Improved accuracy of human cerebral blood perfusion measurements using arterial spin labeling: Accounting for capillary water permeability
Abstract
A two-compartment exchange model for perfusion quantification using arterial spin labeling (ASL) is presented, which corrects for the assumption that the capillary wall has infinite permeability to water. The model incorporates an extravascular and a blood compartment with the permeability surface area product (PS) of the capillary wall characterizing the passage of water between the compartments. The new model predicts that labeled spins spend longer in the blood compartment before exchange. This makes an accurate blood T1 measurement crucial for perfusion quantification; conversely, the tissue T1 measurement is less important and may be unecessary for pulsed ASL experiments. The model gives up to 62% reduction in perfusion estimate for human imaging at 1.5T compared to the single compartment model. For typical human perfusion rates at 1.5T it can be assumed that the venous outflow signal is negligible. This simplifies the solution, introducing only one more parameter than the single compartment model, PS/vbw, where vbw is the fractional blood water volume per unit volume of tissue. The simplified model produces an improved fit to continuous ASL data collected at varying delay time. The fitting yields reasonable values for perfusion and PS/vbw. Magn Reson Med 48:27–41, 2002. © 2002 Wiley-Liss, Inc.
Perfusion imaging by arterial spin labeling (ASL) has proved a very useful technique both for research and clinical practice. With careful application ASL has the potential to measure perfusion more accurately than the gadolinium bolus tracking approach and also to be used repeatedly without the need for or cost of an exogenous contrast agent. It could, therefore, be more useful in a number of applications. For example, tracking perfusion changes with time after stroke or following drug treatment; measuring perfusion differences between subjects; and perfusion-based functional studies. Its full potential has yet to be realized, partly due to poor accuracy and precision leading to a lack of confidence in the technique. In previous work (1) we have shown that the precision of the continuous ASL (CASL) (2) technique is good. In this article we present an improvement in the accuracy of perfusion quantification through the use of a two-compartment exchange model that takes account of capillary wall permeability. A preliminary version of this work has been presented in abstract form (3).
Previous models for the signal from arterial spin labeling experiments have assumed that water is a freely diffusible tracer (4-7). There is evidence suggesting that this is not the case, especially in the brain, where the blood–brain barrier impedes water flow across capillary membranes. In organs other than the brain a major route of water transport from blood to tissue is through the gaps between the endothelial cells of the capillary wall. The blood–brain barrier is a continuous network of junctional complexes that seal these gaps, leaving the main route for water exchange as being directly across the plasma membrane of the endothelial cells.
Permeability of the capillary wall to water is measured in terms of PS, the permeability (P) surface area (S) product of brain capillaries to water, per volume of tissue. It has been measured by a number of different methods in a number of species. Published values in whole human brain vary from 0.9–1.7 min−1 with a mean value of 1.2 min−1 (8-10). Two of the studies (8, 9) sampled the outflow of venous blood after intracarotid injection of tritiated water to determine the extraction fraction of labeled water and from that calculate PS. The more recent study (10) compared the measured perfusion values from PET when using either [11C]butanol (which is more diffusible than water) or [15O]water as the tracer. The differences in perfusion measurements can be attributed to the lower PS of water than butanol. PS values also vary within the brain from 0.8 min−1 in white matter to 1.5 min−1 in central cortex (10), thought to be due to different capillary densities producing differences in surface area S, rather than changes to vessel permeability, P. Other studies (11) have shown that the water extraction fraction during a single capillary transit is less than 1; equivalent to a finite PS. Outside the brain, capillary PS values for water are typically a magnitude larger. One study in the rat measured PS of cerebral cortex as 3.3 ± 0.2 and tongue muscle as 14.1 ± 2.5 ml min−1 g−1 (12), supporting the idea that the blood–brain barrier is a major barrier to water transport.
These studies have firmly established that the capillary wall has limited permeability to water. In this article we demonstrate the effect this could have on perfusion measurements using ASL, considering both CASL (2) and the pulsed technique of flow-alternating inversion recovery or FAIR (6). We re-derive the ASL model with the inclusion of vessel permeability in a two-compartment model. Simplifications to the model reduce the number of free parameters and enable the fitting of in vivo data. CASL data at varying delay times is used to test the accuracy of the current single compartment model and the simplified two-compartment models.
MATERIALS AND METHODS
FAIR
The FAIR (6) technique uses a 180° nonselective RF pulse to invert the longitudinal magnetization of the protons in the head. Inverted or “labeled” protons in the blood outside the image slice flow into the slice, changing the longitudinal magnetization within the slice. After an inversion time (TI) the longitudinal magnetization of this “labeled” image is measured using EPI. For the control image a selective RF pulse is applied, inverting only the protons in the image slice. This time the blood flowing into the slice is unlabeled. The difference in signal between the control and labeled images will depend on the rate at which blood enters the slice, and hence perfusion. Figure 1a shows the tissue magnetization of the difference image ΔM (i.e., control-labeled) with inversion time along with the longitudinal arterial magnetization in the difference image Δma which decays with the T1 of blood, T.
CASL
CASL (2) uses a spatially localized RF field to continuously invert the longitudinal magnetization of the protons in arterial blood water entering the brain. A continuous RF field is applied for labeling time tL, along with a field gradient, such that the RF field is on resonance in a plane through the carotid artery, called the inversion plane. Flowing protons move through the field gradient and their resonant frequency increases. As they pass through resonance the magnetization of the spins is inverted, effectively “labeling” the blood. On arrival at the image slice, after arrival time tA, the magnitude of the longitudinal magnetization of the arterial blood has decreased (i.e., Mz has become less negative) due to T1 relaxation but is nevertheless different from the fully relaxed magnetization of the tissue. The arterial blood reduces the longitudinal magnetization in the slice. At the end of labeling a short delay is introduced, tD, and the longitudinal magnetization in the slice is measured using echo-planar imaging (EPI). A second, control image is taken without arterial spin labeling. Figure 1b shows the tissue magnetization of the difference image ΔM (i.e., control-labeled) with time after the start of labeling. There are three parts to this curve: part 1 has no signal since the labeled blood has yet to reach the tissue volume (t ≤ tA); part 2 shows rising signal as the labeled blood continues to enter the slice (t ≤ tL + tA); and part 3 shows decaying signal due to longitudinal relaxation and outflow after the end of incoming arterial signal (t > tL + tA). The longitudinal arterial magnetization Δma is constant during part 2 due to the continuous labeling pulse.
The subtraction image for both ASL techniques has static tissue signal removed with only the signal from the labeled water protons remaining.
THEORY
There are currently two main approaches in describing this dynamic system mathematically. The first (4) describes the system in terms of modified Bloch equations. The second (13) considers the signal in the difference image to come from a bolus of labeled spins traveling through the vasculature, mathematically convolved with a tissue residue function—similar to the gadolinium bolus tracking models. Progress has been made in extending both types of model to include the effects of delayed blood arrival time (14), magnetization transfer (15) arterial and venous compartments (14, 16), and restricted vessel permeability (11, 16-19).
Earlier work on restricted vessel permeability (11) describes the effect in terms of a reduction in the water extraction fraction, E, during a single capillary transit. A single well-mixed tissue compartment is assumed and E is introduced as a correction factor. More recent work (19, 20) measures E by sensitizing the signal to either blood or extravascular water. The first study, by Silva et al. (11), used diffusion gradients to separate the signal contribution from slowly diffusing water (in the extravascular space) and quickly diffusing water (in the blood). The second study, by Zaharchuk et al. (19), used an intravascular contrast agent to increase the transverse relaxation rate of blood, hence removing the signal from blood water protons. Both studies measured an extraction fraction of less than one, which decreased with increasing perfusion.
More recent work includes restricted permeability into the model through the use of PS, the permeability surface area product. The work of Zhou et al. (17), based on a model by Schwarzbauer et al. (12), extends the Bloch equation model to include both a blood and an extravascular compartment. It is assumed that the blood compartment is well-mixed, i.e., as soon as water enters the blood compartment it instantaneously equilibrates throughout the compartment. This work shows that at high field strengths in an animal model (when flow is higher than in humans), the current single compartment model will underestimate perfusion. Two other studies (16, 18) do not assume a well-mixed blood compartment, allowing the blood signal to change with both distance and time. The work by St Lawrence et al. (16) extends the bolus model to include separate terms in the residue function for blood and extravascular compartments. The model is extended further to include a venous compartment. The model predicts an overestimation of perfusion at low perfusion rates and an underestimation at high perfusion if the current single compartment model is used. The work by Ewing et al. (18) describes the changes to the model in terms of changing the effective T1 of the experiment. For flows up to 150 ml blood min−1 (100 ml tissue)−1 at 7T this work shows that the single compartment model will overestimate perfusion. The widely used model of Alsop et al. (14) should also be mentioned. While this model does not consider the effects of restricted water permeability, the inclusion of an arterial blood compartment in addition to the well-mixed tissue compartment could produce similar signal curves to a true two-compartment model.
Much of this previous work has been theoretical, with little validation of the models. The aim of our work is to use a relatively simple model with few free parameters so that it can be tested with in vivo data. We extend the Bloch equation model to include exchange between the blood and extravascular compartments in terms of PS. It is assumed that the labeled blood water enters the voxel in exchanging vessels and also leaves through exchanging vessels (i.e., there are no “arterial” or “venous” compartments). As Ewing et al. (18) notes, this can include vessels other than the classic definition of a capillary, for example, postcapillary venules. By using a CASL delay time of greater than 0.5 s, most blood passing through the voxel in larger vessels is excluded from the measurement. A second assumption is that the blood compartment is well-mixed, which forces assumptions to be made about the venous magnetization flowing out of the voxel. However, these assumptions may be valid in certain systems (for example, human perfusion rates at 1.5T, the venous magnetization can be assumed to be zero), and allow for further simplification of the model solution. This enables the model to be tested on in vivo data. Before introducing the two-compartment theory, we reformulate the single compartment theory making use of a method first introduced by Alsop et al. (14) that uses the Bloch equation model but considers the signal directly in the difference image. This approach makes for easier handling of the more complex two-compartment system.
Single Compartment Theory
f | Blood perfusion | ml blood/min/100ml tissue |
PS | Permeability surface area product | ml water/min/100ml tissue |
λ | Brain:blood partition coefficient λ = (vew + vbw)/v | ml water(ml tissue)−1/ml water(ml blood)−1 |
tA | Arrival time (from labelling plane to image slice) | s |
tL | Labelling time | s |
G | Gain | signal per unit of magnetisation |
α | Inversion efficiency of labelling pulse | fraction of blood water protons inverted |
M(t) | Tissue magnetisation | magnetic moment (ml tissue)−1 |
M0 | Equilibrium tissue magnetisation | magnetic moment (ml tissue)−1 |
mb(t) | Capillary blood water magnetisation | magnetic moment (ml water)−1 |
me(t) | Extravascular water magnetisation | magnetic moment (ml water)−1 |
m | Equilibrium magnetisation of extravascular water | magnetic moment (ml water)−1 |
ma(t) | Arterial blood magnetisation | magnetic moment (ml blood)−1 |
m | Equilibrium magnetisation of arterial blood | magnetic moment (ml blood)−1 |
mv(t) | Venous blood magnetisation | magnetic moment (ml blood)−1 |
vbw | Blood water volume vbw = vb · v | ml water/ml tissue |
vew | Extravascular water volume vew = ve · v | ml water/ml tissue |
vb | Blood volume | ml blood/ml tissue |
ve | Extravascular volume | ml extravascular space/ml tissue |
v | Blood water content | ml water/ml blood |
v | Extravascular water content | ml water/ml extravascular space |
T1 | Longitudinal relaxation time of water in tissue | s |
T1b | Longitudinal relaxation time of water in blood | s |
T1e | Longitudinal relaxation time of water in extravascular space | s |
Two-Compartment Theory
In reality, the labeled blood water does not exchange immediately with the extravascular water, so using a single well-mixed compartment is inaccurate. As described in several studies (16-18) this leads to two opposing effects when measuring perfusion.
First, if labeled water molecules remain in the blood for some time before entering the extravascular space, they will relax with the longer blood relaxation time, T which is measured as 1.4 s (22) compared to extravascular T1, T which could be as low as 0.6 s in white matter (23) at 1.5T. The signal decay will be slower and the tissue magnetization larger than predicted by the single compartment model. We call this the T1 effect. The application of diffusion gradients which aim to crush the intravascular signal have little effect in such small vessels (24) and the signal we measure is from both capillaries and tissue. Second, some of the labeled blood water molecules will pass directly through the microvasculature without ever exchanging. The venous magnetization mv will be larger and the tissue magnetization lower than predicted by the single compartment model. We call this the outflow effect.
Blood Compartment
Extravascular Compartment
Note that for the CASL technique the magnetization transfer (MT) effects of the labeling pulse will significantly reduce T. The MT effects will reduce with time for t > tL after the labeling pulse is switched off, giving a time-varying T.
Solutions to the Two-Compartment Model
Equations [11] and [12] can be solved to find mb and me and hence the total magnetization using Eq. [8], provided we know Δma and Δmv. Δma can be calculated dependent on the labeling strategy. However, the assumption that Δmv = ΔM/λ as for the single compartment theory cannot be made since we are allowing for the possibility that equilibrium between blood and extravascular water molecules is not reached.
It may seem natural to assume that the venous blood has the same magnetization as the voxel blood compartment. However, this will overestimate the venous magnetization since at short times after labeling the outflow of labeled molecules will be small, as they have not yet had time to travel through the vasculature. For a precise definition of Δmv the distribution of labeled molecules with distance along the vessels within the voxel must be considered (16). In truth, a compartmental model can never deal with this situation properly and assumptions must be made. However, in certain conditions reasonable assumptions can be made about Δmv that will simplify the model solutions.
Slow Solution
We can assume that the labeled water never leaves the tissue voxel during the measurement time, thus giving zero venous magnetization, i.e., Δmv = 0. For a typical human system the mean transit time (MTT) of labeled molecules (MTT = vb/f ≈ 0.05/0.01 = 5 s) is longer than the measurement time (typically 1 s for FAIR, 3 s for CASL) so this assumption is valid. We call this the slow solution because it is more accurate at slow perfusion rates. The model of St. Lawrence et al. (16) supports this assumption, showing that the outflow effect is negligible for typical human flow rates.
Fast Solution
At the other extreme of fast perfusion, we can assume the venous blood has the same magnetization as blood in the voxel with Δmv = Δmbv. For a higher flow animal system the MTT is shorter than the measurement time (MTT ≈ 0.05/0.05 = 1 s) so this assumption is valid. We call this the fast solution because it is more accurate at fast perfusion rates.
Distributed Solution
For a complete solution we can set Δmv = 0 for t < MTT and Δmv = Δmbv for t > MTT. However, this introduces a discontinuity in the signal curve at t = MTT. To smooth the transition around MTT a distribution of transit times centred on MTT can be used. Each transit time will produce a slightly different signal curve, which is averaged to give the final signal curve. This is equivalent to a range of perfusion rates producing an average perfusion f. We call this the distributed solution. In our model we use a Gaussian distribution of perfusion rates with full width at half maximum of 0.7f. This width was chosen to be the minimum width required to adequately smooth the signal curve. Further work could look at the use of more realistic flow distributions.
Using these boundary conditions on Δma and Δmv for all three models, the coupled differential equations (Eqs. [11] and [12]) were solved manually by means of Laplace transforms (25). Mathematica (26) was used to calculate the individual transforms. The solutions for both CASL and FAIR labeling are given in Appendix Eqs. [16]–[18]. We were unable to solve these equations analytically for part 3 of the CASL curve when t > tL + tA (Fig. 1).
Simplified Solutions Neglecting Backflow
One further general assumption is to neglect backflow of labeled water molecules from the extravascular compartment into the blood. That is, to assume that the magnetization in the extravascular water compartment is much less than that of the blood water compartment, i.e., Δme ≪ Δmb. This relies on the fact that the extravascular water volume is much larger than the blood water volume (typically 20:1) and measurement times are short, such that Δme is far from equilibrium. Simulations of the exchange of water protons between the two compartments, ignoring inflow, outflow, and decay, show that at a typical measurement time of 3 s, Δmb/Δme = 11. This is a very useful simplification since it decouples Eq. [11] and Eq. [12], producing simpler solutions. This is particularly useful in the case of CASL since it enables us to find solutions for the complete signal curve, including part 3. The solutions using Δme ≪ Δmb are given in Appendix Eqs. [19]–[22].
The simplification also has benefits for the slow model, giving a solution with only one more free parameter than that of the single compartment theory; this is PS/vbw. This is of particular interest for human imaging at 1.5T, where we are likely to be in the slow flow regime. In cases where PS/vbw is unlikely to change it could be preset from published data, leaving perfusion as the only free parameter to be determined. Alternatively, it is possible to estimate this potentially useful parameter by sampling the signal curve at a number of time points. Since capillary surface area, S, and volume, vb, both depend on the geometry of the vessels, a simple model with vessels of uniform radius r shows that PS/vbw reduces to 2P/vr. v is a global parameter that could be measured from a blood sample and is simply related to hematocrit. Disease could change either permeability P, vessel radius r, or blood water content v, making PS/vbw a useful measurement in a number of applications. In multiple sclerosis, an increase could indicate early damage to the blood–brain barrier causing an increase in water permeability. During neuronal activation a decrease could indicate vessel dilation. Changes would also be expected in stroke and tumours.
Impermeable Solution
It is possible that a model with a single blood compartment (PS = 0, impermeable) is more accurate than a single tissue compartment (PS = ∞). The impermeable solution was found using Eq. [1] with T1 replaced by T and mv = 0 for the slow solution or mv = M/vb for the fast solution. The solutions are given in Appendix Eqs. [23]–[26]. It can be seen that the slow solution has fewer free parameters than the original single compartment Eqs. [5]–[7] with no need for a tissue T1 measurement. For part 3 of the CASL curve (Fig. 1) when t ≥ tA + tL, Eq. [26], the solution is also independent of blood arrival time tA.
Testing the Two-Compartment Model
METHODS
Simulated Data
Accuracy of the Single Compartment and Simplified Solutions
For all simulations we used the parameters in Table 2 with λ = 0.9 [27], α = 0.7 [2], v = 0.7 (28), tA = 0, and tL = 3s. We began by inspecting the form of the two-compartment solutions. Typical parameters were entered into the solutions and the results plotted. Four systems were considered with parameters as shown in Table 2: human gray matter (GM) and white matter (WM) at 1.5T, 4T, and 7T. The distributed model was used to produce simulated data at a range of perfusion rates. The original single compartment solution, the simplified slow and fast solutions (neglecting backflow) and the slow impermeable solution were then used to estimate perfusion values from the data at time points of t = T for FAIR and t = 3 s for CASL since these are typical measurement times that would be used in practice to maximize signal to noise. Note that since tL = 3 s the CASL signal is in the rising phase (part 2 of Fig. 1). The perfusion values were compared to the original values used to generate the data.
Field (T) | 1.5 | 4 | 7 | |||
T1b(s)a | 1.4 | 1.8 | 2.3 | |||
Tissue | Gray | White | Gray | White | Gray | White |
PS (ml water (min)−1(ml tissue)−1)b | 1.5 | 0.8 | 1.5 | 0.8 | 1.5 | 0.8 |
T1e(s)c | 1.0 | 0.6 | 1.35 | 0.8 | 1.8 | 1.1 |
f (ml blood (min)−1(100 ml tissue)−1) | 60 | 30 | 60 | 30 | 60 | 30 |
vb (ml blood (ml tissue)−1)d | 0.05 | 0.03 | 0.05 | 0.03 | 0.05 | 0.03 |
- a 1.5T (22), 7T (36), 4T values interpolated assuming T1b scales linearly with field strength.
- b From Herscovitch et al. (10).
- c 1.5T (23, 37), GM at 4T (38), gray matter at 7T (36). Values for white matter were interpolated assuming the ratio of gray:white T1e remains constant.
- d From Leenders et al. (33).
Errors Due to Fixed Parameters
The inaccuracies of the many fixed parameters in the model (i.e., m, tA, α, T, T, and PS/vbw) will propagate to produce errors in the perfusion values. These errors were determined numerically. The simplified slow solution was fitted, by the method of least squares, to the simulated data of human gray matter at 1.5T using CASL with a slight change to each fixed parameter in turn. The error in perfusion was calculated by comparing the measured value of perfusion with the value used to generate the data. The magnitude of the propagated errors will determine the accuracy to which these fixed parameters must be known if they are not to contribute to the uncertainty of the perfusion estimate.
Imaging
In order to test the accuracy of the two-compartment model, we collected data from three subjects (female, mean age 28 years) using a CASL (2) pulse sequence (described in the Introduction) with gradient echo EPI collection. The imaging was repeated on one subject in order to test reproducibility. In this implementation labeling is psuedo-continuous for a time tL and then there is a postlabeling delay tD before image collection at time t = tL + tD. The postlabeling delay time was varied in order to sample phases 2 and 3 of the signal curve (Fig. 1). Multislice CASL imaging was performed on three subjects with a range of 7 postlabeling delay times from 0 to 1.5 s + ti. ti is the imaging time, which is different for each slice (34–577 ms) depending on the slice collection order. Scanning was conducted on a General Electric Signa 5.6 1.5T scanner with scan parameters: repetition time, TRASL = 4 s (time between the starts of successive labeling pulses), echo time = 34 ms, tL = 1.7 s, 45 averages, 64 × 64 resolution, 24 cm field of view (giving pixel size 3.75 × 3.75 mm), eight slices with 7 mm slice thickness and 2 mm gap, slice collection order 1,3,5,7,2,4,6,8 with slice 1 being the inferior slice. The labeling plane was positioned 4 cm below the inferior slice; control and labeled image collection were interleaved. This perfusion protocol took 42 min.
Analysis
To identify large regions of gray and white matter the difference images were segmented on the basis of T1 values: 0.5–0.9 s (white matter), 1.0–1.8 s (gray matter), and single pixels in the vicinity of large vessels, which were identified as having high signal in both the T1 and perfusion difference images. The single compartment solution, Eqs. [6] and [7], and the simplified two-compartment solutions, Eqs. [21,22], were fitted (using least-squares minimization) to region of interest data from each slice for the free parameters f, PS/vbw, and tA. Fixed parameters were: T = 1.4 s ± 0.1 s, T = 0.75 s (gray matter) 0.52 s (white matter) ± 0.1 s, α = 0.7 ± 0.1, λ = 0.9 ± 0.1. The uncertainties of these values (found from considering a range of published values (2, 14, 22, 27)) were propagated to produce systematic errors in the fitted parameters. For the gray and white matter regions the slow solution was used and for the large vessel pixels the fast solution was used with the extra parameter vb = 0.05 (note that an artery of 1 mm diameter gives vb ≈ 0.05). Note that the T values are chosen to take account of the MT effects of the labeling pulse (14) assuming complete saturation.
Some signal time dependence could come from large vessels that contribute signal at short delay times but have washed through at longer delay times. To avoid this effect only the later six time points with delay times 0.25–1.5 s were used for gray and white matter regions to avoid signal from very large vessels. To test that this effect was negligible the data were refitted using only the later five points (delay time 0.5–1.5 s) and the fitted values were compared to those produced using six data points. At a delay time of 0.5 s all blood flowing at a velocity greater than 1.4 cm/s will have washed through the voxel. This leaves blood in capillaries and larger vessels (up to approximately 0.1 mm diameter (29)) which could go on to perfuse blood in the voxel and so should be included in the measurement.
To calculate random errors a Monte Carlo bootstrapping technique was used to resample the data based on measured errors in the data. These errors were found by performing a scan-rescan test and measuring the difference in signal. This previous study (1) gave, for similar volumes, gray matter raw signal reproducibility (95% confidence limit) of 8% and white matter 14%. One hundred sets of resampled data from each dataset were used to find the standard deviation in the fitted parameters for each original dataset. Average values for perfusion and PS/vbw were found by calculating a weighted mean of values from each dataset. Weights were set as the inverse of the variance for each measurement.
RESULTS
Accuracy of the Single Compartment Solution
The form of the solutions me and mb for pulsed labeling are shown in Fig. 3. They have been multiplied by the relative volumar fractions such that they represent magnetization per unit volume of tissue. It can be seen that mb peaks before me, which is to be expected as the bolus moves through the blood compartment into the extravascular compartment. The average timing of these peaks can be related to the time of exchange, Tex, which was introduced in our initial, crude two-compartment model (30). Figure 4a–d shows the errors of the single compartment model for both pulsed and continuous labeling at a range of field strengths and perfusion rates. At low perfusion rates the single compartment theory overestimates perfusion due to the dominance of the T1 effect, while at higher perfusion rates the outflow effect begins to dominate and perfusion is underestimated. Comparing Fig. 4c with 4a, and 4d with 4b, it can be seen that in white matter the T1 effect is larger due to the greater difference between T and T. These results are in broad agreement with previous studies (16-18) (see the Discussion section).
In addition, the current study also shows the effect of field strength and measurement time. Comparing Fig. 4a with 4b and 4c with 4d, it can be seen how the timing of the measurement affects the interplay of the two effects since the FAIR measurement is taken at an earlier time point (t ≈ 1s) than the CASL measurement (t ≈ 3s). For the later CASL measurement, the outflow effect begins to dominate at a lower perfusion rate because the measurement time is closer to the MTT and unexchanged molecules will begin to exit the voxel. The effect of field strength is to change the T1 values of the blood and tissue, which in turn determine the strength and relative importance of the venous magnetization, Δmv. At higher fields Δmv increases and the outflow effect begins to dominate over the T1 effect at a lower perfusion rate. The distribution of transit times will determine the smoothness of transition between these two effects. For typical human perfusion rates at 1.5T (see Table 2), we find that for CASL the single compartment theory overestimates perfusion by 17% in gray matter and 62% in white matter.
Accuracy of the Simplified Two-Compartment Solutions
Figure 5a–d shows the accuracy of the simplified fast and slow two-compartment solutions at 1.5T. It can be seen that the slow solution is very accurate for the FAIR technique at typical human perfusion rates in both gray and white matter, being almost indistinguishable from the distributed solution. With continuous imaging, the slow solution becomes inaccurate at lower perfusion rates because of the later measurement times that approach MTT. When the MTT is reached, the outflow effect, which is assumed negligible in the slow solution, becomes important. The fast solution is only accurate for higher perfusion rates with the outflow effect being overestimated at slower rates. Using the impermeable solution, perfusion is underestimated by 3% for gray matter and 7% for white matter with FAIR, and 9% for gray matter and 20% for white matter with CASL at 1.5T.
Errors Due to Fixed Parameters
Figure 6 shows the propagated error in the perfusion estimate for each unknown parameter for gray matter at 1.5T using the parameters in Table 2. While PS/vbw need not be known very accurately, the parameters T, m, and α must be more precise since they have a larger effect on the perfusion measurement. For less than 5% error in perfusion measurements in gray matter (an acceptable limit), the parameters must be known to the following accuracy: α: 5%; m: 5%; T: 7%; tA: 21%; T: 27%, and PS/vbw: 100%. These values represent the accuracy we must strive for. While the accuracy of the T1 measurement is not too important, it is clear that accuracy of blood parameters T and m are very important—an issue which has been somewhat overlooked. An accurate measurement of T is difficult due to the movement of the blood. Note that the use of T1app rather than T1 in current single compartment theory (see Eq. [4]) is insignificant in terms of the accuracy of the perfusion estimate. T1app is on the order of 1% smaller than T1, which produces only a 0.2% error in perfusion, which is negligible when compared to other errors.
Imaging
Figure 7a–d shows the typical performance of the single and two-compartment models when fitting for perfusion, arrival time, and PS/vbw. They show that the two-compartment model fits well to both extremes of data, i.e., white matter and tissue containing a large vessel. It can be seen that the single compartment model decays too quickly to fit the white matter data decay (Fig. 7b), presumably because the labeled water remains in the blood compartment with slower longitudinal relaxation. In some cases the single compartment model compensates for the inability to fit by shifting the curve and overestimating the arrival time, as shown in Fig. 7c. For the single pixels containing large vessels (Fig. 7d) the single compartment model decays more slowly than the data, presumably because of outflow of nonexchanging molecules from the blood compartment.
Table 3 shows the results of the fitting for perfusion and PS/vbw averaged over all datasets (as described in the Methods section) for three subjects. PS/vbw is found to be significantly different in different tissue types, being larger in gray matter. There is a wide spread in both perfusion and PS/vbw measurements. The repeat scan on Subject 1 shows high reproducibility of perfusion measurements (6% change in whole brain value) but poor reproducibility of PS/vbw (59% change in whole brain value). This perfusion change is within intrasubject biological variability found by repeat scanning at a single delay time throughout the day (21). The errors on the PS/vbw measurements are large, reflected in the poor reproducibility of this measurement. The results for fitting only the later five data points are not significantly different than with six data points, suggesting that the effect of large vessel through-flow is negligible.
Tissue | Subject 1 | Subject 1 repeated | Subject 2 | Subject 3 | ||||
---|---|---|---|---|---|---|---|---|
gray | white | gray | white | gray | white | gray | white | |
f (ml blood (min)−1(100 ml tissue)−1) | 98 | 30 | 92 | 29 | 71 | 22 | 84 | 23 |
Random errora | 3 | 3 | 2 | 2 | 2 | 1 | 2 | 1 |
Systematic errorb | 15 | 6 | 14 | 5 | 11 | 4 | 13 | 3 |
PS/vbw (ml water (min)−1(ml tissue)−1) | 26 | 7 | 12 | 2 | 16 | 2 | 22 | 0.7 |
Random errora | 4 | 8 | 3 | 6 | 4 | 4 | 5 | 4 |
Systematic errorb | 11 | 3 | 7 | 1 | 7 | 1 | 10 | 0.4 |
f, 5 data points | 97 | 30 | 93 | 28 | 73 | 24 | 82 | 24 |
PS/vbw, data points | 33 | 9 | 12 | 3 | 21 | 5 | 31 | 1.5 |
- a 2 × standard deviation due to noise in the data.
- b Uncertainty due to uncertainties in the fixed parameters.
Figure 8 shows the arrival time results averaged over all subjects. We find arrival time to be approximately linear with distance from the labeling plane with a slope in agreement with published values (31). The offset with the origin could be due to the blood traveling faster in the larger arteries closer to the labeling plane. The value at 4 cm agrees with estimates of arrival time (22) of approximately 200 ms based on artery blood velocity of 20 cm s−1. The white matter arrival time is found to be longer than the gray matter arrival time, which could be due to the lower white matter perfusion reflecting a lower blood velocity in feeding vessels (14, 31).
DISCUSSION
Simulations
This work shows that a single compartment model can significantly underestimate or overestimate perfusion, depending on the perfusion rate of the system and the field strength. In the human regime at 1.5T the single compartment theory overestimates perfusion by 62% (white matter) and 17% (gray matter). This error is strongly dependent on the blood T1, T. Recent work (32) has measured blood T1 as 1.4–1.7 s at 1.5T, depending on hematocrit, which will increase the errors of the single compartment model. The work by St. Lawrence et al. (16) shows almost identical results for their single pass approximation model, without the assumption of a well-mixed blood compartment. This helps to verify our handling of the venous outflow (Δmv = 0) to correct for the assumption of a well-mixed blood compartment. Interestingly, the inclusion of a venous compartment in the St. Lawrence model significantly changes the simulation results, reducing the outflow effect. It is therefore important to establish if this venous compartment is required. The work by Zhou et al. (17) does not show an overestimation by the single compartment model at low perfusion rates. There are two possible reasons for this discrepancy. First, Zhou et al. assumed a smaller value of T, which will reduce the T1 effect. Second, the model used did not take account of the overestimation of venous outflow that results from assuming a well-mixed blood compartment. This will overemphasize the outflow effect, and therefore mask the T1 effect. Ewing et al. (18) predict the crossover between the two effects occurs around 150 ml blood min−1 (100 ml tissue)−1 at 7T. From Fig. 4a this study gives a similar value of 130 ml blood min−1 (100 ml tissue)−1.
In addition, we find that the balance of the T1 effect and the outflow effect depends not only on perfusion and relaxation parameters but also on the measurement technique and the measurement time. FAIR is more sensitive to the T1 effect because measurements are made about 1 s after labeling, when there is negligible outflow of labeled protons. The T1 effect dominates up to flow rates in excess of 200 ml blood min−1 (100 ml tissue)−1 at 1.5T. CASL imaging is typically around 3 s after the start of labeling, by which time there is some outflow of labeled protons. The outflow effect dominates for flow rates in excess of 100 ml blood min−1 (100 ml tissue)−1 at 1.5T. In either case, for human scanning at 1.5T it is the T1 effect that dominates and the single compartment model will overestimate perfusion. During human neuronal activation it is likely that the increased perfusion will change the balance of these two effects. It is important to consider what effect this will have on the analysis of functional imaging studies. Note that the simulations assumed zero arrival time. An increased arrival time would tend to delay the outflow effects, causing the T1 effect to dominate for a wider range of perfusion rates.
The impermeable solution simplifies the model even further to a single blood compartment. For FAIR, errors are found to be only 3% in gray matter, much smaller than assuming the original well-mixed single compartment model. For CASL, at the later measurement time, errors are larger but still an improvement on those of the original single compartment model. The advantage of the impermeable model is that the tissue T1 map is no longer required for quantification. This is particularly attractive since it will reduce the scan time and the need for registration of the perfusion images to a T1 map. The error analysis shown in Fig. 6 supports this finding, showing that the accuracy of T is not very important for an accurate gray matter perfusion measurement. The error propagation ratio will, however, be larger for white matter. This model is clearly a good option for FAIR modeling. It may be advantageous to reduce the CASL measurement time in order to improve the accuracy of the impermeable model. Conversely, the error analysis highlights the importance of an accurate measurement of T and m. This has been somewhat overlooked in the current modeling of ASL data.
In Vivo Modeling
The two-compartment model provides a better fit to CASL decay-time data than the single compartment model over all extremes of data. It is important to note that the model can fit to all types of blood flow, from true capillary perfusion to blood flow in a single large vessel, and will simply measure “perfusion” as the total volume of blood flowing through a tissue volume per unit time. The insensitivity of perfusion measurements to increasing delay time after 250 ms (shown in the comparison of fitting five and six data points, Table 3) suggests that after this time all of the blood present in the voxel will go on to perfuse that voxel. It is important to remember that any blood which will eventually perfuse the measurement voxel should be included, even if it is in larger arterioles at the time of image collection.
Our assumption that labeled blood water enters the voxel in exchanging vessels is therefore probably incorrect. The inclusion of an arterial compartment may improve the accuracy of the model, but would also increase the number of free parameters in the model. In terms of modeling in vivo data, the term PS/vbw could encompass a number of other factors that would have a similar effect on the signal curve. It is instructive to think of PS as an average PS for all of the vessels in the imaging voxel that contain labeled blood water.
The simplified slow solution allowed the fitting of arrival time and PS/vbw as well as perfusion. However, this method is not suggested as a good way to measure PS/vbw since the reproducibility is poor. We simply show that a realistic, noninfinite, value of PS/vbw is required to correctly model the perfusion signal. The measurements show a tissue-type dependence of PS/vbw. This was initially unexpected since the parameter reduces to 2P/rv, with all components being tissue-independent. However, this contradiction could be explained if we consider the location of the blood in the different tissue types when the measurement is taken. For white matter it is thought (14, 31) that the blood takes longer to reach the capillary bed than in gray matter, so we could be measuring PS/vbw of the arterioles rather than the capillaries, which, as we find, would be lower. This is supported by the trend towards higher PS/vbw values when only the later five data points are fitted, since at later times more of the labeled molecules will have reached the capillaries. It would be interesting to see if the addition of mild diffusion gradients, which would remove some of the arteriole signal, would increase the estimated PS/vbw in white matter. The values for PS/vbw measured from the later five time points are likely to be more accurate estimates of true capillary permeability. The gray matter values of around 25 min−1 agree reasonably well with published values. For example, Herscovitch et al. (10) found PS of 1.5 min−1 (gray matter), giving PS/vbw values of approximately 30 min−1 using typical blood volume of 5% (33).
There are a number of effects besides restricted vessel permeability that could also produce changes to the difference signal with delay time. These include MT and T effects. The MT effects of the labeling pulse will shorten T and this effect will diminish with increasing delay time. In our simulations we used a fixed value for T taken from a measurement at zero delay time. At maximum delay time T could be up to 50% larger (the value without MT effects). If included in the model this time-varying T would produce a more slowly decaying signal curve, which could perhaps be imitated by a lower value for PS/vbw. As a result, measurements of PS/vbw could be underestimated. Simulated gray matter data with varying T (0.75 s to 1.0 s, i.e., saturated to unsaturated) show that assuming a fixed T of 0.75 s estimates perfusion accurately but underestimates PS/vbw by 55%.
There have been a number of studies comparing ASL perfusion estimates using a single compartment model to those using other techniques. The study by Zhou et al. (17) in cats at 4.7T made comparisons with microspheres, finding ASL underestimated perfusion with increasing deviation at higher flow rates. This is in agreement with our predictions (Fig 4a). A study by Ye et al. (35) in humans at 1.5T made comparisons with PET and found good agreement in gray matter but an underestimation of white matter perfusion with ASL, contrary to our model predictions (Fig. 4d). As the authors suggest, this may have been due to an underestimation of white matter arrival time. These results, and other comparison studies, do not all agree with our model predictions, but they are also all fairly inconsistent. This highlights the problem of comparing different perfusion techniques, in that none of them can be assumed to be the “gold standard.” With each technique there are various factors which could cause an error to perfusion quantification. Our approach of testing perfusion models internally on data from a single technique offers an alternative method for model validation.
CONCLUSION
Vessel permeability to water is an important factor that must be taken into account if we wish to obtain truly quantitative estimates of perfusion using arterial spin labeling.
Acknowledgements
We thank David Alsop for the use of the CASL sequence, Gareth Barker for help with implementation, Keith St. Lawrence for helpful suggestions, and the subjects who agreed to take part in this study.
APPENDIX
General Solutions of the Two-Compartment Model
FAIR
Distributed solution: J = A + D for t < MTT and J = A + D + f/vb for t ≥ MTT. Distribution of transit times centered on MTT.
Slow solution: J = A + D.
Fast solution: J = A + D + f/vb.
CASL
Here we consider the solution for the rising part of the curve only, before the trail end of the labeled bolus reaches the imaging slice. We use Δma = 2mα(exp(−tA/T1b) for t ≥ tA and Δma = 0 for t < tA where tA, the arrival time, is the time taken for the labeled blood to travel from the labeling plane to the tissue. We find solutions:
Simplified Solutions to the Two-Compartment Model, Neglecting Backflow
Here are the solutions when using the assumption that me ≪ mb. We use the same constants and the same expressions for Δma as in the previous section with the addition of tL, the duration of labeling used in continuous labeling such that Δma = 0 for t > tA + tL.
FAIR
CASL
Impermeable Solution
Below are the solutions to Eq. [1] with T1 replaced by T with the same expressions for Δma as in the previous sections.
FAIR
CASL
The equations given above are for the slow solution. For the fast solution replace 1/T with 1/T + f/vb. Note that these equations are the same as for the single compartment model Eqs. [5]–[7] with T replaced by T.