Introduction
Time-resolved vessel-selective cerebral angiography provides crucial information on arterial morphology, hemodynamics and flow patterns. However, the current clinical gold standard, X-ray digital subtraction angiography, requires the use of an invasive procedure and injection of a contrast agent, resulting in some risks to the patient [
1]. Non-invasive alternatives, based on arterial spin labeling (ASL) MRI, show promise, but may be prohibitively slow for many clinical settings because additional measurements are required to obtain vessel-selective information from multiple arteries [
2‐
6]. The investigation of accelerated vessel-selective ASL methods is, therefore, warranted, since these could allow phenomena such as occlusions, stenoses, collateral flow and blood supply to lesions to be visualized non-invasively within a clinically acceptable time frame.
Acceleration inevitably leads to loss of signal-to-noise ratio (SNR), so in order to achieve significant scan time reductions it is important that the SNR efficiency is as high as possible. Vessel-encoded pseudocontinuous arterial spin labeling (VEPCASL) [
7] is considerably more SNR efficient than single-artery selective methods [
8,
9] when multiple arteries are of interest, since all arteries contribute signal to all measurements, such that the SNR efficiency is comparable to non-selective ASL [
10]. In addition, it has been shown that VEPCASL can be combined with a balanced steady-state free precession (bSSFP) readout [
11], in which transverse magnetization is “recycled” from one excitation to the next [
12], unlike spoiled gradient echo (SPGR) techniques, further improving SNR-efficiency in angiographic acquisitions. However, bSSFP suffers from sensitivity to magnetic field inhomogeneity [
12]. This can lead to significant ASL signal loss when imaging is performed over large regions where B
0 shimming is more challenging [
11].
If high SNR efficiency can be achieved, then acceleration through undersampling should be possible whilst maintaining reasonable image quality. However, undersampling with traditional Cartesian trajectories results in strong, coherent ghosting artifacts unless the data are reconstructed with parallel imaging algorithms [
13,
14]. This requires additional scan time for calibration (acquired as a single pre-scan for the whole examination or specifically for each scan) and results in noise amplification, which is typically most severe at the center of the head where many cerebral vessels of interest reside. Radial acquisition schemes can be angularly undersampled, resulting in aliasing artifacts that take the form of noise-like signal variations and streaks [
15]. Undersampled radial readouts are particularly applicable for cerebral angiography because the signal is sparse, so streak artifacts are minimal and often relegated to the image periphery [
16,
17]. Radial trajectories also result in more manageable flow or respiration artifacts than Cartesian acquisitions [
15], can give more accurate timing information [
18], and have previously been used with success for 3D non-vessel-selective ASL angiography [
18‐
23].
In this work, we compare undersampled radial and Cartesian acquisition schemes for the acquisition of accelerated dynamic vessel-selective ASL cerebral angiograms. In the context of rapid 2D dynamic protocols, we also compare bSSFP and SPGR, as bSSFP artifacts can be scan time and trajectory dependent [
24]. We then explore the potential for 3D whole brain acquisitions: over a large imaging region bSSFP artifacts are problematic and the required acceleration factors are so high that conventional Cartesian approaches are not feasible, so we instead demonstrate the potential of a SPGR radial approach. This work follows on from a previously presented conference abstract [
25].
Discussion
The ability to accelerate image acquisition is crucial for vessel-selective dynamic angiography with ASL where additional measurements are required beyond non-selective approaches. In this study, we have shown that radial trajectories have considerable benefits over conventional Cartesian techniques in this context. These include (1) the lack of need to acquire calibration data for parallel imaging, allowing all the available scan time to be used for imaging; (2) higher apparent SNR and distal vessel visibility; and (3) no noise amplification in the middle of the brain or residual aliasing arising from the parallel imaging reconstructions. The sparse nature of the angiograms meant that radial undersampling artifacts were relatively benign and most prominent at the edges of the field of view, away from the vessels of interest. The ability of radial trajectories to accelerate the acquisition was demonstrated by the acquisition of 2D vessel-selective angiograms in less than one minute. In 3D, even greater acceleration factors are possible, allowing the first demonstration of whole-brain vessel-selective dynamic angiography with a 3D radial trajectory within a reasonable scan time (16 min).
For vessel-selective dynamic 2D angiography, it was found that bSSFP resulted in images with higher SNR than SPGR images, as expected from previous work [
11]. No increase in flow-related artifacts were observed using radial trajectories compared to a conventional Cartesian approach. The Cartesian readout with high GRAPPA factor suffered from reduced SNR towards the center of the brain, around vessels of interest, which is to be expected due to the higher g-factor in this region [
40]. Radial undersampling, however, led to increased noise or signal aliasing which was most apparent at the edges of the FOV, away from these vessels. As a result, the time-matched radial acquisition had a two-fold better SNR than the Cartesian acquisition within the major cerebral vessels. Increasing the undersampling factor for the radial acquisition to achieve scans times of less than one minute resulted in vessel-selective angiograms with increased streak artifacts, but the SNR was still higher and the distal arteries still better delineated than in accelerated Cartesian acquisitions (Fig.
4). Image quality could perhaps be further improved through the use of parallel imaging in the radial image reconstruction [
41]. As well as increased noise amplification in the centre of the brain, accelerated Cartesian acquisitions also require additional calibration data, the measurement of which can take up a significant proportion of the scan time in this kind of rapid acquisition, further reducing the SNR efficiency. This has less impact in cases where a single pre-scan is used to measure coil sensitivity profiles and applied to all subsequent acquisitions. However, such an approach may be more sensitive to subject motion during the examination.
For a 3D acquisition, it was not possible to use bSSFP due to the greater B
0 inhomogeneity present across the whole brain, which has previously been shown to be problematic for ASL angiography [
11]. However, the increased SNR of a 3D acquisition still allowed good image quality to be obtained using an SPGR readout, despite the high undersampling factor required (×13). This degree of undersampling was not feasible using a conventional Cartesian parallel imaging reconstruction, although in future work the application of more advanced approaches such as blipped-CAIPI sampling [
42] may allow higher acceleration factors. The use of a true 3D radial approach in this study allowed the benefits of radial trajectories to apply in all three dimensions and gave whole-brain coverage within a reasonable scan time. However, particularly when only a reduced field-of-view is of interest, comparison of this approach with a stack-of-stars trajectory [
21,
23] would be useful in future work, as well as validation in a greater number of subjects.
The results of this study are consistent with previous work that has used various radial approaches for non-selective ASL angiography. For example, Koktzoglou et al. [
19,
22] and Wu et al. [
20] have demonstrated the excellent image quality achievable with radial trajectories, although no direct comparison with a matched Cartesian approach was performed. Song et al. [
21] and Cong et al. [
23] showed comparable image quality with an accelerated stack-of-stars trajectory combined with a
k-space filtering approach compared to a fully sampled Cartesian acquisition. Wu et al. [
18] also showed the improved timing information obtained from a radial approach compared to a Cartesian method in a digital phantom, although no experimental comparison was performed. In this study, we built upon this prior work by performing a direct experimental comparison of time-matched accelerated Cartesian and radial trajectories in the context of vessel-selective ASL angiography, where acceleration is particularly important, and utilize the benefits of a true 3D radial trajectory to perform whole-brain vessel-selective dynamic ASL angiography.
One downside of using bSSFP for the 2D acquisitions in this study was the appearance of artifacts in the proximal arteries in some subjects (Fig.
4), which were not observed in the SPGR data (Figs.
2,
5,
6). These may be a result of pulsatile blood and/or cerebrospinal fluid flow, or the presence of B
0 inhomogeneity. Similar artifacts were not observed in previous work on VEPCASL angiography with bSSFP [
11]. One possible explanation is the difference in the VEPCASL pulse train duration: here, the shorter tag duration was approximately the same as the cardiac cycle, perhaps giving poorer image quality in the initial frames due to data acquisition during systole. In addition, the TR used here was about 20% longer than the previous study and data were acquired on a different scanner, so the increased sensitivity to off-resonance effects combined with a different shimming setup could have led to artifacts in this region. Further investigation of this phenomenon is necessary before such a protocol could be deployed clinically. However, these artifacts were not specific to the trajectory, so we anticipate that the benefits of radial imaging over conventional Cartesian methods found in this study should generalize to other protocols where such artifacts have been minimized.
The measure of SNR used in this study is somewhat qualitative. It allowed aliased signal from undersampling to be included, which affects the estimates of both the vascular signal and the noise measure. The effect on the mean signal was likely to be small, as the aliasing within the brain was not severe. However, the aliased signal will contribute to the apparent noise in ROIs close to the arteries in such a way as to include information on signal fluctuations not due to local vascular signal, but which will still hinder image interpretation and are, therefore, important to consider. In addition, the noise ROIs were manually selected to avoid vessels, but it is possible that they still included some very small arteries that were not clearly visualized in the data, increasing the measured noise standard deviation. However, such an effect would have impacted all the protocols in the same way. Therefore, it is possible to use such an SNR metric to compare the relative performance of different protocols, but determination of the absolute image SNR would require a more sophisticated approach.
Whilst the dynamic 2D protocols described here can take as little as one minute to acquire, in this proof-of-concept study the setup time, including TOF and field map data of the labeling plane, and use of an iterative shimming procedure all contributed to the total scan time. In future work, scan time could be minimized by (1) fixing the location of the labeling plane relative to an anatomical landmark [
43] or using a planning-free approach [
44], removing the need for a TOF acquisition; (2) removing the field map scan used for off-resonance correction, since magnetic field inhomogeneity is typically small at this level in the neck, even when shimming is only performed over the imaging region [
32]; and (3) reducing the sensitivity of bSSFP to field inhomogeneity using a shorter TR and reduced flip angle, or by moving to a lower field strength (e.g., 1.5 T), or switching to an SPGR approach, thereby removing the need for an iterative shim. However, since these setup procedures were matched between the acquisitions, these factors should not have affected the observed improvements in image quality obtained with a radial trajectory compared to a conventional Cartesian approach.
The next stage of this work would be to increase the robustness of the 2D bSSFP approach in proximal vessels and explore the potential for further scan time reductions in the 3D SPGR acquisition. Comparisons with a gold standard, such as X-ray-based methods, and assessment through clinician scoring would help to establish the usefulness of these approaches in a clinical setting. The inclusion of the protocol on a patient population would better evaluate its ability to provide dynamic vessel-selective information in diseased cerebral vasculature where vessel caliber, hemodynamics and motion characteristics will likely differ from those in healthy volunteers. Additionally, the focus of this work has been time-resolved angiograms with vessel-specific information for four arteries, but it could easily be extended to more than four vessels, albeit with the requirement for a longer scan time. The optimized encoding scheme technique with off-resonance correction [
33] is able to generate a minimal number of optimized encodings for any number of vessels. Consequently, the number of encoding cycles can be tailored to the number of vessels of interest and kept to a minimum. The result of this would be relatively short total scan times for many-vessel dynamic angiographic information. In addition, further acceleration could be achieved through the use of advanced image reconstruction techniques such as compressed sensing [
45], which are well suited to sparse angiographic data and radial trajectories [
46]. Additional flexibility in the reconstruction could also be provided through the use of golden ratio spoke ordering [
47,
48], allowing the temporal resolution and undersampling factors to be chosen retrospectively.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.