Though widely used in neurological research (1-5), the clinical utility of myelin imaging based on “gold standard” multi echo spin echo (MESE) T2 relaxometry is currently impeded due to the requirement of high SNR and the need to account for contributions from stimulated pathways. Use of GRASE sequence as an alternative to MESE in context of myelin imaging has been explored before (6). In this context, our goal is to generate better voxelwise match between MWF-maps extracted using both sequences by the use of multivoxel spatial regularization algorithm with stimulated echo correction (7).

Intensities of T2-decay can be written as a linearized form of the EPG model (7): y = A_EPGx +ε, with A_EPG(i,j) = intensity at echo-time TE(i) with discrete T2(j) values and the white Gaussian noise vector ε. The noise-robustness of the reconstruction can be enhanced by imposing spatial smoothness of solutions(4,7):

$$ \widehat{x} = arg min_x {||A_{ex} \overline{x}- \overline{y} ||}^2 + M_T {||\overline{x}||}^2 + \mu_s{||D_s \overline{x}||}^2 $$ where column vectors $$$\overline{x}, \overline{y} $$$ are multi-voxel equivalent of x, y and A_ex is the block diagonal matrix, with A_EPG as its block. M_T is the diagonal matrix with voxelwise temporal regularization µ_T along its diagonal and µ_S is spatial regularization parameter. Matrix D_S is first difference operator and ||D_SX|| penalizes non-smooth solutions.

Experiment: QT2R data was acquired from healthy volunteers using CPMG based non-selective MESE and GRASE sequences (3T Philips-Ingenia) with: axial FOV 230x190 mm, voxel resolution 2 x 2 x 3.5 mm3, receiver bandwidth 355 kHz, 12 slices, TR 2000 ms, 32 echoes, SENSE-factor: inplane = 2 & Slice-encoding = 2; echo spacing 6 ms; Average 2. Additionally, EPI-factor of 3 was used for GRASE sequence. It took ~14 and ~42 minutes to acquire MESE and GRASE data with limited coverage. Average of 2 was essential for the FID correction. The same GRASE sequence (no slice over sampling, #slices ~45) with full brain coverage data could be acquired in 25 minutes.

Processing Algorithm: For nominal flip angles of 90^o, 180^o, the EPG model is only sensitive to the magnitude of FAE. 60 possible candidates for FAEs are considered between 0% to 30% at regular intervals. The first step involves generating a series of L-curves with one for each value of possible FAEs. Among multiple elbows of those L-curves, one closest to the origin is selected as the best compromise between the data fidelity and the prior. The second step involves spatial regularization and since entire data cannot be processed in one go due to excessive memory and CPU demand resulting from sparse nonnegative least square (SNNLS) solver, 10 x 10 data selecting window (DSW) is processed and an overlap between successive DSWs is ensured to avoid tiling effect (3). For any particular DSW, matrix Aex is constructed using the EPG model and voxelwise FAE-values. With known FAE-map, MT and µS (= 3000 x median of µT-map), eqn [1] can be written as expression with single L2-norm (7) and can be solved by sparse nonnegative least square solver (8). A Matlab implementation is also available online (9).

There is a good qualitative match between myelin quantifications using either of sequence; though MWF values from GRASE sequence are significantly suppressed on the left side of front cortex (Fig. 1). The linear regression between both set of MWF-maps is found to be y = 0.82x + 0.0073 (Fig. 2), indicating satisfactory match.

The acquisition of a very large number of k-space points following each excitation makes the underlying relaxation process in T2-prep based method significantly different from gold standard MESE, leading to divergent MWF-values across white-matter (figure 4 vs figure 5 of (10)). A small EPI factor used for GRASE sequence preserves T2W contrast similar to MESE. The study by Prasloski et al. (6), comparing quantifications from both 3D MESE and 3D GRASE, gave similar ROI-averaged MWF-values for most of WM and GM structures, though the congruence between respective voxel wise maps were far from perfect (absolute difference ~0.05-0.15). Our method shows better voxel wise match between both set of MWF-maps due to improved noise robustness and better B1-resolution. Suppressed MWF-values in frontal brain in case of GRASE sequence can be attributed to enhanced field inhomogeneity due to nasal cavity which signficantly alters T2-decay in those parts.The good agreement between MEGE and GRASE based MWF maps suggests that GRASE sequence based MWF-imaging has potential to be clinically feasible.