While the consensus statement on the recommended implementation of arterial spin labeling (ASL) has advanced ASL to clinical application, variations in labeling efficiency, longitudinal relaxation time of blood and arterial transit times can cause significant quantification errors. With simulation experiments, it is shown that sacrificing ASL scan time for measurements of these parameters improves the estimation reproducibility of the cerebral blood flow on a population level. Furthermore, multi-delay ASL modalities in combination with these extra measurements can compete with or outperform the single-delay consensus implementation in terms of estimation accuracy and precision, depending on the underlying distribution of transit times.
In 2015, the ISMRM perfusion study group and the European consortium for ASL in dementia delineated a recommended implementation of arterial spin labeling (ASL) MRI for clinical application1. While this has pushed ASL towards clinical adoption, several potential quantification confounders remain, such as variations in labeling efficiency ($$$\small{\alpha}$$$), longitudinal relaxation time of blood ($$$\small{T_{1b}}$$$) and arterial transit times ($$$\small{\Delta}$$$$$$\small{t}$$$)2. These parameters can be estimated from additional experiments ($$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$)3,4 or by a modified acquisition of ASL data ($$$\small{\Delta}$$$$$$\small{t}$$$)5,6. However, both the ASL scan as the supporting measurements come with limited SNR posing the question how to distribute scan time in say a five-minute protocol between averaging of ASL data and acquiring supporting measurements. In this work, this dilemma is studied by means of simulations.
The simulation experiment consisted of three parts (Fig.1). First, a physiological state associated with a perfusion process was defined by a set of parameters $$$\small{\boldsymbol{\theta}_i}$$$ drawn randomly from a population distribution $$$\small{p(\boldsymbol{\theta})}$$$ (see Table 1) and by the cerebral blood flow (CBF), which was kept constant at 50mL/100g/min. With these parameters, pseudocontinuous ASL (PCASL) data can be simulated accurately by a convolution of a rectangular arterial input function7 and a single-pass approximation (SPA) impulse residue function8. In the second step, a five-minute experiment was simulated in which the scan time is divided between the ASL acquisition and support-scans to estimate $$$\small{\alpha_i}$$$ and $$$\small{T_{1b,i}}$$$. Three options were considered for the ASL acquisition: single-post-labeling-delay PCASL (single-PLD PCASL) with labeling duration $$$\small{\tau}$$$=1800ms and PLD=1800ms, time-encoded free-lunch PCASL6 (te-FL PCASL) with the duration of block one 1800ms followed by ten blocks of 175ms and a 50ms shortest PLD, and multi-PLD PCASL with $$$\small{\tau}$$$=1800ms and PLDs=200,400,...,2000ms. Number of averages of the ASL acquisition was set based upon the available scan time. Estimates of $$$\small{\alpha_i}$$$ and $$$\small{T_{1b,i}}$$$ were simulated by a random draw from normal distributions $$$\small{\mathcal{N}(\alpha_i,\sigma_\alpha)}$$$ and $$$\small{\mathcal{N}(T_{1b,i},\sigma_{T_{1b}})}$$$, with $$$\small{\sigma_\alpha}$$$ and $$$\small{\sigma_{T_{1b}}}$$$ scaled inversely by the square root of the respective assigned scan time to reflect averaging. Realistic values for the temporal SNR of ASL data, and $$$\small{\sigma_\alpha}$$$ and $$$\small{\sigma_{T_{1b}}}$$$ for certain reported scan times were adopted from literature3,4,9. Thirdly, the single-compartment model was used to quantify CBF1,7. For te-FL and multi-PLD PCASL, $$$\small{\Delta}$$$$$$\small{t}$$$ is estimated alongside the CBF.
This entire protocol was repeated on three levels: 1) varying distributions of the five-minute scan time 2) for each scan time distribution, 10000 physiological conditions ($$$\small{i}$$$=1,...,10000) were simulated, and 3) for each of these, 1000 noise realizations of the PCASL data and $$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$ measurements were generated
($$$\small{j}$$$=1,...,1000).
Splitting scan time between acquisition of ASL data and estimating $$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$ drastically improves the CBF estimation reproducibility in the general population compared to the consensus-statement ASL experiment. Sacrificing ASL scan time for estimating $$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$ increases the noise level in the ASL data, yet it captures the underlying physiological state more accurately. On a population level, the latter clearly outweighs the former (Fig.3). For the parameter distributions in Table 1, combining multi-PLD PCASL with estimating $$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$ proves most reliable (Fig.3). For the assumed $$$\small{\Delta}$$$$$$\small{t}$$$ distribution, a large amount of signal is lost in single-PLD and te-FL PCASL, impeding their performance. The fact that they outperform multi-PLD PCASL for longer $$$\small{\Delta}$$$$$$\small{t}$$$ values (Fig.4) shows the dependence of this simulation on the chosen distribution of $$$\small{\Delta}$$$$$$\small{t}$$$ values and might point to the necessity of including dispersion effects in simulation studies to capture a more realistic spread of arrival times.
The results from any simulation experiment depend on the parameter distributions used to generate data. The benefit of sacrificing ASL scan time for $$$\small{\alpha}$$$ and $$$\small{T_{1b}}$$$ estimation is directly related to the assumed uncertainties for both. Confidence in the conclusions drawn from this experiment can be attributed to the reported distributions being representative for the general population and the use of an accurate data-generation model8. In the future, dispersion will be included to further increase the accuracy of the signal generation process of PCASL.