Highly time-resolved 4D MR angiography using golden-angle radial sparse parallel (GRASP) MRI

Current dynamic MRA techniques are limited by temporal resolution and signal-to-noise penalties. GRASP, a fast and flexible MRI technique combining compressed-sensing, parallel imaging, and golden-angle radial sampling, acquires volumetric data continuously and can be reconstructed post hoc for user-defined applications. We describe a custom pipeline to retrospectively reconstruct ultrahigh temporal resolution, dynamic MRA from GRASP imaging obtained in the course of routine practice. GRASP scans were reconstructed using a custom implementation of the GRASP algorithm and post-processed with MeVisLab (MeVis Medical Solutions AG, Germany). Twenty consecutive examinations were scored by three neuroradiologists for angiographic quality of specific vascular segments and imaging artifacts using a 4-point scale. Unsubtracted images, baseline-subtracted images, and a temporal gradient dataset were available in 2D and 3D reconstructions. Distinct arterial and capillary phases were identified in all reconstructions, with a median of 2 frames (IQR1-3 and 2–3, respectively). Median rating for vascular segments was 3 (excellent) in all reconstructions and for nearly all segments, with excellent intraclass correlation (range 0.91–1.00). No cases were degraded by artifacts. GRASP-MRI obtained in routine practice can be seamlessly repurposed to produce high quality 4D MRA with 1–2-s resolved isotropic cerebrovascular angiography. Further exploration into diagnostic accuracy in disease-specific applications is warranted.

Golden-angle RAdial Sparse Parallel (GRASP) is a relatively new technique for fast and flexible dynamic contrast-enhanced MRI that combines compressed sensing, parallel imaging, and golden-angle radial k-space sampling 15 . GRASP allows for continuous, rather than sequential, data acquisition and thus straightforward, user-defined post hoc reconstructions that can be optimized for application-specific use. Because radial sampling is inherently robust to motion, initial applications focused on dynamic visceral imaging to enable free-breathing examination of patients unable to suspend respiration. Further, because varying parenchymal enhancement phases could be reconstructed retrospectively from the continuously acquired data, the technique mitigates errors resulting from inaccurate timing of scans, hence simplifying clinical workflow significantly [16][17][18][19][20] .
While initial applications of GRASP used low temporal resolutions targeted for use in visceral imaging, we have observed that the continuous and uniform golden-angle radial sampling is particularly well-suited to capturing more rapid dynamic processes, owing to its potentially sub-second temporal resolution. Moreover, it relaxes the demand for precise synchronization of the bolus injection with sampling of the k-space center because each acquired k-space line covers the k-space center. By reconstructing the continuously acquired data with extreme under-sampling, volumes with a narrow temporal footprint can be generated for any time point within the acquisition window. Naturally, the lesser the data incorporated into the reconstruction, the larger the costs of reduced signal and the greater the potential vulnerability to artifacts. Nevertheless, we posit that the compressed-sensing reconstruction with robust temporal total-variation (TV) regularization and the strong sensitivity to gadolinium-induced T1 shortening are conditioned well to assessing dynamic cerebrovascular processes 15 . While past approaches have leveraged radial and other non-cartesian readouts with compressed sensing and golden-angle ordering to achieve uniform sampling, the application to large or real-world cohorts has yet to be fully realized [21][22][23] . The flexibility of GRASP as a technique easily implemented in routine practice, by comparison, allows for complementary TR-MRA reconstruction from data that is routinely acquired for volumetric T1-weighted structural imaging due to the high robustness to motion.
In this study we aim to demonstrate this possibility of utilizing GRASP for fully retrospective TR-MRA in a cohort of patients undergoing structural GRASP brain imaging in routine clinical practice. We hypothesize that GRASP bares untapped potential for simple and robust TR-MRA, which can be realized through a few clinically practicable post-processing steps. We present a custom reconstruction and visualization pipeline for creating multi-planar, subtracted, and time-resolved 1 mm isotropic whole brain MR angiography with 1-2 s temporal resolution.

Methods
Study subjects. After approval by the institutional review board (IRB) of NYU Langone Medical Center with waiver of informed consent, a retrospective query of the radiological information system was performed to identify subjects > 18 years of age with reportedly normal brain MRI obtained in routine practice with intravenous contrast. At our institution, GRASP is prescribed in standard brain MRI protocols during the first pass of gadolinium with reconstruction of 6-10 nominally time-resolved, post-contrast structural brain series, separated by ~ 20-25 s per volume depending upon the protocol. Immediately following the acquisition, raw k-space data from GRASP scans are exported to an external processing server within the hospital network, where retrospective GRASP processing can be performed using an offline reconstruction framework (Yarra Framework, https:// yarra-frame work. org). Parameters for the retrospective reconstruction, such as the number of radial views per volume (see below), can be selected using a graphical user interface under direction of the examiner.
All exams were obtained on clinical 3 T whole-body scanners (MAGNETOM Vida, MAGNETOM Trio, or MAGNETOM Skyra, Siemens Healthineers, Erlangen, Germany) using combined head/neck RF receive coils. A weight-based gadolinium dosing regime was employed using a high relaxivity agent (Gadobutrol, 0.1 mmol/ kg) injected at 4 ml/s and followed by a normal saline flush, which corresponds to the injection rate used at our institution for routine imaging. No a priori assumptions with respect to contrast kinetics or cerebral hemodynamics, nor modifications to the injection or acquisition protocol were made in any case. GRASP data were acquired using a fat-saturated 3D gradient-echo pulse sequence with slab-selective RF excitation and radial stack-of-stars sampling of k-space according to the golden-angle scheme 24 . The contrast agent was injected 20 s after the start of the sequence. Following sequence parameters were used for the scans: TR 3.65 ms; TE 1.73 ms; flip angle 12°; pixel bandwidth 630; radial views 600; field-of-view 256 mm; base resolution 1.0 mm × 1.0 mm × 2.0 mm, reconstructed to 1.0 mm isotropic voxels; slices 170-208 per slab (interpolated); slabs 1; total acquisition time 3:20. Exact acquisition parameters, including TR and slices per slab, varied slightly due to use of different MR systems and requirement for full anatomic coverage of the brain in all patients.
After initial screening of formal radiologic interpretations to identify reportedly normal exams, the studies were reviewed by a second-year neuroradiology fellow under the supervision of a subspecialty certified neuroradiologist with > 10 years of experience in neurovascular imaging. Cases with structural abnormalities, masses, developmental anomalies, or prior injury were excluded from analysis. Additional exclusions included the absence of full GRASP data sets or missing GRASP raw k-space data, failed gadolinium injections, and examinations unrecoverably corrupted by artifacts such as from orthodontic hardware or extrinsic devices and leads. The electronic medical record was reviewed to identify any unanticipated diagnoses or abnormalities in the final clinical interpretation that might affect cerebral hemodynamics (e.g., stroke, heart failure, peripheral vascular/atherosclerotic disease, etc.).
To preliminarily assess the potential utility of GRASP MRA in the setting of intracranial pathology, subjects with intracranial AVF or AVM and undergoing routine perioperative digital subtraction angiography (DSA) contemporaneous with their MRI (within 3 months) were identified. MRA processing. Angiographic GRASP images were reconstructed retrospectively on an external processing server using a C/C++ implementation of the GRASP algorithm. Reconstructions used 5 radial views per frame, producing an effective frame rate in the range of 1.58-2.07 s per whole 3D brain volume, depending on the acquisition parameters TR and number of slices. Using an Ubuntu Linux server with two Intel Xeon E5-2698 processors, a total of 32 cores, and 256 GB memory, the average reconstruction time per case was 3 h. Subtractions and further post-processing steps of the reconstructed images were performed using a custom pipeline implemented in a publicly available DICOM processing toolbox (MeVisLab, MeVis Medical Solutions AG, Bremen, Germany, https:// www. mevis lab. de) running on a separate workstation. Briefly, GRASP images were converted to the compact MeVisLab format to facilitate file handling and post-processing and were made available for review as contiguous, multi-planar reformatted (MPR) dynamic 2D slices and 3D rendered maximum intensity projections (MIP) with and without baseline subtraction in all cases. Baseline frames were selected by the user, typically between the 16th-18th temporal frame to ensure a clean steady state of the signal. In order to better isolate the parenchymal blood pool in the capillary phase and to recapitulate catheter-angiographic tissue hemodynamic signatures, dynamic images were also separately processed in MeVisLab by calculating the gradient of the subtraction images with respect to the time dimension and displaying the magnitude value of the gradient, which accentuates temporal changes in the subtraction images.
Image review and rating. Twenty examinations were evaluated by three subspecialty-certified neuroradiologists for image quality, vascular dynamic imaging/isolation of angiographic phases, and presence of reconstruction artifacts. Readers were instructed to identify peak arterial and venous enhancement phases, when sufficiently isolated, and to score the performance in 5 bilateral vascular segments including: (1) fair, evaluation possible but limited; (2) good angiographic quality; and (3) excellent angiographic quality 25 . High temporal resolution native (i.e. standard structural) GRASP and subtracted MPR (Fig. 1) and subtracted, 3D rendered MIP angiographic reconstructions (Fig. 2), as well as the intensity-based gradient MPR images (see below) were made available to readers on the workstation running MeVisLab. Raters reviewed images independently and subsequently generated a consensus result due to the complexity of the data structure. An initial test batch of three cases was compared to a contemporaneously acquired 3D TOF intracranial MRA to ensure that the rating methodology was harmonized across raters for each vessel segment of interest. While the concurrent availability of TOF-MRA and GRASP occurred in only a small subset of patients, the TOF in those subjects were used to establish a qualitative benchmark for standard-of-care arteriographic visualization and to ensure a uniform methodology among raters regarding vascular segments to be interrogated. Raters were further instructed to record the number of distinct arterial phases in all reconstruction algorithms,  In order to assess the potential influence of circulating blood volume (based on Nadler's formula) and its consequent modulation of circulating gadolinium concentration using weight-based dosing, subject height and weight were recorded, and any interactions with angiographic quality were determined 26 .

Statistical analysis.
Categorical variables are reported as number and percent. Continuous variables were tested for normality using visual inspection of histograms and/or the Shapiro-Wilk test. When not normally distributed, continuous variables were reported as median and interquartile range (IQR). Intraclass correlation coefficients (ICC) were calculated using two-way mixed-effects models to determine consistency between 3 readers and assessed using published thresholds 27 . Spearman correlation was used to determine the association between vessel scores and other continuous variables. A two-tailed p-value of < 0.05 was designated for statistical significance. Statistical analysis was performed using Stata version 12.1 (StataCorp, College Station, TX).

Results
Twenty patients without significant pathology identified on brain MRI were included in the final analysis from a cohort of 71 consecutive cases obtained at 3 T and with standard contrast dosing of 0.1 mmol/kg (subjects were excluded for tumor and/or postoperative changes n = 23, infection n = 15, inflammation/demyelination n = 10, and vascular malformation n = 3). Patient characteristics are summarized in Table 1. The median age of the cohort was 63 (IQR 32-73) with 15 (75%) female subjects. Using standardized contrast dosing, median contrast concentration was 6.9 (IQR 5.7-7.3). The most common indication for the brain MRI was evaluation for metastatic disease (70%), with other indications including headache, syncope, encephalopathy, and a suspected mass. The ratings of performance for all vascular segments by the three neuroradiologists are summarized in Table 2. For native, structural GRASP imaging volumes without subtraction, the median rating across readers was 3 (IQR 3-3) indicating excellent angiographic quality, with ICC of 1.00 (95% CI 1.00-1.00), for all arterial and venous segments of interest. Median rating of subtracted images was 3 (excellent) for all arterial and venous segments, with the exception of a score of 2 (good) rendered for the MCA-M4 segments.
ICC ranged from 1.00 (95% CI 1.00-1.00) in the M1, M2, and venous segments (excellent reliability) to 0.80 (95% CI 0.62-0.91) for the M3 segment (good reliability). Median rating for temporal gradient images (Fig. 3) was 3 (excellent) including for the depiction of a capillary/parenchymal phase of enhancement and angiographic performance, apart from the cavernous/supraclinoid ICA, for which a median of 2 (good) was recorded. ICC for gradient images was 1.00 (95% CI of 1.00-1.00) for all vascular segments, apart from the M2 segment, which was 0.93 (95% CI 0.85-0.97), suggesting excellent reliability. Median height in meters was 1.6 (IQR 1.6-1.7) and weight in kg was 69 (IQR 57-73), with no significant correlation identified between vascular segment ratings and either subject weight, height, sex, or contrast dose (p > 0.05) to suggest a direct dependency of angiographic performance on circulating blood volume.
All reconstructions were found to have a median of 2 distinct arterial phases (IQR 1-3). A median of 2 distinct frames of capillary arterial phase (IQR 2-3) were found in the hemodynamic-accentuated temporal gradient images. No artifacts were reported as degrading evaluation, including either motion, ghosting, aliasing or other artifacts related to parallel acceleration, with the exception of a sub-optimal subtraction in a single case, which did not affect angiographic quality. Streak-like signal spikes related to radial k-space sampling were observed in the image periphery in the majority of cases without degradation of imaging or angiographic quality.
A representative case of TR GRASP-MRA in the setting intracranial AVM (Figs. 4 and 5) demonstrates potential practical application of GRASP MRA to establish the presence or obliteration of residual arteriovenous  www.nature.com/scientificreports/ shunting. GRASP MRA reconstructed post hoc and using routine structural scanning parameters clearly demonstrates the residual small nidus and arteriovenous shunting, which were confirmed on DSA.

Discussion
Using an automated offline reconstruction and streamlined image processing pipeline, GRASP enables highly temporally-resolved dynamic angiography that can be obtained, even retrospectively, in routine practice without demand for further optimization of individual acquisition or timing parameters. Readily attainable 4D (i.e., time-resolved 3D) data sets were easily produced and yielded consistently good or excellent imaging in a threerater review across both proximal and distally arborized cerebrovascular segments. The golden-angle acquisition scheme of GRASP proved particularly well-suited to this objective because images with the desired temporal footprint can be flexibly reconstructed post hoc and for any desired time point, allowing for excellent depiction of enhancement dynamics. While other 4D dynamic MRA techniques have been previously developed, including highly accelerated non-cartesian techniques using golden angle sampling and compressed sensing, the flexibility of GRASP to be used as anatomical/structural sequence, producing pre-and post-contrast T1-weighted images, makes it highly suitable for routine application as reported here. The continuous radial stack-of-stars acquisition of k-space mitigates timing errors while motion and flow artifacts are inherently suppressed due to the robustness of the radial k-space trajectory 28 . After data acquisition, an arbitrary number of consecutive radial views can be combined into single temporal frames to create 4D angiography catered to specific applications, thereby relaxing the need for a priori determination of final temporal resolution. Even higher frame rates beyond the 1.58-2.07 s temporal resolution reconstructed in our study should be attainable simply by further reducing the number of radial spokes per volume, although our experience at those extremes remains limited.
Several existing techniques for TR-MRA, including TRICKS and TWIST, utilize center-weighted keyhole imaging while under-sampling the periphery of k-space 1-8 . However, view sharing may engender temporal smearing between ostensibly time-resolved volumes, confounding assessment for arteriovenous shunting.   www.nature.com/scientificreports/ Signal-to-noise penalties inherent to conventional TR-MRA place additional constraints upon attainable spatial resolution, which in our experience limits its utility for small shunts. Compressed sensing methods like the approach used here can, by comparison, enable acceleration beyond the Nyquist limit by exploiting sparsity of the information content of the acquired data, although the suitability of such techniques for routine clinical use, despite promising performance, remains to be established [20][21][22] . Together with its performance as a structural brain imaging technique, GRASP is therefore uniquely adaptable for routine clinical workflow and general canvas for ad hoc dynamic imaging applications. Several limitations to our study require discussion, including its retrospective nature. Importantly, however, the completely post hoc extraction of GRASP MRA in this cohort of brain imaging studies, all of which were obtained in routine practice for post-contrast structural T1-weighted imaging, underscores its versatility for dynamic imaging and secondary 4D angiography in a real-world setting. Similarly, sequence parameters were not modified or optimized for angiographic applications. In dedicate use as a prospective TR-MRA technique, optimizations including, for instance, greater flip angles to further accentuate vascular enhancement may be appropriate. Second, the standardized acquisition of our routine GRASP protocol in the axial plane introduces potential vascular inflow effects related to the steady state of the magnetization in the spoiled gradient-echo (fast low-angle shot, FLASH) scheme, in which non-saturated spins entering the inferior slices of the imaging volume may be hyperintense. This effect may be reduced through acquisition of sagittal 3D volumes as used commonly in other 3D brain imaging, or with use of non-selective excitations, although neither could be experimentally validated here due to the retrospective nature of the study. Nevertheless, subtraction of pre-contrast baseline images, which are similarly affected by inflow effects, provided excellent angiographic depiction and with faithful isolation of the arteriographic phase uncorrupted by inflow enhancement. The retrospective study design furthermore precluded any direct comparison between GRASP MRA and other angiographic techniques, with the exception of the limited initial comparison to TOF MRA during the reader calibration phase of the study. The number of vascular segments studied, including across multiple hemodynamic phases and using multiple reconstruction/rendering modes all in an entirely custom post-processing and visualization pipeline, also precluded a truly independent methodology for the reader ratings. However, as reported in similarly designed past studies, a consensus read from three subspecialized neuroradiologists represents a reasonable and generalizable initial experience from which to study the technique 2,29,30 . To limit variability, the current study analyzed exams only at 3 T and only using a single high-relaxivity contrast agent (gadobutrol). Imaging at lower field strength or with contrast agents exhibiting weaker relaxation enhancement could impact angiographic quality in non-trivial ways. In this respect, further exploration across varying imaging conditions is necessary to ensure generalizability 25,31 . The long calculation times currently required for post hoc GRASP reconstructions at extremely high temporal resolution, which can take several hours, presently pose a challenge for widespread use in acute settings. However, we anticipate that significant acceleration of the reconstruction can be achieved by using dedicated graphics processing units (GPUs), by utilizing cloud computing resources, or by using Deep Learning-based techniques, which are being explored by our group for future integration 32 . Similarly, the image postprocessing and visualization steps presently require direct user involvement in our pipeline, but can be shortened or potentially automated with tailored and optimized software tools. Most importantly, the actual diagnostic utility and accuracy of 4D GRASP MRA was not studied, and dedicated investigation in disease-specific applications is therefore necessary to establish its full diagnostic performance.
In conclusion, GRASP MRI obtained as part of routine clinical care can be simultaneously or even retrospectively utilized for production of 4D MRA with high spatial and temporal resolution, reliably demonstrating excellent angiographic quality and a low tendency for angiographic artifacts. The streamlined nature of the acquisition and the flexibility for post hoc angiographic reconstruction make it an attractive option for potential clinical TR-MRA. Further exploration into diagnostic accuracy in disease-specific applications is warranted.

Data availability
The datasets generated and/or analyzed during the current study are not publicly available due to reasons of sensitivity (human data) but are available from the corresponding author on reasonable request. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.