Comparison of two methods for atrophy-correction in perfusion imaging: Partial-volume correction versus gray matter volume covariate

Purpose / Introduction
Partial volume (PV) effects, caused by the mixing of different tissue signals within a voxel, are a well-recognized confounder in arterial spin labeling (ASL) imaging1,2, and are especially relevant in populations with atrophy3,4. Although several PV-correction algorithms exist5,6,7, many investigators continue to account for PV by using gray matter volume (GM-vol) as a covariate in the analysis of PV-uncorrected cerebral blood flow (CBF) images. To gain insight into this issue, we compared the performance of PV-correction5 vs GM-vol covariate using data based on acquired images that were simulated to reflect decreases in GM-vol and CBF.
Subjects and Methods
A total of 88 3D T1w images with 1x1x1mm3 resolution were acquired on twenty-two healthy volunteers8 (22.6±2.1 years, 9 men) who were scanned twice on two 3T scanners (GE Discovery and Philips Intera). T1w images were segmented, and CBF images were simulated to reflect GM-vol and CBF decrease with age, which was randomly and uniformly assigned to each subject (range 40-80 years). Sixteen different combinations of linear GM-volume and CBF decrease of 0, 0.25, 0.5 and 1% per year were simulated. The GM-vol decrease was simulated as cortical thinning using the high-resolution PV maps (Figure 1). The PV maps were then downsampled to the 3x3x7mm3 resolution. CBF images were simulated assuming baseline GM-CBF of 80±4 mL/min/100g and age-independent GM/WM-CBF ratio of 3. Additionally, pixelwise Gaussian noise (s.d. ±4 mL/min/100g) was added to the simulated CBF images11 (Figure 2). The mean CBF in voxels with GM content above 70% was analyzed with the two methods using a multivariate linear regression with age as fixed effect. The estimated slope of CBF change with age was compared with the simulated (ground truth) value.
Results
The PV-correction outperformed the GM-covariate method in all cases, and the PV-correction relative error was under 0.5% (Table 1). The error of the GM-covariate method increased with increasing atrophy and reached a maximum value of 10%.
Discussion / Conclusion
In the presence of atrophy, PV correction allowed for detection of CBF changes relatively independent of atrophy and with higher accuracy than the GM-volume covariate method. The reason is probably that the underestimation of PV-uncorrected CBF in the presence of atrophy does not perfectly correlate with the GM-vol decrease as it depends on additional factors such as cortical thickness. Further studies should evaluate the effect of global and local segmentation errors, and focal atrophy on the efficacy of both methods.