’s. The dotted lines in c) and d) indicate the detected optimal P_OW and P_CA, respectively.

Step 2

Selecting the optimal P_OW: For 3DPR, each projection travels through k-space center and can be regarded of approximately equal contribution to the image SNR. Therefore, the optimal P_OW should correspond to the outer k-space window covering the set of projections with maximum overall coronary artery enhancement. Overall coronary artery enhancement obtained with a P_OW could be estimated by integrating self-timing samples within the corresponding outer k-space window (200 consecutive heartbeats starting from P_OW). The P_OW resulting in the maximum outer-window integration was selected as the optimal P_OW.

Step 3

Selecting the optimal P_CA: Choice of P_CA determines the acquisition time of center k-space samples and thereby blood signal intensity and tissue contrast in the reconstructed image. Similar to step 2, self-timing samples within different center k-space apertures (10 consecutive heartbeats starting from P_CA) were integrated to estimate coronary artery enhancement in different time frames. The P_CA providing the maximum center-aperture integration was regarded as the optimal P_CA, which corresponds to the time frame with predicted peak SNR for coronary artery visualization. Due to concomitant enhancement of myocardium, optimization of the contrast-to-noise ratio (CNR) is relatively complicated and will be discussed later in this paper.

Analysis

The calculated optimal P_OW was used to select a set of k-space projections for image reconstruction. Next, by sliding P_CA in the temporal direction, contrast-kinetics-resolved images with different blood signal intensity and tissue contrast were reconstructed. Coronary artery delineation was evaluated in the reconstructed time frames based on calculations of SNR and CNR. Signal intensity of blood and myocardium was estimated as the average signal at the root of left artery descending (LAD) and in a left ventricular area adjacent to LAD, respectively. For 3DPR, besides intrinsic noise from data acquisition, sub-Nyquist sampling in outer k-space generates noise-like streaking artifacts in reconstructed images. Stochastic noise and noise-like streaks can both adversely affect image quality and were considered together as pseudo-noise in image quality evaluations in this study. The level of this pseudo-noise was measured by calculating signal standard deviation in an area of air outside of thoracic trunk (19). All the above calculations were performed at the same positions in all time frames for each volunteer. Pseudo-SNR (pSNR) and pseudo-CNR (pCNR), defined as blood signal intensity and the difference between blood and myocardium signal intensities divided by the pseudo-noise level respectively, were calculated in all reconstructed time frames. These measurements were compared with self-timing prediction to investigate the effectiveness of using self-timing to select optimal time frames.

RESULTS

Fig. 2.a shows the original signal variation (thin gray) derived from the change of center k-space magnitude during a coronary MRA scan. After low-pass filtering, high-frequency artifacts are eliminated and the smoothened curve (bold black) depicts dynamics of the cardiac blood signal induced by contrast injection. Those samples in NAV-accepted heartbeats are extracted for self-timing reconstruction (Fig. 2.b). Fig. 2.c and d show integration of the self-timing signal in Fig. 2.b over different outer k-space windows and center k-space apertures, respectively. According to these two integration curves, the 62nd outer k-space window (P_OW = 108 sec) results in data acquisition with maximum overall contrast enhancement and the 85th center k-space aperture (P_CA = 150 sec) corresponds to the time frame with the highest cardiac blood signal, respectively.

Fig. 3.a and b shows an axial slice at the root of right coronary artery (RCA) obtained from the same scan reconstructed using a full k-space dataset (P_OW = 108 sec) and the designed optimal tornado filter (P_OW\P_CA = 108\150 sec), respectively. Fig. 3.b clearly exhibits much higher blood enhancement as well as slightly increased streaking artifacts. The same data used for Fig. 3.b were reprocessed using the iterative streak-suppressing algorithm. As shown in Fig. 3.c, the new method effectively reduced the streaking artefact level throughout the entire image without compromising coronary artery depiction. As illustrated in the vessel profile plot (Fig. 3.d), the optimal tornado filter substantially increases the signal intensity at the measured RCA segment and the iterative algorithm preserves the vessel profile very well.

Fig. 4.a shows reformatted contrast-kinetics-resolved images of RCA from the same subject reconstructed with the optimal P_OW and different P_CA’s. The first 9 time frames were obtained during the first 4 min after contrast injection with a temporal interval of ~30 sec. To characterize the entire contrast kinetics, two later time frames (P_CA = 375 and 510 sec) are also displayed. Contrast kinetics of RCA is captured in this image series. Fig. 4.b shows the RCA image reconstructed from the same set of k-space projections without applying tornado filtering. RCA in Fig. 4.b exhibits much lower signal intensity than the time frames in between 90 and 210 sec but higher intensity than the other earlier or later time frames. Signal intensities of coronary artery blood (stars) and myocardium (circles) measured in different time frames are plotted in Fig. 4.c. Clearly, the entire contrast kinetics consists of three distinctive phases: 1. ‘pre-enhancement phase’ during which there is insufficient contrast agent in cardiac blood to cause visible contrast enhancement (0~30 sec); 2. ‘ascending phase’ during which the contrast agent accumulates leading to increasing blood enhancement (30~120 sec); 3. ‘descending phase’ during which blood enhancement decreases as the contrast agent washes out (>120 sec). Noticeably, myocardium signal also increases substantially during the scan slightly later than blood enhancement. By visual comparison, the contrast enhancement predicted using self-timing (bold solid) is closely synchronized with the measured blood signal dynamics with a slight temporal delay. In this example, peak pSNR was achieved at P_CA = 120 sec, whereas peak pCNR was reached at P_CA = 90 sec. The dashed and dotted lines in Fig. 4.c represent signal intensities of coronary artery blood and myocardium in full k-space reconstruction (Fig. 4.b), respectively. Obviously, they both approximate the average enhancement within the entire outer k-space window. The level of pseudo-noise exhibits very small variations in different time frames (3.94±0.19) and is 7% higher than that in Fig. 4.b. However, compared to Fig. 4.b, blood signal intensity in the peak-pSNR time frame is increased by 26% and accordingly pSNR is increased by 24% in this time frame. The peak-pCNR time frame provides a pCNR increment of 18% relative to Fig. 4.b.

Fig. 4.

RCA images of contrast-kinetics-resolved reconstruction (a) and full k-space reconstruction (b). The number at the left-lower corner of each image in a) indicates the P_CA of that time frame (unit: sec). Note the enhancement and decay of blood signal in contrast-kinetics-resolved imaging. c) shows the predicted contrast enhancement (solid bold) and measured signal intensity of coronary arteries (stars) and myocardium (circles) in different time frames. The dashed and dotted lines represent signal intensity of coronary arteries and myocardium in full k-space reconstruction, respectively.

Fig. 5.a shows the predicted contrast enhancement and measured signal intensities in another subject. The signal change in coronary artery blood and myocardium exhibits a pattern similar to the previous case and the predicted contrast kinetics nicely follows the measured dynamics of blood enhancement. In this case, peak pSNR and pCNR are achieved at P_CA = 135 and 111 sec, respectively. Coronary artery images obtained from this subject are shown in Fig. 5.b and c. Image reconstruction was performed during the predicted ascending phase (50~160 sec) with a temporal interval of ~25 sec. For comparison, a pre-enhancement frame (P_CA = 0 sec) and two frames during the descending phase (P_CA = 211 and 511 sec) are also shown. Enhancement of coronary arteries and myocardium can be clearly appreciated in these time frames. The peak-pSNR frame (P_CA = 135 sec) provides the best visualization of RCA (Fig. 5.b). Although with slightly reduced pSNR, an earlier frame (P_CA = 111 sec) provides higher pCNR and better depiction of LAD (Fig. 5.c).

Fig. 5.

a) Contrast enhancement predicted using self-timing (solid bold) and true signal intensity of coronary arteries (stars) and myocardium (circles) measured in contrast-kinetics-resolved imaging. b) and c) show RCA and LAD images in different time frames, respectively. Frames 2~5 in b) and c) were reconstructed during the ascending phase of the predicted contrast enhancement. The first and the last two images show a pre-enhancement and two post-peak-enhancement time frames. The numbers on the images show the P_CA values of the corresponding time frames (unit: sec). Note the apparent myocardium enhancement in c).

Another example is shown in Fig. 6. Four time frames were reconstructed during the ascending phase (50~135 sec) detected in the predicted contrast kinetics. The progress of contrast enhancement in LAD and RCA can be clearly observed. Maximum blood enhancement and best visualization of RCA were achieved at P_CA = 112 sec (Fig. 6.a). However, similar to the previous case, LAD in this time frame is obscured by myocardium enhancement and its best delineation was achieved in an earlier frame (P_CA = 86 sec) with higher blood-myocardium contrast (Fig. 6.b).

Fig. 6.

RCA (a) and LAD (b) images reconstructed during the predicted ascending phase (left to right: P_CA = 60, 86, 112, 136 sec). Best overall delineation of RCA and LAD are achieved at 112 sec and 86 sec, respectively.

Table 1 summarizes the pSNR, pCNR, P_CA and P_OW for full k-space reconstruction (without tornado filtering) and the peak-pSNR and peak-pCNR frames from contrast-kinetics-resolved reconstruction (with tornado filtering). Compared to full k-space reconstruction, contrast-kinetics-resolved reconstruction increases both pSNR and pCNR substantially in both selected frames. However, peak-SNR and peak-CNR are not achieved in the same time frame and the latter comes earlier. Among different subjects, the optimal P_OW and the P_CA corresponding to peak-SNR and peak-CNR exhibit large standard deviations.

Table 1.

comparison of full k-space and contrast-kinetics-resolved reconstructions

	without tornado filtering	with tornado filtering
		peak pSNR frame	peak pCNR frame
pSNR	36.16 ± 10.13	44.92 ± 10.56	40.52 ± 9.61
pCNR	23.58 ± 8.99	28.55 ± 8.26	30.75 ± 7.99
PCA (sec)	N/A	116.83 ± 10.53	91.00 ± 10.43
POW (sec)	105.38 ± 9.73	105.38 ± 9.73	105.38 ± 9.73

DISCUSSION

Results presented in this study illustrated that blood signal still varies substantially during slow contrast infusion used for whole-heart coronary MRA. Furthermore, the delay to peak contrast enhancement exhibits significant inter-subject variability. These observations indicate that coordinating data acquisition with contrast enhancement needs to be addressed for each subject. This work developed a new approach enabling retrospective reconstruction of sequential time frames, capturing the entire contrast kinetics during a scan. With this capability, data acquisition can be simply started simultaneously with contrast injection, eliminating the difficulty in predicting contrast enhancement needed for conventional protocols (6,7).

Noticeably, myocardium also presents significant enhancement during contrast-enhanced whole-heart coronary MRA. This concomitant myocardium enhancement generally occurs slightly later than cardiac blood enhancement, due to the additional time needed for the contrast medium to perfuse into myocardium. Accordingly, peak blood-myocardium contrast arrived earlier than peak blood enhancement in this study. Ideally, selection of the optimal time frame should be specific to coronary artery segments residing in various environments. For segments primarily surrounded by epicardial fat (e.g. medial RCA), vessel delineation is dominated by SNR. However, for those adjacent to myocardial muscle (e.g. proximal LAD), both SNR and CNR must be considered and usually better visualization is obtained in the peak-CNR frame using the current injection scheme.

This work also presented a self-timing method to guide selective reconstruction of optimal time frames. By blood signal dynamics estimated using the self-timing method and measured in time-resolved images, we can see that the change of center k-space magnitude is closely synchronized with true blood enhancement in coronary arteries. Nonetheless, a slight temporal delay of 20~30 sec between them was also noted. This delay is presumably caused by late enhancement of background tissues, mostly in abdominal organs covered by the whole-heart imaging volume. Considering this delay and the concomitant myocardium enhancement, a robust reconstruction scheme is to reconstruct a few (usually ~4) frames during the ascending phase of the predicted contrast kinetics, defined as the period between the estimated moments of initial and peak contrast enhancement. These time frames should cover the entire progress of contrast filling in coronary arteries and thereby both peak-SNR and peak-CNR frames.

Compared with conventionally Cartesian k-space sampling, 3DPR employed in this study intrinsically increases noise variance due to its nonuniform k-space sampling (20). However, the proposed method also possesses a few desirable features for coronary MRA. First, radial sampling is less sensitive to motion than Cartesian sampling (21), which can mitigate motion artefacts for free-breathing imaging (22,23). Second, 3DPR generates true isotropic spatial resolution, which is beneficial to offline reformatting for coronary artery visualization (24). Furthermore, undersampling in radial k-space manifests as noise-like streaks in image domain instead of malign foldover artefacts. A gain in imaging speed can be achieved by reducing the projection number with a tolerable compromise in image quality (23,25). For contrast-enhanced coronary MRA, 3DPR is especially insusceptible to highly undersampling in k-space, because thoracic trunk and abdominal tissues, the major source of streaking artifacts in the region of the heart, are substantially suppressed by IR pulses. Moreover, combining with tornado filtering enables retrospective coordination of data acquisition with contrast enhancement. In contrast, the conventional Cartesian sampling-based approach (6,7) suffers from suboptimal SNR and CNR due to usage of fixed empirical acquisition delays. Also, it requires a longer scan time for the same spatial resolution and coverage, which entails slower contrast injection and consequently further sacrifices SNR.

Usage of tornado filtering can potentially increase both streaking artifacts and noise compared to unfiltered reconstruction including all collected data. However, the increase of pseuso-noise measured in our studies was minor (6.72±1.57%) due to the following two reasons. First, by preserving all outer k-space samples, the streaking artifact level is similar to unfiltered reconstruction. Second, tornado filtering reduces the unevenness of k-space sampling, which can to some extent offset the reduced number of k-space samples in reconstruction and the resulting increase in noise variance (20). Furthermore, from the prospective of blood signal intensity, tornado filtering can target reconstruction at peak contrast enhancement, contrary to averaging signal throughout the entire outer k-space window for unfiltered reconstruction. Therefore, as illustrated in this study, time-resolved reconstruction can improve both pSNR and pCNR.

In our current work, data acquisition was repeated within a fixed long duration of 9 min to ensure covering the optimal acquisition window. The optimal dataset was selected offline using sliding window. Clearly, this scheme leads to redundant data acquisition. Alternatively, self-timing signals can be calculated in real-time. Detection of the optimal P_OW indicates complete acquisition of the optimal dataset and can be used to automatically terminate a scan.

In conclusion, this work demonstrated the feasibility of a time-resolved 3DPR approach for contrast-enhanced whole-heart coronary MRA. This new method enables reconstruction of contrast-kinetics-resolved images and retrospective selection of the optimal time frame for each coronary artery segment. Also, a self-timing signal extracted from imaging data can be employed to monitor cardiac blood enhancement and automatically determine the optimal period for image reconstruction. The proposed method is promising for resolving acquisition-timing-related challenges in conventional contrast-enhanced coronary MRA.