Synopsis

Keywords: Data Processing, Spectroscopy

The automated voxel placement (AVP) framework has been proposed to optimize accuracy and reproducibility in single-voxel brain proton spectroscopy. AVP is based on the affine transformation of brain coordinates from standard to subject-specific space defined on a 3D T1-weighted structural scan. Here we evaluate the robustness of the AVP approach by evaluating the displacement of voxel center coordinates across 3D T1w MPRAGE sequence variants, including compressed sensing (CS) accelerated protocols. We show that AVP gave small but significantly higher voxel displacements for MP2RAGE and CS-MP2RAGE. There were no significant differences between multi-echo MPRAGE (6 min) and CS-MPRAGE (1 min).

Introduction

Single-voxel proton magnetic resonance spectroscopy (MRS) results are crucially dependent on the accurate and reproducible positioning of the voxel that includes the target tissue of interest for metabolic evaluation [1]. Previous studies have shown that precision and reliability of brain MRS can be highly increased by using an automatic voxel placement (AVP) strategy [2]. This approach consists of acquiring the subject-specific T1-weighted (T1w) anatomical image, exporting it from the MR scanner to a computer that runs the AVP tool, setting the AVP to a predefined voxel coordinate in MNI space, computing the corresponding target voxel coordinates in subject space, and finally feeding these coordinates to the MR scanner console for the MRS acquisition. The AVP strategy has been demonstrated at 3T using a standard 3D T1w MPRAGE as input [2]. However, it remains unknown how robust the AVP is when different input T1w images are used, which is especially important with the availability of highly accelerated versions. In this study, for three different target MRS voxels, we study the variability of the AVP subject-specific coordinates as a function of MPRAGE sequence variants, both for standard (MPRAGE, multi-echo MPRAGE, MP2RAGE) and compressed sensing (CS) accelerated versions (CS-MP2RAGE and CS-MPRAGE).

Methods

MRI acquisition
Six healthy adults (age=25.7±2 years) underwent 3T MRI (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) with a 64-channel head-neck coil. T1w images were acquired with three “standard” and three CS-accelerated 3D MPRAGE sequence variants using a research application sequence, all with 1mm isotropic voxels, same spatial coverage, active prescan normalize and no image filters: a) multi-echo MPRAGE (TR/TI=2530/1100, TE1-4=1.69/3.55/5.41/7.27ms, α=7°, GRAPPA=2, TA=6:03min) [3]; b) MP2RAGE (TR/TE=5000/2.98ms, α=4°/5°, GRAPPA=3, TA=8:52min) [4]; c) MPRAGE (TR/TE/TI=2310/3.48/1200ms, α=12°, GRAPPA=2, TA=5:32 min); d) CS-MP2RAGE (TR/TE=5000/2.88ms, α=4°/5°, undersampling factor 4.6, regularization factors 0.0006/0.0004, TA=3:40min) [5]; e) CS-MP2RAGE as in d), with online removal of background noise; f) CS-MPRAGE (TR/TE/TI=2300/2.88/900ms, α=9°, regularization=0.0006, undersampling factor 3.6, TA=2:04min) and g) CS-MPRAGE as in f) with undersampling factor 6.6, TA=1:14min [5]. More details on the CS reconstructions can be found in reference [5].

Single voxel placement for MRS
T1w images of each subject were given as input to an optimized version of the AVP tool [1] to enable faster coregistration of three hypothetical MRS voxels (templates) relevant for various diseases: right-hemispheric dorsolateral prefrontal cortex (rDLPFC, MNI=[36, 44, 20] [6], right-hemispheric basal ganglia (rBGAN, MNI=[20, 10, -2] [7]) and posterior cingulate cortex (PCC, MNI=[0, -53, 26] [8]). For each T1w sequence and voxel, the AVP algorithm translates standard MNI coordinates into subject space location.

Statistical analysis
The goal of the analysis was to quantify, for each MRS voxel location, the within-subject differences in the voxel position estimation as function of the various input T1w sequences. Relative displacements were measured computing the Euclidean distance between the affine transformation of voxel coordinates returned by AVP and the template voxel coordinates defined in standard space. The statistical analysis was performed as a two-way repeated measures ANOVA.

Results

Figure 1 shows, for a representative subject, the similarity of image contrast properties for the various sequences and acquisition times. As multi-echo MPRAGE we used the root mean squared calculated across echoes. As MP2RAGE we used the uniform image, which includes background noise, where the CS-MP2RAGE version provided a uniform and denoised image. Figure 2 shows the anatomical localization of the three template voxels in MNI space. Figure 3 shows the Euclidean distance differences (mm) for the three voxels across T1w sequences. Significant effect of sequence was found for the standard and CS-accelerated MPRAGE variants, regardless of the particular voxel location (p<0.01). Non significant effects were reported both for voxel location and sequence x voxel interaction (p=0.62; F(12, 105)=0.31, p=0.98, respectively).

Discussion and conclusions

Our study shows that at the single-subject level, the AVP results depend on the details of the anatomical T1w sequence. With respect to the choice of the MPRAGE sequence, which was proposed in the original study [2], we found that MP2RAGE sequence variants, both standard and with CS acceleration, provide significantly different target voxel coordinates. The background noise from uniform MP2RAGE image, if not removed, potentially introduces a major source of error in estimating the affine transformations between spaces. A further step of denoising may diminish these differences, as suggested by the results of the uniform denoised version of the CS-MP2RAGE. There were no significant differences of voxel localization when using a fast CS-MPRAGE implementation of about 1 min acquisition time. Even though further acceleration may be possible combining CS-acceleration and reduced matrix size for AVP, higher resolution enables different applications - e.g. morphometry - that would otherwise require a separate acquisition. These results strongly support the use of fast CS-MPRAGE strategies for voxel positioning in brain MR spectroscopy.