Recent Technical Developments in ASL: A Review of the State of the Art

2.1 Improvements in 3D segmented readouts

2.2 Cartesian FSE

Segmented FSE acquisitions with Cartesian encoding, where 1 line of k-space is acquired after each refocusing pulse, are workhorses of volumetric imaging, having excellent off-resonance robustness and anatomical fidelity. Whereas this makes FSE particularly attractive for body ASL and high-resolution ASL, long acquisition times are a major limiting factor if large volume coverage is required.

Using a reduced Field of view (FOV) with selective excitation allowed the benefits of volumetric Cartesian encoding for renal imaging to be demonstrated.²⁰ More time-efficient acquisitions with spiral re-ordering on a Cartesian grid,²¹ variable-density sampling combined with compressed sensing reconstruction for body,²² and brain imaging²³ have also been demonstrated.

2.3 Radial trajectories

Whereas conventional trajectories (e.g., 2D/3D EPI or spirals) are very efficient at covering large amounts of k-space quickly, they generally have a fixed spatial/temporal resolution and suffer from artifacts due to off-resonance effects and motion between shots. Radial k-space trajectories, which acquire a single line of k-space at a time through its center with different orientation, sample fewer k-space points but allow the retrospective choice of spatial and temporal resolution for reconstruction when using golden ratio sampling^{24, 25}; are intrinsically robust to motion; do not suffer from significant distortion, blurring, or signal dropout artifacts; and tolerate relatively high levels of undersampling, particularly when combined with advanced reconstruction techniques (see below).

Radial trajectories have been used fairly extensively for ASL angiography (see below), and more recently for assessing the labeling efficiency of velocity-selective ASL preparations²⁶; however, a few new methods have also explored their use for ASL perfusion imaging. In the Combined Angiography and Perfusion using Radial Imaging and ASL (CAPRIA) approach,²⁷ a PCASL preparation is followed by a continuous golden ratio readout. Dynamic angiographic images are reconstructed using a small number of radial spokes from early timepoints, while the labeled blood still resides within the arteries. This results in a high temporal resolution and a high undersampling factor, but the sparse nature and high SNR of the angiographic signal means good quality images can still be reconstructed. Using the same raw data, perfusion images can be reconstructed from later time points once the labeled blood arrives at the tissue.

A golden ratio readout can also be combined with a time-encoded ASL preparation: this means fewer excitation pulses are needed to span a range of effective postlabeling delays,²⁸ allowing higher flip angles to be used without causing excessive signal attenuation. This boosts the SNR, in addition to the noise-averaging benefit of time-encoding. Although potentially more time-efficient than separately acquired angiography and perfusion imaging, further studies are required to refine these techniques and test them in clinical cohorts.

2.4 Cardiac triggering

The variability of blood flow velocity in the brain-feeding arteries affects the ASL labeling efficiency (in CASL and PCASL) and arterial transit time. These effects have been tested with cardiac gating in pulsed,^{29, 30} pseudo-continuous,³¹ and velocity/acceleration-selective ASL.³² For example, shorter bolus arrival time and a 16% higher perfusion signal in gray matter (GM) were found when triggering a pulsed ASL (PASL) labeling module at systole compared to diastole, although the signal was similar at long TIs.²⁹ Larger signal variations across the cardiac cycle have been demonstrated for velocity selective ASL (36%) and acceleration selective ASL (64%) compared to PCASL (25%).³² Similarly, stability gains were found in vessel-selective ASL by triggering.³³

A PCASL study³¹ triggered the end of the labeling period to a specific cardiac phase with a long labeling duration (> 7 s) and found no significant differences in vivo in the mean ASL signal and its stability. However, a second study³⁴ tested a nontriggered- versus a cardiac-triggered standard PCASL sequence with the parameters suggested in the consensus paper.³ The nontriggered PCASL sequence showed signal fluctuation near large vessels in single-shot acquisitions and also more artifacts in segmented acquisitions, whereas the cardiac triggered sequence demonstrated higher temporal SNR.

Cardiac triggering improves stability at a cost of increased dead-time in the sequence while waiting for the next cardiac trigger. Triggers should be applied to the start of labeling because triggering of the readout would lead to differences in postlabeling delay (PLD) between acquisitions and thus imperfect subtraction of static signal between label and control condition.

2.5 Suggestions

The use of moderately segmented 3D readout schemes continues to be recommended for ASL due to their high efficiency and SNR, as well as their ability to achieve spatially uniform background suppression. The use of parallel imaging with relatively low acceleration factors (e.g., 2 or 3) is also recommended when available (e.g., for Cartesian trajectories), particularly when combined with low g-factor methods such as CAIPIRINHIA. We encourage the further development and validation of newer techniques before they are used for clinical research applications. At this time, there is not a sufficient amount of evidence to recommend the general use of cardiac triggering with ASL.

3.1 Advanced reconstruction techniques

ASL-perfusion imaging has some inherent properties that make it well suited for acceleration and reconstruction using compressed sensing methods. Particularly, compressed sensing has been shown to perform well when applied to ASL difference images by leveraging sparsity across the averages²² or using a total generalized variation constraint in combination with a time-dependent CAIPIRINHA sampling pattern.³⁵ Multi-delay ASL images can be further improved by additionally exploiting the redundancy among images (temporal sparsity) with different labeling duration and postlabel delays. For example, an over-complete dictionary was built from the perfusion model and was used to sparsify the acquired ASL signal.³⁶ This helped reject noise and motion artifacts that could not be described by the perfusion signal model.

3.2 Noise reduction

Many strategies have been developed to improve ASL SNR using image-processing techniques. Spatial smoothing is a routine procedure for suppressing random noise in MRI and has been used frequently for ASL¹⁴; however, this further reduces the already low spatial resolution and blurs perfusion differences between tissue types. This can be partly addressed by using Wavelet denoising³⁷ or by a spatial kernel as part of partial volume correction approaches.^{38, 39} High-pass filtering can remove temporal noise³⁶ as the perfusion signal encoded in the label-control acquisition paradigm is located in the high-frequency band.^{40, 41}

Outliers, caused by physiological fluctuations or subject motion, are a major challenge for ASL MRI, especially due to the limited number of samples.³⁹ Robust fitting⁴² can address outliers at the voxel-level, although it does not take spatial information into account. Several empirical algorithms were introduced to remove outlier volumes or slices before calculating the final CBF map, which can be identified based on motion parameters and variation in the CBF time series³⁹ or using a M-estimator.⁴³ An adaptive outlier cleaning algorithm (see Figure 2) can iteratively identify outlier volumes based on the correlation of each remaining volume to the current mean image.⁴⁴ This approach can be improved using structural information regularization,⁴⁵ using a prior-guided slice-wise adaptive outlier cleaning method,⁴⁶ or by accounting for relative motion.⁴⁷

Alternatively, spatial priors can be used on the resulting CBF and/or arterial transit time (ATT) maps^{48, 49} to reduce the effect of outliers, or a total generalized variation regularized spatial–temporal filtering algorithm can be used for directly denoising the raw ASL images.¹⁹

Another strategy to denoise ASL data is to decompose the signal into components and then regress out the “noise” components. One approach is to use independent component analysis with manual or automatic classification of components (e.g., by assessing if the spatial/temporal variations match the expected perfusion signal), which results in improved SNR and repeatability^50-52 (Figure 3). Similarly, the component-based noise correction method extracts principal components from noise regions of no interest, which can be used as covariates in a general linear model and improve the stability of the perfusion signal.^{53, 54} Alternatively, a low-rank and sparse decomposition can separate the ASL image series into slowly changing perfusion and spatially sparse noise component.⁵⁵

Recently, DL has been utilized for simultaneous denoising and resolution improvement in ASL,^56-58 and various approaches have allowed a significant acquisition time reduction without sacrificing CBF quantification quality.^{59, 60} Unsupervised DL ASL denoising algorithms using autoencoder networks have also been proposed,⁶¹ reducing the burden of generating large amounts of training data. Deep convolutional neural networks have been used to enhance image quality of multi-timepoint ASL data acquired with a low number of averages,⁵⁹ showing a 40% higher accuracy than the conventional averaging method when tested on ASL data of stroke patients.

Different ASL acquisition strategies introduce different noise patterns, making it necessary to fully evaluate the capability of a model to transfer from one type of ASL data or population to another. However, it is important to be careful not to “over-denoise” functional ASL images, as sometimes the activation itself is correlated with components identified as noise, and suppressing too many temporal components may artificially increase functional connectivity.

3.3 Partial volume correction

ASL spatial resolution is typically much lower than the cortical thickness (average value ∼2.5 mm vs. typical ASL resolutions of 4 × 4 × 4 mm³). In superficial brain regions, individual voxels are therefore highly likely to contain a mixture of GM, white matter (WM), and cerebrospinal fluid (CSF), which is known as the partial volume (PV) effect. Given that GM perfusion is approximately 2–5 times greater than WM perfusion^{62, 63} PV will have a large effect on CBF quantification. In ASL, the primary focus is often on GM-CBF. PV effects bring 2 issues here: 1) actual GM content is still variable in nominally “GM voxels” causing potential GM-CBF underestimation; and 2) the spatial distribution of predominantly GM voxels varies between subjects, causing a potential evaluation bias (Figure 4). The importance of PV effects grows in longitudinal and cross-section studies where cortical thickness varies in time and across groups.^{64, 65} Several algorithms have been proposed to correct for PV-effects at the voxel level using fractional GM and WM maps obtained from segmenting structural images. These algorithms either assume a locally homogeneous GM and WM CBF,⁶² leveraging the different kinetics in GM and WM (along with spatial regularization),³⁸ or use GM volume as a covariate in the statistical analysis.⁶⁶

The quality of the fractional GM and WM maps, along with coregistration, distortion correction, and resolution errors⁶⁷ also propagate into the PV correction. However, these errors would have a similar influence on non-PV corrected GM-CBF evaluation using a GM mask,⁶⁸ or alternative approaches using tissue classification from inversion recovery or a similar readout sequence.⁶⁹ It is important to note that the partial volume effect is a methodological artifact. Correcting for it allows the investigator to examine changes in perfusion and GM volume as separate effects, even in patients where both are changing concurrently. For the latter, GM volume could be a covariate in statistical analyses; for the former, PV correction is more appropriate. These 2 issues are currently usually not separately addressed.

3.4 Suggestions

When processing ASL data, we recommend the use of motion correction (unless very strong background suppression is performed) and consideration of at least 1 denoising technique (such as adaptive outlier cleaning or component-based methods) if there are sufficient measurements to support them. Partial volume correction as an additional analysis is strongly recommended for studies focusing on specific tissue types, such as GM, especially if a difference in tissue volumes, for example, due to atrophy, is expected between participants or cohorts.

4.1 Blood oxygenation and oxygen consumption

Spin labeling methods can be creatively applied to measure venous oxygen saturation (S_vO₂, or Y_v are commonly used in the literature), from which Oxygen Extraction Fraction (OEF) and cerebral metabolic rate of oxygen (CMRO2) can be subsequently estimated. All 3 parameters are important indicators of brain health and function and are often perturbed in states of disease.

One class of methods to estimate Y_v first measures the T₂ of venous blood, which is then calibrated to Y_v using empirical or theoretical relationships because blood T₂ is directly related to the blood oxygenation fraction.⁷⁰ OEF can then be estimated using the derived venous oxygenation (Y_v) along with a measured or assumed value of arterial oxygenation (OEF is defined as the ratio of the extracted oxygenation to arterial oxygenation). Rate of metabolism is calculated as the product of the assumed arterial oxygenation, OEF, and CBF.

An effective way to measure T₂ values of blood in vivo is to apply T₂ weighting “preparation modules,” which consist of ±90° hard pulses enclosing a train of refocusing pulses with different TEs, immediately before image acquisition.⁷¹ This approach has been applied to determine blood T₂ of coronary veins,⁷² brain sagittal sinus,^{73, 74} and internal jugular veins.^75-77 For abnormal blood composition such as sickle cell anemia, T₂-based oximetry may require disease-specific calibrations.^{77, 78}

The main challenge of this approach, however, is isolating signal solely from venous blood without contamination from tissue, CSF, or blood from other vascular compartments. Spin labeling methods provide a natural option to isolate vascular signal because the intrinsic subtraction can eliminate signal from unwanted voxel constituents.

T₂-Relaxation Under Spin Tagging (TRUST) was the first spin labeling technique to target venous blood signal.^{73, 79} TRUST modifies the pulsed ASL experiment by placing the inversion band above the imaging slab (instead of below) to invert venous spins flowing inferiorly. Control-label subtraction yields high signal exclusive to medium-to-large size veins within the imaging slab. A T₂ preparation module or FSE readout generates multiple echoes to fit for venous blood T₂, ultimately yielding high-SNR global oxygenation measurements in short scan times.

The Quantitative Imaging Of Extraction Of Oxygen And Tissue Consumption (QUIXOTIC) method expands on TRUST by employing velocity-selective pulse trains to label blood accelerating from capillaries into the venous system. This allows T₂ measurement of venous blood on a voxel-by-voxel basis, and generation of Y_v, OEF, and CMRO2 maps. QUIXOTIC, however, is limited by low SNR and error introduced by CSF contamination.⁸⁰ The Velocity Selective Excitation and Arterial Nulling (VSEAN) technique mitigates these limitations by applying a unique velocity-selective excitation to acquire signal directly from slow-moving venous spins, thereby improving SNR and reducing CSF contamination.⁸¹

4.2 MR fingerprinting ASL

A dynamic time series of images, in which the acquisition settings are varied in a pseudo-random (but known) pattern, can be used to identify the underlying MR parameters of the tissue (e.g., its relaxation times).^87-90 The specific combination of tissue MR parameters at each voxel produces a unique dynamic MR signal for that specific acquisition, and this signal can be predicted in simulation. In MR fingerprinting, the parameter fits are carried out by identifying the signal from a precomputed database, or “dictionary,” of signals that matches the observed signal most closely. The entry that is most correlated with the observation corresponds to the appropriate combination of MR parameters.

The key features and advantages of the fingerprinting approach are that it produces joint parameter estimates from a given signal and is robust to spurious signals as long as their effect is not correlated with the parameter of interest. Joint parameter estimation of variables, like T₁ and T₂ relaxation, eliminates coregistration and other biases from separate measurements or assumptions. The dictionary matching process is generally very fast, but generating the dictionary is a computationally expensive process and can result in coarse granularity of the parameter estimates.

Fingerprinting is an appealing strategy in the context of quantitative ASL for several reasons. Primarily, ASL is intrinsically low SNR,⁹¹ and the robustness of fingerprinting to noise offers a major benefit. Second, quantification of ASL requires multiple parameters to be measured or assumed a priori. This can introduce biases into the measurement if assumed, or coregistration errors and additional scanning time if those additional parameters are measured separately. In contrast, ASL fingerprinting has been successfully implemented by collecting a single time series of PCASL prepared images in which the labeling duration varies according to a pseudo-random, predetermined schedule, and the control/label condition of the PCASL preparation train is also randomized. A postlabeling delay is not necessary because the control PCASL periods serve as variable postlabeling delays for modeling as reduced flip angles are used to preserve some ASL signal from previous TRs. From this time series, multiple parameters can be estimated by matching the signal to a precomputed dictionary, usually T₁ relaxation, perfusion, arterial blood volume, and bolus arrival time.

In several studies, ASL fingerprinting with dictionary matching was able to estimate the hemodynamic parameters of interest, showing good agreement with more established ASL techniques.^92-96 Recently, however, DL methods have been shown to be a powerful alternative to dictionary matching.^{94, 96} Whereas the data acquisition portion of the method remains the same, the parameter estimation portion can be accomplished more efficiently using neural network regression (Figure 5).

4.3 ASL angiography (ASL-MRA)

ASL angiography (ASL-MRA) has many advantages over conventional contrast-enhanced MR/CT methods⁹⁷: it allows vessel-selective labeling (especially useful for assessing arterial supply to, for example, arteriovenous malformations/fistulas^98-104) and has excellent flexibility in temporal and spatial resolution because the labeling and associated imaging readout can be repeated until the desired resolution is reached, unconstrained by the necessity to image the first passage of a contrast bolus.

To achieve high spatial resolution, however, the entire scan time is often used to acquire a large k-space matrix without signal averaging. When vessel-selective labeling is employed targeting multiple arteries, the total scan time can become very long. Therefore, the use of acceleration techniques should be considered: for example, undersampled golden-angle stack-of-stars¹⁰⁵ and 3D radial “koosh-ball” acquisitions,^{106, 107} in conjunction with advanced image reconstruction techniques such as CS and k-space weighted image contrast.¹⁰⁸ Fortunately, ASL-MRA is well-suited for undersampled reconstruction because of its high sparsity in the image domain after subtraction, particularly when it is vessel-selective.¹⁰⁹

Both PASL and PCASL can be used for ASL-MRA. PASL with a Look-Locker readout has already proved its clinical usefulness in several studies^{101, 104, 110, 111} and is particularly good at visualizing the early inflow phase of the proximal arteries. However, vessel-selective PASL has some difficulties (see below), which makes PCASL a preferred option for vessel-selective MRA. PCASL can also be combined with subtraction techniques to visualize blood inflow.^112-114

For static 3D-MRA, in contrast, PCASL's long labeling duration is more advantageous for visualizing the whole arterial tree, and a hybrid of PCASL and PASL helps to minimize the signal loss in proximal vessels caused by fresh unlabeled blood flowing into the imaging volume.^{107, 115}

Recently, velocity selective static 3D-MRA^116-119 has also been demonstrated by utilizing Fourier transform-based velocity selective saturation pulse trains, which set the flowing spins in the pass-band and static spins in the saturation-band before acquisition as a nonsubtractive method.

The typical readout for ASL-MRA is based on 3D gradient-echo sequences. However, with a Look-Locker readout, the repetitive excitation pulses can strongly attenuate the ASL signal when the flip angle is high. This can be mitigated through the use of a balanced steady-state free precession readout (Figure 6) that recycles the transverse magnetization for the next excitation,^{113, 120} or the use of a segmented EPI readout to reduce the number of excitation pulses while making the interval between RF-pulses longer.^{97, 114} However, off-resonance effects can cause loss of vessel depiction with balanced steady-state free precession,¹¹³ so high B₀ homogeneity is required (e.g., using a small FOV or lower B₀ field strength) and segmented EPI can suffer from ghosting due to strong pulsatile flow,¹²¹ typically at the M1 section of the middle cerebral artery, although this is reduced when using right–left phase-encoding.

ASL-MRA can be combined with perfusion imaging in a single sequence by sharing the labeling module, providing both macrovascular and microvascular information: besides CAPRIA²⁷ (described earlier), time-encoded PCASL can be combined with a segmented EPI 4D-MRA readout, minimizing the number of excitation pulses required and preserving magnetization for a separate perfusion-weighted readout.¹²²

4.4 ASL functional MRI (fMRI)

Although hampered by its low SNR and acquisition speed, early work demonstrated that ASL offered several important advantages over blood oxygen level-dependent (BOLD) fMRI. These include its quantitative nature and the temporal stability of the measurement—that is, it is not subject to 1/f noise that plagues BOLD fMRI.^{40, 41, 123-126} These features make it more suitable for fMRI experimental paradigms that span longer periods of time (e.g., blocked designs with durations greater than a minute), such as applied in pharmacological fMRI or when studying conditions like sleep deprivation. For example, in an extreme case, images of the control and active conditions were taken 30 days apart and reliable activation maps of the motor cortex could still be obtained.¹²⁴

Another advantageous feature of perfusion-based (and blood volume-based) fMRI is that CBF and cerebral blood volume changes are more specific to the parenchyma where the neural activity takes place, rather than the draining veins. This feature makes it particularly appealing for layer-specific fMRI, where BOLD imaging is unable to differentiate activity between cortical layers.^127-130

ASL is also advantageous for fMRI in regions of high susceptibility-induced static field inhomogeneities, such as the orbito-frontal cortex, the amygdala, or the medial temporal lobe, where BOLD techniques are prone to signal loss because ASL does not depend on susceptibility contrast, and thus ASL images can be acquired using sequences with low T₂* sensitivity.¹³¹ This feature of the technique makes it attractive for fMRI studies of spoken language because it is less sensitive than BOLD to speech-related motion and susceptibility confounds.^132-134

ASL-based fMRI sequences typically avoid acquiring segmented readouts to ensure a sufficient temporal resolution. Besides the traditional 2D multislice EPI readout, 3D stack of spirals¹³⁵ and 3D-GRASE readouts¹³⁶ are efficient approaches to collect all of k-space after a single labeling/control period. An attractive acquisition strategy, using a pseudo golden-angle stack-of-spirals 3D Rapid Acquisition with Relaxation Enhancement (RARE) readout and CS reconstruction, has been recently proposed that yields high spatial resolution time-averaged CBF maps and low spatial resolution measurements of CBF fluctuations.¹³⁷ More recently, velocity selective labeling pulses have been shown to allow faster sampling and improved sensitivity¹³⁸ and could become more widely adopted for perfusion-based fMRI.

ASL has also found some use for assessing resting-state functional connectivity. Early on, it was shown that connectivity of the sensorimotor network could be detected with ASL by evaluating fluctuations in the CBF signal.¹³⁹ Since then, several studies performed to identify resting state networks, applying different analysis methods, such as seed-based connectivity,^140-142 independent component analysis^143-147 and whole-brain voxel level connectivity,^{143, 148} have found similar brain networks as those observed in resting state BOLD studies. As in the case of task activation studies, resting-state functional connectivity measured with ASL can potentially provide better localization of resting state networks than BOLD, despite the lower spatial resolution of the ASL images. The lower temporal resolution of ASL is not so much of a disadvantage because resting state connectivity is based on the correlation of low frequency signal fluctuations.

5.1 Slab-selective single artery labeling

Some of the original techniques for vessel-selectivity restricted the spatial region over which an ASL inversion pulse acted, thereby only labeling a single vessel at a time. The most common approach is to use a conventional slab-selective inversion pulse but to angle it in such a way as to only cover the artery of interest.^151-153 Efficient postlabeling saturation must then be used to remove any effect of the angled labeling pulses on tissue magnetisation within the imaging region. However, orientating the slab to cover only the artery of interest, which is often tortuous, is challenging. In addition, if only a limited vessel segment can be covered, then the bolus of labeled blood created is relatively small, the SNR of the resulting images is impaired, and perfusion quantification is challenging.

5.2 Superselective methods

Vessel-selective labeling based on (P)CASL avoids some of the drawbacks of the slab-selective PASL-based methods by using a secondary gradient perpendicular to the main labeling gradient axis. If the gradient is rotated dynamically during the labeling period instead of applying this gradient in a continuous fashion, one can achieve a small labeling region. Early vessel selective work using CASL essentially created a labeling plane that was not perpendicular to the flow direction and rotated about a target artery such that only the spins flowing through that artery would experience the adiabatic inversion that underlies CASL.^{154, 155}

A similar idea can be applied to PCASL methods by inserting in-plane gradient pulses between the individual RF labeling subpulses that make up a balanced PCASL labeling train (see Figure 7). The effect is a phase distribution of the spins determined by their location along the in-plane gradient direction. Matching the phase of the individual pulses in the labeling train to the phase of the spins at a specific vessel location allows the creation of a “labeling stripe” that tags spins flowing through that location by adiabatic inversion similar to nonselective PCASL. The periodic nature of phase accrual means that if the in-plane gradient pulses were the same each time, these conditions would be met at a number of stripes within the labeling plane. In superselective PCASL, the in-plane gradient is rotated at varying increments between RF pulses in the PCASL train (in a continuous or pseudo-random fashion), and the RF phase adjusted such that only the spins flowing through 1 location in the plane will experience the adiabatic inversion process.^{156, 157}

The amplitude of the in-plane gradient blips determines the effective “labeling spot” size and must be chosen as a compromise between labeling efficiency/insensitivity to motion (larger spot size) and the potential for labeling other nearby arteries (smaller spot size). Moreover, the labeling plane needs to be oriented approximately perpendicular to the artery, intersecting at a straight part of the artery and without intersecting the tissue in which the relevant imaging is performed.

Superselective PCASL has already shown some promising results in patients with a range of cerebrovascular diseases, including steno-occlusive disease and arteriovenous malformation.^{150, 158} Recent work on correcting for off-resonance effects and pulsatility is likely to further improve robustness.³³

5.3 Vessel-encoding

Given a limited scan time, labeling methods with higher time efficiency are preferred, that is, methods that can label several feeding arteries simultaneously either by pulsed¹⁵⁹ or (pseudo-) continuous labeling methods.^{160, 161} In this type of approach, perfusion images are acquired in a few “encoding” steps. As described earlier (see Figure 7), including an additional gradient blip within the PCASL labeling plane in a consistent direction, along with associated RF phase modulations, creates spatial labeling bands within the plane without labeling the blood in other regions. The encoding of arteries is achieved by labeling different subregions of the labeling plane over a series of readouts. For PASL-based approaches, this involves positioning the labeling slab to cover more than 1 artery at a time, although the difficulties in positioning this slab to cover tortuous arteries still remain, so PCASL-based approaches are generally preferred.

Using columns from a Hadamard encoding matrix (with elements of 1 or − 1) to construct the encoding matrix,¹⁶⁰ such as the one shown above, maximizes encoding and SNR efficiency.¹⁵⁹ This leads to vessel-encoded ASL, sometimes being referred to as Hadamard-encoded ASL, although this should not be confused with time-encoded methods, which also use Hadamard encoding.¹⁶² To distinguish $$$ N $$$ vascular territory regions, the SNR for each feeding artery using Hadamard encoding is improved by a factor of $$$ \sqrt{N} $$$ compared to labeling each feeding vessel individually,¹⁵⁹ given the same total acquisition time.

Due to variation in the geometry of the feeding arteries and scanner hardware limitations, Hadamard encoding schemes may not always be feasible, or the planning/calculation process could be slow, although some automated methods to optimize the encodings have been proposed.^163-165 Optimization of the labeling parameters can also improve the separation of arteries selected to be in label or control conditions.^{161, 166}

Due to field inhomogeneities such as B₁ variation or off-resonance at the labeling sites, the actual labeling status of the feeding arteries may deviate from the designed values (e.g., encoded as 0 if the signal is saturated) and should be estimated from the data to accurately decode the vascular territory information.¹⁶⁰ This can be done by estimating the encoded labeling efficiency of the ASL signal in each perfusion territory by k-means clustering and linear analysis,¹⁶⁰ or by using Bayesian inference framework with improved accuracy.¹⁶⁷

Some applications of vessel-encoded ASL include detecting/assessing collaterals,¹⁶⁸ or producing vessel-encoded angiograms¹⁶⁹ that can be used to assess the blood supply to arteriovenous malformations.¹⁷⁰

5.4 Suggestions

PCASL is the recommended method for vessel-selective ASL. When choosing between the vessel-encoded and superselective labeling schemes, the purpose of the scan should guide the decision: when there is need to have insight into all (or the main) flow territories, vessel-encoded labeling using a Hadamard scheme is the most efficient method and will yield the highest SNR. However, when there is specific interest in the flow territory of a single or a few arteries, especially in cases where these arteries are located intracranially or are part of an unusual vascular anatomy, superselective labeling is the method of choice: it allows the labeling plane to be optimally positioned for each artery and is perhaps the simplest to implement. However, in both methods, imperfect labeling efficiency must be accounted for when trying to quantify CBF or mixed perfusion fractions.

6.1 ASL quantification

DL provides a powerful way for solving complex nonlinear inverse problems, such as the one posed by ASL, particularly in the fingerprinting application (described above). In the case of ASL fingerprinting, neural network regression can be used to estimate multiple parameters independently, 1 at a time, without assuming the value of the other parameters.

The general strategy is to generate a database of synthetic signals based on a physics-based model, the pulse sequence parameters (e.g., labeling duration schedule, PLD, TR), and many parameter combinations. This database of signals is then used to train a set of neural networks to output the desired parameters. Once trained, each of the networks will take the observed signal as input and yield a parameter estimate as its output. Alternatively, experimental ASL data from a high-quality data set in which the underlying parameters were known a priori can be used to train the neural networks instead of using purely synthetic data from Bloch simulations.⁹⁶

Training the neural networks requires a large database of signals, which is computationally expensive to synthesize and store. However, the network needs only to be trained once. After training, computation of the output (i.e., the parameter estimates) is extremely fast. This approach offers an important advantage over dictionary learning: it allows for much finer granularity of the parameter estimates, whereas the dictionary entries are computed on a coarser grid of parameter values because the size of the dictionary grows exponentially with the grid size and the number of parameters (dimensions) that one wishes to estimate.

In terms of ASL, this strategy has been demonstrated to estimate hemodynamic parameters from ASL fingerprints quite effectively,^{94, 96} although a fingerprint's sensitivity to perfusion and other hemodynamic parameters can be limited in some cases.⁹⁴ Optimizing the fingerprint readout schedule to maximize the sensitivity to perfusion (using an objective metric of sensitivity, such as the Cramer-Rao bound) is crucial to obtaining reliable estimates. As a result, perfusion, arterial transit time, and arterial blood volume can be estimated reliably in addition to T₁ relaxation time and the effective flip angle, giving good agreement with standard measurements.^{94, 96}

6.2 Machine learning and ASL for diagnosis

ML and DL give us the means to study regional and voxel-wise patterns of pathological perfusion changes in more detail than a simple evaluation at specific pathology-related regions. Two distinct approaches are generally used for ASL: (i) evaluation of regional mean CBF in anatomical regions based on atlases and then working in the vector space defined by these regions to, for example, separate healthy controls from patients with a major depressive disorder¹⁷³; and (ii) process the full voxel-wise CBF maps either using DL based on neural networks, or using a feature space reductions methods (such as PCA) and traditional ML algorithms (such as a support vector machine). Although DL-based methods can achieve higher performance and are not bound to predefined anatomical regions, such methods have numerous shortcomings: Much larger datasets are needed for training, and they suffer from interpretability issues, can cue on nonperfusion-based artifacts such as motion, and are computationally more demanding. The major hurdle is, however, the sensitivity of the ASL protocol: variations in acquisition parameters (commonly present in ASL) can render a well-performing machine learning method useless on another acquisition protocol.

Despite the first examples of ML/DL applications emerging, they are still pilot studies conducted on a limited number of patients from a single cohort without an external validation and are thus far from wider adoption in clinical research. Standardizing image processing to decrease the between-center differences in data¹⁷⁴ is a way to gather larger datasets, necessary for both the ML and DL training.

6.3 Suggestions

Whereas ML offers great promise, this field is still evolving. We anticipate the continued development and validation of these techniques for ASL, particularly those that are robust to differences between sites, scanners, and acquisition protocols.

7.1 Suggestions

UHF PASL using appropriately optimized inversion pulses could be considered when very high spatial resolution is required, particularly for layer-specific functional imaging, although this becomes more challenging in inferior brain regions. Whereas UHF PCASL shows great promise, technical challenges such as B1 inhomogeneity and power deposition have thus far hindered its implementation, so further work in this area is encouraged to allow optimal labeling durations and background suppression to be achieved, perhaps utilizing full parallel transmission capabilities.