Artifact Reduction by Target-Driven Reconstruction

Introduction Magnetic Resonance Spectroscopic Imaging (MRSI) is a powerful technique to gain insight into physiological processes in the human body. Since 2D MRSI examinations typically last excessively long, the use of acceleration techniques is of great importance. Here, we use a SENSE approach [1, 2, 3]. The Spatial Response Function (SRF) indicates which spatial areas contribute to the final voxel content. Ideally, it vanishes outside the voxel of interest (VOI). However, since the SRF is a superposition of the encoding functions used, whose number is finite, it typically exhibits noticeable side lobes. In parallel MRSI, where the number of k-space sampling points is extremely low, this detrimental effect is particularly pronounced. As a result, voxel spectra often contain unwanted contributions by signal originating elsewhere, e.g. subcranial fat signal in brain matter voxels. In this work, we demonstrate a novel reconstruction approach that allows to address this issue by minimizing a cost function that contains the deviation from a predefined SRF target. Theory and Methods Algorithm: Reconstruction was performed by applying a reconstruction matrix F to raw k-space data. In our proposed method [4], F is calculated as the minimum of the cost function Δπ=(FΨF)π,π + ||(FE – T)π||2 for each voxel π. Here, E is the encoding matrix and Ψ the noise covariance matrix; H indicates the Hermitean adjoint. The first term optimizes SNR, whereas the second term minimizes the deviation of the resulting SRF FE from an initially chosen target T. Target functions were centered on the VOI and are displayed in Fig. 1. Note that T=Boxcar mimics standard SENSE reconstruction. T=Gaussian is expected to be a more realistic approximation to the actual SRF, while T=|Sinc| was chosen for comparison. No spatial apodization or spectral filtering was applied. Data acquisition: Transversal slice SELOVS MRSI [5] (FOV: 240 mm x 240 mm, voxel size 10 mm x 10 mm x 12 mm) of a volunteer brain (see Fig. 2) were acquired on a 3T MR system (Philips Medical Systems, Best, The Netherlands) with an 8-channel head coil along with the coil sensitivity information. Outer volume and VAPOR water suppression was employed, and twofold cartesian k-space undersampling was used.