Arterial Spin Labelling shows great promise for perfusion measurements; however, despite numerous volunteer reproducibility studies, comparisons have not been made using a phantom to establish differences due to the acquisition hardware and pulse sequences. We present data from a multi-site study using a perfusion phantom, targeting 3T MRI systems from a single vendor running the same software version.
The measurement of Cerebral Blood Flow (CBF) using Arterial Spin Labelling (ASL) has recently seen a renewed interest following the publication of the Position Paper providing a set of very clear guidelines on how to set up an ASL experiment1. In particular, it has been shown to be valid as a biomarker of neurological disease onset2 and response to therapy3. Indeed, based on numerous reproducibility studies4, the coefficient of variation of CBF has been well established, enabling its use as a biomarker in cross-sectional studies5. However, while the sources of potential physiological confounds are well established6, it has so far not been possible to compare all ASL implementations depending uniquely on potential hardware differences.
In this study, we set out to assess the effective reproducibility of CBF estimates by ASL using a recently developed Perfusion phantom7 at 11 different sites with a range of scanner manufacturers (total 17 systems) working on various software levels. We present here the preliminary data from the first 5 scanners from 3 sites working on Philips 3T MRI scanners with software release R5.3.
A perfusion phantom was transported by car to 3T MRI imaging centres in the Netherlands over the course of a week and scanned on 5 Philips 3T MRI systems running software release R5.3 (detailed in Figure 1.a). On each system, ASL measurements were made using the product ASL sequence, comprising of pCASL labelling with a 4-shot 2D-EPI segmented acquisition. Measurement parameters are detailed in Figure 1.b. Data were acquired with ‘normalisation’ turned on; the first dynamic a long-TR without background suppression or labelling pulses for an M0 image, followed by 3 dynamics of control-label pairs. Measurements were made at two flow rates; 200ml/min and 350ml/min, and computer software monitored and recorded the phantom flow rates during scanning. At each site, care was taken to ensure the phantom was reproducibly placed on the patient couch (see Figure 2), and the FOV was centred at a landmark at the centre of the porous material and rotated into alignment with the phantom.
Analysis was performed in Matlab R2016b (The Mathworks, Natick, MA, USA). Dicom images were converted to NIFTI dicm2nii8, and CBF maps calculated using the single subtraction equation for pCASL1:
$$\text{CBF}={6000 \cdot \lambda \cdot (\text{SI}_{\text{control}}-\text{SI}_{\text{label}} )\cdot e^{\text{PLD} \over T_{1b}} \over {2 \cdot \alpha \cdot T_{1b} \cdot \text{SI}_{\text{PD}} \cdot (1-e^{-{\tau \over T_{1b}}} ) }}$$
where α=0.85, λ=0.32, T1b=1800ms, τ=1800ms, PLD=1800ms. The M0 image from dataset was registered to a structural atlas image of the phantom, from which an ROI mask of the entire porous material was generated (see Figure 3). The mean CBF and standard deviation within this ROI were then calculated.
Anita Harteveld (UMCU, Utrecht, Netherlands) for time and assistance scanning.
Koen Baas (AMC, Amsterdam, Netherlands) for time and assistance scanning.
Pieter Vandemaele (Ghent University, Belgium) for advice on standardised phantom positioning.
Figure 1
Summary of MRI Systems (a). ASL Sequence Parameters (b).
Figure 2
The perfusion phantom consists of an MRI compatible pump that delivers a liquid at a controlled flow rate to a perfusion chamber. The liquid is distributed in a ‘vascular’ network to a porous material cylinder that simulates the capillary bed of diameter 116mm and height 28.5mm. To minimise any variance in measurements due to placement of the phantom, on each MRI system the laser marker was set to the same reference point on the phantom, the phantom was aligned with the laser in the head-foot direction, and the phantom was levelled using foam pads.
Figure 3
The creation of the image masks is automatic, using the registration of an 'atlas' image of the phantom to the data to be analysed. First, a Hough transform provides candidates for known circular features to be detected which are then exhaustively matched w.r.t. their size and relative location to one another (top). If consecutive slices detect features in a similar location, this location is used as an initialisation to a multi-resolution, multi-transformation (including B-spline) registration to the phantom 'atlas' image (bottom-left), which provides the porous material mask (in red, bottom-right).
Figure 4
Calculated CBF maps of slice 5 at each flow rate in each data set from all systems (a). Images have been masked with the porous material mask to remove background signal.
Histograms of the CBF value distributions within the mask at a flow rate of 200ml/min (b) and 350ml/min (c) for MRI System 5. The distributions have two components: a Gaussian distribution centred around zero, corresponding to noise in voxels where there is no perfusion signal; and a broader, positive non-zero centred Gaussian distribution of values from voxels that do have a perfusion signal.
Figure 5
Mean CBF (a) and standard deviation of the CBF values (b) within the porous material masks. Trends in variation between the MRI systems visually correspond at flow rates, with variation at the higher flow rate more pronounced. In all cases, the mean CBF at 350ml/min is more than double that measured at the 200ml/min, despite the flow rate ratio being 1.75. This is because at 200ml/min not all of the labelled bolus has reached the porous material at PLD=1800ms.