Abstract
Purpose/Introduction: MRSI suffers from the voxel bleeding effect, where the spectrum in a given voxel exhibits contributions from other spatial regions. Intra-voxel variation of coil element sensitivities, especially at ultra-high fields, may cause significant additional contamination in parallel MRSI [1,2,3,4]. We extend the SENSE technique by introducing spatial overdiscretization and combining it with direct optimization of the Spatial Response Function (SRF) [5,6,7] to reduce residual aliasing and suppress SRF side lobes.
Subjects and Methods: A 20x16 voxel 1H FIDLOVS [8] MRSI was acquired at 7T with a 16channel receive array from a transversal slice of a healthy volunteer’s brain with coil sensitivity information. The reconstruction operator
F for voxel π is calculated from the encoding matrix E by minimizing the cost function Δπ=||(FE – T)π||22 + α(FΨFH)πwhere the regularization parameter α weights SRF optimization (first term) against noise minimization (second
term) and T contains the SRF target functions [5,7]. E and T are newly sampled at multiple positions per voxel (overdiscretization).
Results: SRF side lobes reflecting voxel bleeding into the voxel of interest abound at a simulated R=4-fold acceleration (Fig. 1). For R=9, even residual aliasing peaks are discernible. With a slightly widened Gaussian SRF target (width σ=1.5 in units of the underlying spatial grid) and threefold overdiscretization, very efficient side lobe and aliasing peak suppression are combined with an acceptably small increase of the effective voxel size.
Minimal noise regularization is desired for SRF shape and unfolding but causes numerical ill-conditioning. We find, in our example, α=1 to be a good compromise (Fig. 2).
Spatio-spectral quality improves drastically over standard SENSE (Fig. 3, 4).
Even with the inherent SNR decrease, acceptable spectra are obtained at R=9, which corresponds to an acquisition time of only 4.5 min.
Discussion/Conclusion: We demonstrate up to 9-fold accelerated FIDLOVS MRSI at 7T with intrinsic and efficient suppression of voxel bleeding contributions as well as good maintenance of the nominal spatial resolution in contrast
to conventional k-space apodization methods.
