Identifying systematic errors in quantitative dynamic-susceptibility contrast perfusion imaging by high-resolution multi-echo parallel EPI

Abstract

Several obstacles usually confound a straightforward perfusion analysis using dynamic-susceptibility contrast-based magnetic resonance imaging (DSC-MRI). In this work, it became possible to eliminate some of these sources of error by combining a multiple gradient-echo technique with parallel imaging (PI): first, the large dynamic range of tracer concentrations could be covered satisfactorily with multiple echo times (TE) which would otherwise result in overestimation of image magnitude in the presence of noise. Second, any bias from T(1) relaxation could be avoided by fitting to the signal magnitude of multiple TEs. Finally, with PI, a good tradeoff can be achieved between number of echoes, brain coverage, temporal resolution and spatial resolution. The latter reduces partial voluming, which could distort calculation of the arterial input function. Having ruled out these sources of error, a 4-fold overestimation of cerebral blood volume and flow remained, which was most likely due to the completely different relaxation mechanisms that are effective in arterial voxels compared with tissue. Hence, the uniform tissue-independent linear dependency of relaxation rate upon tracer concentration, which is usually assumed, must be questioned. Therefore, DSC-MRI requires knowledge of the exact dependency of transverse relaxation rate upon tracer concentration in order to calculate truly quantitative perfusion maps.

Figure 1

(a) AIF calculated from different echoes. The AIF averaged over the time course of ~30 automatically selected voxels. The FWHM of in-plane smoothing was 1.0. Please note that, owing to the differing input to the automatic selection of AIF voxels, different clusters of voxels were used for AIF calculation. The integral over each time course is given by Σ in arbitrary units. (b) AIF calculated from weighted fit with different error-weighting exponents w (see text). Other parameters were the same as in (a). (c) AIF calculated from the maximum of the AIF from the first echo and the weighted fit with w = 1. Please note that the actual values of the combined AIF are slightly different from those of the constituent curves because there is a minor variation of the locations of automatically selected voxels. (d) Influence of the number of AIF voxels on the shape of the AIF. Note that the actual number of voxels is less than N (which is specified in brackets) because of the next-neighbor analysis as described in the text. The integral over each time course is given by Σ in arbitrary units. (e) The effect of in-plane smoothing with a Gaussian kernel of different width on the shape of the AIF. Data is from 33 automatically selected voxels

Figure 2

(a) $\mathrm{\Delta }{R}_2^{\ast }$ time course within GM and WM regions for the last echo with TE ms = 57.1 and for a weighted fit with w = 1. (b) $\mathrm{\Delta }{R}_2^{\ast }$ time course within GM regions for different TEs

Figure 3

CBV maps, generated from weighted fit (left column), and from echo with TE = 57.1 ms (right column) of every other slice. Gray-scale coded values are in ml/100 g

Figure 4

CBF maps, generated from weighted fit (left column), and from echo with TE = 57.1 ms (right column) of the same slices as in Fig. 3. Deconvolution was performed using an AIF calculated by the maximum method as described in the text. Gray-scale coded values are in ml/100 g/min

Figure 5

The gain (or loss) of SNR when using PERMEATE in comparison to single-echo parallel imaging with different TEs as a function of tracer concentration