To assess the possibility and the potential impact of partial volume correction (PVC) in dynamic susceptibility contrast MRI (DSC-MRI).

Tissue partial volume effects (PVEs) constitute an important source of bias in measurements of absolute brain perfusion. In particular, low spatial resolution tends to result in under- and overestimation of cerebral blood flow (CBF) estimates in gray matter (GM) and white matter (WM), respectively. Thus, disentangling the effects of tissue volume and perfusion is important, especially in diseases in which atrophy and perfusion alterations are common.

The impact of tissue PVEs in perfusion measurements with DSC-MRI has not previously been thoroughly investigated. In this work, two PVC methods were applied to DSC-MRI data, and, for comparison, to the corresponding arterial spin labeling (ASL) maps.

The DSC-MRI signal was assumed to be given by $$$S(t)=S_0e^{-t_E\cdot r_2^*\cdot c(t)}$$$, where $$$S_0$$$ is the baseline signal, $$$t_E$$$ is the echo time, $$$r_2^*$$$ is the transverse relaxivity of the contrast agent, and $$$c(t)$$$ is the tracer concentration. The signal can also be described as the sum of signal contributions from different components according to $$$S(t)=S_0\sum_ip_{v,i}e^{-t_E\cdot r_2^*\cdot c_i(t)}$$$, where $$$p_{v,i}$$$ is the partial volume of component $$$i$$$, $$$c_i(t)$$$ is the tracer concentration in component $$$i$$$, and $$$i=$$$ {GM,WM,CSF}. By combining the two equations above, acknowledging that cerebrospinal fluid (CSF) is non-perfused, and that partial volumes sum up to 1, Taylor expansion to the first order yields: $$$c(t)\approx p_{v,GM}\cdot c_{GM}(t)+p_{v,WM}\cdot c_{WM}(t)$$$. Preliminary simulations with plausible assumptions about $$$t_E$$$, $$$r_2^*$$$ and $$$c(t)$$$ indicate that a first-order Taylor expansion is not unreasonable (data not shown). Still, when second- and higher-order terms are non-negligible, this approximation fails. By further assuming similar temporal tracer-kinetic patterns in GM and WM we can assume that the measured perfusion $$$f\approx p_{v,GM}\cdot f_{GM}+p_{v,WM}\cdot f_{WM}$$$, and this can be used for post-quantification PVC (this relationship was also used for ASL). Here, $$$f_{GM}$$$ denotes GM perfusion and $$$f_{WM}$$$ denotes WM perfusion.

Twenty healthy volunteers underwent prebolus DSC-MRI¹, pseudo-continuous ASL, and quantitative T1-based (IR Look-Locker) partial volume (PV) estimation using the fractional signal approach². DSC-MRI and PV data had a voxel size of 1.72×1.72×5 mm³ and ASL data 2.3×2.3×5 mm³. (See Ref. 3 for details on protocols.)

PVC was performed using (i) linear regression⁴ (LR) and (ii) modified least trimmed squares⁵ (mLTS), with circular kernels (diameter 5 pixels), yielding partial-volume-free GM and WM perfusion maps. Uncorrected and PV-corrected CBF maps were analyzed, where GM and WM ROIs were extracted from the PV estimates ($$$p_{v,i}=$$$ 10–20%,...,90–100%). Visual inspection and calculation of the relative range (rR) of CBF values over the range of partial volumes (for no PVC, mLTS and LR) allowed for assessment of the degree of volume dependence of estimated CBF. Spuriously high values from large vessels were removed by masking out voxels with CBV>10 ml/100g.

Example PV and PVC maps are shown for DSC-MRI (Fig. 1) and ASL (Fig. 2). (Note that PVC CBF maps are masked at PV>10%.) LR yielded more apparent smoothing in the PVC maps than mLTS, and WM perfusion maps were less noisy in DSC-MRI compared to ASL. Figure 3 shows results of the ROI analysis. PVC decreased the volume dependence of CBF estimates for DSC-MRI as well as ASL, and for both tissue types. Notably, for DSC-MRI, the LR method decreased the rR from 44.4% (for uncorrected data) to 4.2% in GM and from 55.3% to 14.2% in WM. As expected, the PVEs were larger in the ASL data, with a rR of 60-80% for uncorrected data.

In this work, a linear approximation of the DSC-MRI signal was applied to allow for post-quantification PVC of perfusion estimates. PVEs were clearly identified in DSC-MRI data, and PVC decreased the volume dependence of CBF estimates. However, the first-order Taylor approximation needs to be further evaluated, and it might be preferable to expand the concept by, for example, nonlinear regression or, alternatively, by modelling the tissue signal as a combination of nonlinear tracer responses from GM and WM (i.e., modelling the PVE within the quantification process).

The mLTS algorithm performed less convincing than LR with regard to removing volume dependence, although it preserved more detail in the corrected maps due to the use of adaptive matrix trimming. However, the low volume dependence in LR could, at least partially, be an effect of the blurring. PVC can also be applied to CBV estimates, and therefore, in principle, allow for separate GM and WM MTT estimates. In conclusion, PVEs significantly affected CBF estimates in DSC-MRI, and application of PVC should be considered when absolute values are of interest.