This study aimed to optimize a processing pipeline for QUASAR ASL. We have focused on 3 aspects: the assignment of AIFs in voxels; dealing with voxels having excessive values potentially due to an EPI artifact; and minimization of partial volume effects. Simulations showed that GM CBF values closer to the ground truth are obtained by using AIFs in a distance double than the nearest-neighbor AIF to every voxel. In terms of an EPI artifact present in the analysis, we have shown that identification and exclusion of influenced voxels with a developed algorithm, results in values closer to the expected ones.
To develop and evaluate using a combination of real and simulated data a processing approach for QUASAR ASL targeted to:
a) reduce the effect of low SNR and presence of chemical shift artifacts and
b) reduce the effects of the partial volume phenomenon.
The crushed (ΔMcr) and non-crushed (ΔMncr) ASL signals were simulated based on the following equations:
ΔΜcr=2*CBF*M0a*AIF(t)⊗ R(t)
ΔΜncr=ΔΜcr*(PVGM+PVWM)+2*M0a*AIF(t)*aBV
Where the AIF(t) and the residue function were simulated as described in [2] and CBFGM=60ml/100g/min, CBFWM=20ml/100g/min. M0a, aBV and partial volume (PV) maps for grey matter(GM) and white matter (WM) were extracted from a subject in order for the data to be more realistic. A random number in the interval [0.2-0.4s] was added to the crushed signal, to reflect the delay from the arterial input to the tissue. AIFs were assigned to every voxel using a distance-based weighted average of local AIFs. The closest local AIF to every voxel was determined and its distance was used as a reference (ρ). We evaluated the impact of using a weighted average of local AIFs by assigning weights in AIFs situated within a n*ρ distance from every voxel in various n (Fig.1). For each combination 5 repetitions were run. The metrics used to evaluate the performance for the chosen values were the mean errors in: GM CBF, WM CBF, voxel-wise total CBF, voxel-wise GM CBF for GM>50% and GM>80%.
In order to account for an EPI artifact present in the analysis, an algorithm detecting edges based on a user-defined threshold and excluding voxels influenced by the artifact in the ΔΜncr-ΔΜcr image was developed. A bright “ring” was added in the crushed and non-crushed experiments slightly misaligned to simulate the artifact (Fig.2). 7 threshold combinations (slightly higher thresholds were used for the lower slices) were evaluated in our simulations, n=2 was used. The same metrics as before were used to assess the performance for every value. For PV correction we used the method developed by Asllani et al [3]. PV maps were generated using spatial fuzzy c-means (SCFM) clustering [4] of a high resolution MPRAGE image of a subject. The PV maps were down-sampled and registered to the QUASAR images.
1. Petersen, E.T., T. Lim, and X. Golay, Model-free arterial spin labeling quantification approach for perfusion MRI. Magnetic Resonance in Medicine, 2006. 55(2): p. 219-232.
2. Chappell, M.A., et al., Comparing Model-Based and Model-Free Analysis Methods for QUASAR Arterial Spin Labeling Perfusion Quantification. Magnetic Resonance in Medicine, 2013. 69(5): p. 1466-1475.
3. Asllani, I., A. Borogovac, and T.R. Brown, Regression algorithm correcting for partial volume effects in arterial spin labeling MRI. Magn Reson Med, 2008. 60(6): p. 1362-71.
4. Chuang, K.S., et al., Fuzzy c-means clustering with spatial information for image segmentation. Comput Med Imaging Graph, 2006. 30(1): p. 9-15.