Arterial Spin Labeling Perfusion MRI Signal Processing Through Traditional Methods and Machine Learning

Abstract

Arterial spin labeling (ASL) perfusion MRI is a non-invasive technique for quantifying and mapping cerebral blood flow (CBF). Depending on the tissue signal change after magnetically labeled arterial blood enters the brain tissue, ASL MRI signal can be affected by several factors, including the volume of arrived arterial blood, signal decay of labeled blood, physiological fluctuations of the brain and CBF, and head motion, etc. Some of them can be controlled using sophisticated state-of-art ASL MRI sequences, but the others can only be resolved with post-processing strategies. Over the decades, various post-processing methods have been proposed in the literature, and many post processing software packages have been released. This self-contained review provides a brief introduction to ASL MRI, recommendations for typical ASL MRI data acquisition protocols, an overview of the ASL data processing pipeline, and an introduction to typical methods used at each step in the pipeline. Although the main focus is on traditional heuristic model-based methods, a brief introduction to recent machine learning-based approaches is provided too.

Introduction

Cerebral blood flow (CBF) is a fundamental physiological quantity that reflects regional brain function and neurovascular condition. Regional CBF changes have long been essential for neurological and neuropsychiatric disease assessments. CBF can be measured using different methods, but arterial spin labeling (ASL) perfusion MRI remains the only non-invasive technique for measuring regional cerebral blood flow (CBF) [1, 2]. As illustrated in Fig. 1, ASL perfusion MRI uses radiofrequency (RF) pulses to modulate arterial blood signals in the feeding arteries magnetically in a place proximal to the image site. After the labeled blood is transmitted to the place to be imaged, it will exchange with the tissue water and reduce the tissue’s signal. That signal change is proportional to the volume of perfusion. It can be converted to the quantitative CBF in units of ml/100g/min after normalization by the fully relaxed MR signal-M0 and taking into account the signal decay.

Fig 1.

Illustration of continuous arterial spins labeling (CASL). The radiofrequency pulses are spatially focused at the place as labeled by the green line. When fresh arterial blood (red color) reaches the labeling plane, its MR signal is immediately flipped from positive (red) to negative (blue color), transported to the imaging place, and exchanged with tissue water. PCASL differs from CASL by using a chain of short radiofrequency pulses rather than a long continuous radiofrequency pulse. The same scheme can be used to illustrate a type of PASL by enlarging the thin green line into a rectangle and changing the radiofrequency pulse train into a short inversion pulse.

To extract the perfusion weighted signal from the background, ASL MRI often uses a pair-acquisition scheme, with each acquisition consisting of one spin- labeled image and another spin-unlabeled image. These two images are often named by label (L) and control (C) images, respectively. The control image is acquired with the same ASL MRI sequence but after canceling the spin labeling function of the RF pulses using phase manipulations. Because labeled cerebral blood water maintains the T1 ratio of the decay, the adequate vortex labeling time is often limited to 1-2 secs, and the signal change associated with the arrival of labeled arterial blood water spins in only approximately 1-5% of mean MR signal intensity [3] in the absence of background suppression (BS) [4]. This relatively weak signal induces a low signal-to-noise (SNR) of ASL data, which is even worse when the average CBF value decreases in conditions such as Alzheimer’s disease (AD) [5, 6]. A typical strategy is to acquire multiple pairs of label/control images and average the perfusion weighted image obtained at different time points, but this basic strategy can only suppress random noise.

Mathematically, perfusion signal extraction in ASL MRI can be described by, y = xβ + ε, where y is the signal intensity of an arbitrary voxel from all L/C image pairs, x=[−1, 1, −1, 1, …] is the spin non-labeling and labeling paradigm, β is the fitting coefficient which is proportional to the mean perfusion signal at that voxel, and ε is the residual noise. Post-processing here involves various aspects. This can be the L/C image denoising, L/C pair difference denoising, or different approaches to estimate the fitting coefficient. The most widely used approach to obtain the perfusion signal is to take the mean of the L/C image difference, similar to multiplying the above equation precisely by the transpose of x. The pairwise subtraction is the same as the Haar transform[7]. The standard subtraction or fitting-based perfusion signal extraction approach assuming a constant zig-zag labeling paradigm, which however is often not true in real ASL data. An alternative approach is to identify the labeling and non-labeling paradigm induced signal oscillation from the first intrinsic mode function using an empirical mode decomposition method such as the Hilbert-Huang transform[8]. In the following, we will focus on the traditional subtraction based approach.

In the past decades, post-processing strategies have been proposed for ASL MRI [9-30]. Several software packages have also been published, including ASLtbx [13] and BASIL[31]. This review summarizes the current standard ASL MRI data processing and highlights advances in machine learning-based processing strategies. An Introduction to recent advances in ASL imaging technique development can be found in [32].

ASL MRI data acquisition protocols

The entire imaging process for obtaining a label or a control image is illustrated in Figure 2 and consists of three modules. Different ASL MRI sequences differ by differences in some or all of these modules. Once a specific ASL MRI sequence is selected, the three constituent modules will be repeated to acquire the label images and the control images . The only difference between the label image and control image acquisition procedures is that there is additional phase or gradient polarization modulation in the spin labeling module. This modulation is designed to keep the atrial blood signal unchanged at the end of the spin labeling module. The protocol for acquiring ASL MRI data can vary depending on the sequence used. The following paragraphs provide an empirical reference for setting up a protocol to suit the needs of routine ASL MRI data acquisition.

Fig. 2.

Illustration of the entire ASL MRI imaging procedure. ASL MRI consists of a spin labeling module, a waiting time that can use background suppression (BS) pulses to suppress the background tissue signal, and an image acquisition module. The total time for ASL MRI is the repetition time (TR). In order to acquire a control image without arterial spin tagged, the spin labeling pulses in the spin labeling module are modulated to a 0 net arterial blood MR signal change at the end of the module.

1). Choice of spin labeling techniques.

For CBF measurement in adults, pseudo-continuous ASL (PCASL) [33] is considered the current state-of-art spin labeling technique [34] and is now widely available from many MR vendors. The spin labeling plane (the green line in Fig 1) should be positioned to be approximately orthogonal to the big feeding arteries . The velocity-encoded ASL [35] can theoretically avoid post-labeling delay (PLD) and may become another state-of-art spin labeling technique after it fully matures. To measure CBF in infants or babies, the pulsed ASL (PASL)[34, 36, 37] rather than PCASL is recommended because arterial blood flow in these populations is very high and PCASL is difficult to label the fast moving arterial blood. In addition, PCASL in infants or babies will be subject to the strong arterial pulsations because the brain is very close to the heart. For organ perfusion imaging, PASL may also be a good choice due to the low blood flow and the complex tomography of the feeding arteries.

2). Choice of ASL MRI sequences.

ASL is a preparation module in ASL MRI sequence. After spin labeling and the post-labeling delay time, MR images with or without spin labeled can be acquired with different regular imaging sequences. Depending on the imaging readout, they can be either a 3D imaging sequence or a 2D one. A 3D imaging sequence is preferred because it has higher SNR than 2D sequences and can achieve better background suppression. Two types of 3D readouts have been used in ASL MRI: the stack or spirals[33, 38, 39] and the gradient- and spin-echo (GRASE)[40, 41]. The spiral-out readout has shorter TE than GRASE and relatively shorter readout time, resulting in higher SNR and less T2-weighting induced blurring. Single-shot is widely used in 3D ASL MRI and the typical parameters can be: labeling time=1.5-2s, PLD=1.5-2s (longer PLD is recommended for old people, patients with neurovascular disease or AD), BS=on, shortest TE, minimal TR, within-plane resolution= 2.5-3.4 mm, number of partitions=36-50, partial Fourier encoding along partition encoding direction=5/8 or 6/8, slice thickness=3-5 mm, within-plane imaging matrix=96x96 or bigger, number of C/L pairs=5-30. Parallel imaging with an acceleration factor of 2 or 3 should be used to shorten the total readout time for 3D ASL. Inflow suppression should be turned on if available. To achieve high resolution, e.g. 2x2x2, multi-shot can be used with each shot covering one part of the 3D encoded space. Between-shot motions will be problematic in multi-shot ASL. 2D ASL MRI is still used in many centers. For the 2D echo-planar imaging readout-based ASL MRI, a typical protocol can have the following parameters: labeling time=1.5-2s, PLD=1.5-2s (longer PLD is recommended for old people, patients with neurovascular disease or AD), partial Fourier along phase encoding direction=5/8 or 6/8, shortest TE, minimal TR, within-plane resolution=3-3.4 mm, number of 2D slices=18-28, slice acquisition order=ascending, slice thickness=4-7 mm, gap between slices=5-10%, within-plane imaging matrix=64x64 or 96x96, number of C/L pairs=30-50. Parallel imaging with an acceleration factor of 2 or 3 should be used to shorten the total readout time for each slice. Inflow suppression should be turned on if available. The multi-band 2D imaging sequence[42, 43] has been incorporated into ASL MRI but is not widely used at this moment.

3). Single PLD or multiple-PLD.

ASL MRI is often played with a single PLD by assuming that the PLD is sufficiently long to allow the labeled spins reach the imaging site. However, this assumption could be wrong especially in old individuals or patients with AD or neurovascular disease as the arterial transit time (ATT) is often much longer than usual in those populations. Obviously, a PLD shorter than ATT will lead to inaccurate CBF quantification because not all labeled spins have reached the imaging place. To accurately estimate the arterial transit time (ATT) and CBF, ASL MRI can be repeated several times, each time with a different PLD. The perfusion signal at the multiple-PLDs can be fitted to a kinetic model[44] to get ATT and CBF. For multiple-PLD ASL, the imaging parameters can be chosen similarly to what have been described above except for PLD and number of C/L pairs. Depending on the total available time, 5-7 PLDs can be chosen. For example, 0.8s, 1.1s, 1.6s, 2s, 2.5s can be used in a 5-PLD protocol for healthy adults; 0.8s, 1.2s, 1.8s, 2.2s, 2.8s can be used for old subjects or patients with AD or neurovascular disease. To limit the total scan time, few C/L pairs can be used for the short PLDs and more C/L pairs should be used for long PLDs because of the reduced SNR when PLD increases. For example, if the 3D ASL MRI sequence is used, the first several PLD ASL scans can take 4-6 pairs of C/L images, and the last 2 PLDs can take 10-16 pairs. A time efficient alternative approach to the multiple-PLD ASL imaging protocol is the Hadamard encoded ASL[45, 46], but it needs a new sequence and may not be available at every imaging center.

4). M0 scan.

ASL CBF quantification needs a M0 scan, which is often performed using the same data acquisition sequence but without the spin labeling module, BS, and PLD. TR should be set to be a big number such as 6s. TE should be set to be the minimum. M0 scan is often played before the ASL scan.

Traditional ASL MRI processing pipeline and strategies for each processing step

As mentioned above, ASL data often contain many pairs of C/L images. As shown in Fig 3, a typical processing pipeline contains the following steps: 1) motion correction, 2) denoising and pairwise subtraction, 3) arterial transit time (ATT) estimation and CBF quantification, 4) outlier cleaning and mean CBF calculation, 5) partial volume correction, 6) spatial registration and regional CBF extraction.

Fig. 3.

A typical ASL MRI data processing pipeline consists of several steps: 1) starting from the acquired C/L (control/label) images, the first step is to correct motions in relative to the mean image (by default) or suppress and temporal noise; 2) the C/L image pairs are then successively subtracted to obtain a perfusion-weighted image series, which are then converted to quantitative CBF in units of ml/100g/min; 3) The outlier CBF images in the entire time series are removed using outlier cleaning, and the remaining CBF images are averaged to get the mean CBF map; 4) partial volume correction is used to correct grey matter CBF values due to contamination from white matter and CSF each voxel (the voxel size in ASL MRI is often >3x3x3 mm³); 5) the partial volume corrected mean CBF and the CBF image series (if time series analysis is required) are recorded on the fixed brain and used for group level analysis.

Motion correction

Assuming the shape and structure do not vary across the scan (minor to modest changes can be caused by the interactions between motion or physiological changes and MR field), rigid body motions can be estimated and corrected for ASL MRI using the methods established in the blood-oxygen-level-dependent (BOLD) fMRI [47, 48]. However, the spin labeling and non-labeling scheme induced systematic signal modulation may be artificially picked up by the affine-transform based motion estimation method. This can be corrected by regressing out the zig-zagged label-control paradigm from the estimated motion parameters before correcting them from the images[13, 23].

Denoising

ASL data contains random noise, residual motion artifacts, physiological noise. For most of ASL MRI data, the resolution is often low (around 3.4x3.4x6 mm³), random noise can be readily suppressed with spatial smoothing (Jiong J. Wang et al. 2005). For high resolution ASL data, Wavelet denoising can be used to simultaneously suppress random noise while reserve spatial resolutions for ASL MRI[19]. Temporal noise in ASL MRI L/C image series can be removed using high-pass filtering as the perfusion signal encoded in the label-control acquisition paradigm locates in the high frequency band [9, 23]. More advanced temporal denoising can be achieved through the component based noise correction method (CompCor) [49] and global signal regression[23].

CBF quantification

CBF quantification is typically based on the one-compartment model[34] which converts the M0-calibrated L/C difference into CBF in a unit of ml/100g/min. Two approaches can be used to calculate the mean CBF. One is to convert the mean perfusion weighted L/C difference image into the mean CBF image; the other is to calculate a CBF image for each L/C difference image and then take the average. Instead of directly calculating the mean difference map through averaging, a mathematically equivalent approach is to fit the L/C image series to the zig-zag labeling and control labeling paradigm and calculate the mean L/C difference from the fitted signal [16]. A robust fitting approach can be adopted to improve the fitting robustness [50].

Outlier cleaning

In data analysis, outliers are those showing large offset, e.g. 3 standard deviations, from the mean or median of the entire data distribution. Outliers can significantly affect the mean value for a small sample size. They can also dramatically affect data fitting stability and generalizability. For image analysis, outliers are those images that differ dramatically from the average image or image patches that differ dramatically from the average patches of the corresponding patch locations. Outliers can be a big issue in ASL MRI due to the relatively low SNR, fluctuations of labeling, the inevitable head motions, and especially the limited number of samples. The robust fitting method [50] can mitigate the influence of outliers at voxel-level but can not consider the spatial information into account. Similar issues exist in the M-estimator based outlier signal clamping method[51]. A more effective strategy is to identify the outlier perfusion difference maps or the CBF maps from the entire acquired time series and remove them before calculating the final mean difference or CBF image. A few methods have been proposed over the past decade. In [13], outliers were defined based on motion patterns (amplitude and first derivative) and whole brain CBF time series (mean and standard deviation). Tan et al. [18] defined the outliers based on the mean and standard deviation of each CBF map. These methods may not be sufficient to find all outliers because outliers can be caused by non-motion related sources and the mean CBF is already contaminated before outlier removal. To be more flexible and independent of specific prior information, an adaptive outlier cleaning algorithm (AOC) was proposed to iteratively find outlier CBF volumes based on the correlation of each remaining CBF volume to the current mean CBF image[24]. AOC was later improved by Dolui et al. by incorporating structural information regularization in the Structural Correlation based Outlier Rejection (SCORE) method [26] and by Li et al. in a prior-guided slice-wise AOC (PAOC) method [27, 52]. Fig. 4 shows an example of the effectiveness of different outlier cleaning methods. Both SCORE and PAOC improved mean CBF image quality and PAOC showed more improvement.

Fig. 4.

Outlier cleaned ASL CBF maps from a representative ADNI subject. NAOC means no outlier cleaning; SCORE stands for structural correlation based outlier cleaning (SCORE); PAOC stands prior-information guided adaptive outlier cleaning. Green ovals indicate areas with the most apparent image quality enhancement.

Partial volume correction

Because of the low spatial resolution used in a typical ASL scan, voxels in ASL images are more likely to be composed of different types of tissues such as grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF). As a result, CBF in a GM voxel is weighted sum of CBF of the pure GM and CBF of pure WM and is smaller than the real GM CBF. This resolution-induced CBF underestimation is called the partial volume effect (PVE). Post-processing based PVE correction is often performed using the proportion of GM and WM estimated by the concurrently acquired high resolution structural MRI and the GM/WM ratio estimated from PET [53, 54]. A data driven approach has been proposed to correct PVE through within-neighborhood data regression by assuming that GM and WM baseline MR intensity and CBF are locally homogeneous [55]. This approach represents a better solution as it can estimate the region specific GM/WM CBF ratio. Meanwhile, by assuming a constant GM/WM MR signal intensity ratio and GM/WM CBF ratio, it inevitably introduces spatial blurring to the PVE corrected CBF map. Instead of explicitly correcting PVE from the CBF images, Chen et al.[56] proposed to include tissue concentration map estimated from the structural MRI can be included as a nuisance covariate during statistical analysis. This however may substantially remove the effects of interest if the trend of tissue density correlates with the regressor. Certainly, every model has its drawback and the ultimate solution for PVE is to increase the spatial resolution of ASL MRI.

Spatial registration

This can be done by registering mean control image to the structural MRI and then to the standard brain such as the Montreal Neurology Institute (MNI) brain. In background suppressed ASL MRI or spiral readout-based ASL MRI, the control image might be too noisy or have excessive artifacts. An alternative approach is to directly register the mean CBF image to the structural MRI and then to the standard brain.

Machine learning-based ASL MRI data processing

Traditional signal processing methods depend on theoretical or empirical models for either the signal or noise which may not be sufficient for the high complex human brain data such as ASL images. By contrast, machine learning approaches are independent of prior-knowledge and can remove artifacts that are difficult to model in an analytical way. Two challenges of machine learning algorithms are the parameter selection and generalizability of the machine learned model. Parameter selection in machine learning often needs exhaustive trials-and-errors and may need additional tuning for new dataset. Large data are required to train a generalizable machine learning algorithm but may not be available. Due to these two challenges, machine learning in ASL MRI is still in an exploratory stage. Over the past decade, there have been many publications in this direction, and we were only able to cover some of them.

Principal component analysis (PCA) was likely the first data-driven approach to be applied in ASL MRI either as a way to reduce the dimension of the confounding temporal artifacts or directly suppress spatial-temporal noise [49, 57]. Independent component analysis (ICA) has been applied to extract noise components and subsequently remove them from the CBF image series[58]. Using PCA to extract the most prominent components from the confounding variables and regressing them out from the acquired data has been well accepted but using PCA or ICA to directly remove noise is still subject to the criteria for selecting the “noisy” components.

Because the systematic spin labeling and non-labeling automatically group the ASL raw images into two categories: the L and C images, finding the C-L difference in a multivariate way (across all voxels) can be done through a two-class image classification process. Based on this observation, the support vector machine (SVM) was adopted in [25] to extract the feature map from the classifier that maximally separates L images from the corresponding C images. This multivariate machine learning-based approach showed better performance than traditional ASL CBF quantification method. However, the model needs to be trained for each subject separately; quantification still depends on a scale estimated from the original L and C images which may have already affected by noise. The low-rank and sparse decomposition was adopted as a spatio-temporal denoising tool in ASL MRI [59, 60]. This method separates the ASL CBF image series into a slowly changing component (the low-rank part) and a spatially sparse component, and removing the noise from the sparse component. Similar process can be applied to the L and C image series separately before CBF quantification. While this method showed increased sensitivity of ASL MRI for task activation detection and functional connectivity analysis, it is challenging to find the optimal parameter for balancing the extent of splitting the original signal into the low-rank and the sparse components. The empirically identified parameter value may subject to change for different subject. In [30], Spann et al. proposed a total generalized variation (TGV) regularized spatial-temporal filtering algorithm for denoising the raw ASL images. The TGV is designed to calculate the denoised version of the L and C images by penalizing the spatial temporal total variations of the residuals of the L or C images and simultaneously penalizing the spatial temporal total variations of the perfusion difference maps. Using simulated ASL data with different level of noise, this algorithm has shown better denoising performance than several state-of-art image denoising methods. However, there are four parameters in TGV for controlling the total variations and image sparsity to be tuned and the entire denoising process takes longer time than other methods due to the iterative process involving the 4D data.

The most recent development of machine learning in ASL MRI highlights the deep learning (DL) neural networks. Over the years, DL has been used for ASL denoising and resolution improvement[61-63], denoising and scan time reduction[64-76]. Fig. 5 shows CBF calculation results based on a representative healthy subject’s 2D PCASL data. The DL method was the super-ASL designed to extract CBF signal directly from the label images without using the control images[67]. Compared to the traditional method, super-ASL showed improved CBF map contrast in the mean CBF image. Remarkable improvement was seen at each time point as evidenced by the substantially improved image contrast and much less artificial negative CBF value (the black holes) as shown in Fig. 5D. These results certainly need to be further repeated using different types of ASL MRI. A great benefit of DL for ASL MRI processing is that it can potentially solve several problems at the same time. For example, it may simultaneously increase spatial and temporal resolution and increase SNR. It may also generate CBF maps from non-ASL data such as PET FDG [61, 77] or BOLD fMRI [78, 79]. The big challenge, however, is to establish the model generalizability given the many complex situations that can affect CBF, such as age, disease, different ASL MRI sequences and MR scanners. Thus far, only one paper has explicitly addressed this kind of DL generalizability in ASL MRI. Nevertheless, DL is certainly a highly promising direction in ASL MRI. Since this direction is still in rapid investigation and many explorations are in the early stage, substantially reviewing the progress and state-of-arts in this direction should be delayed to a future date.

Fig 5.

Representative results of traditional ASL CBF calculation method (A, C) and DL-based ASL CBF extraction method (B, D) based on 2D PCASL data of a healthy subject. 45 C/L pairs were acquired. A and B are four slices of average CBF maps from the traditional method and super-ASL. C and D is one image slice at four different time points.

Conclusion

ASL MRI post-processing should contain motion correction, raw data denoising, CBF quantification, outlier cleaning. Denoising is in general beneficial given the inherently low SNR of ASL MRI but a tradeoff should be taken between suppressing noise and reserving signal especially when we often do not know the difference between signal and noise. While traditional methods based on empirical models have been shown to be effective, machine learning has started to show new promise for further separating noise or artifacts from the acquired ASL data or offering new functions such as CBF generation from other types of neuroimaging data or totally removing the control scan which are not possible using traditional methods.