Single Voxel Proton Spectroscopy for Neurofeedback at 7 Tesla

Abstract

Echo-planar imaging (EPI) in fMRI is regularly used to reveal BOLD activation in presubscribed regions of interest (ROI). The response is mediated by relative changes in T2* which appear as changes in the image pixel intensities. We have proposed an application of functional single-voxel proton spectroscopy (fSVPS) for real-time studies at ultra-high MR field which can be comparable to the EPI BOLD fMRI technique. A spin-echo SVPS protocol without water suppression was acquired with 310 repetitions on a 7T Siemens MR scanner (TE/TR = 20/1000 ms, flip angle α = 90°, voxel size 10 × 10 × 10 mm³). Transmitter reference voltage was optimized for the voxel location. Spectral processing of the water signal free induction decay (FID) using log-linear regression was used to estimate the T2* change between rest and activation of a functional task. The FID spectrum was filtered with a Gaussian window around the water peak, and log-linear regression was optimized for the particular ROI by adoption of the linearization length. The spectroscopic voxel was positioned on an ROI defined from a real-time fMRI EPI BOLD localizer. Additional online signal processing algorithms performed signal drift removal (exponential moving average), despiking and low-pass filtering (modified Kalman filter) and, finally, the dynamic feedback signal normalization. Two functional tasks were used to estimate the sensitivity of the SVPS method compared to BOLD signal changes, namely the primary motor cortex (PMC, left hand finger tapping) and visual cortex (VC, blinking checkerboard). Four healthy volunteers performed these tasks and an additional session using real-time signal feedback modulating their activation level of the PMC. Results show that single voxel spectroscopy is able to provide a good and reliable estimation of the BOLD signal changes. Small data size and FID signal processing instead of processing entire brain volumes as well as more information revealed from the acquired total water spectrum, i.e., direct estimation of the T2* values and B0 changes, make SVPS proton spectroscopy suitable and advantageous for real-time neurofeedback studies. Particular challenges of ultra-high field spectroscopy due to the non-linearity in the spectral information, e.g., poor main magnetic field homogeneity and the absence of motion correction for the SVPS sequence may lead to the special artifacts in the control signal which still need to be addressed. The contrast to noise ratio (CNR), experimental statistic (t-values) and percent signal change were used as quality parameters to estimate the method performance. The potential and challenges of the spectroscopic approach for fMRI studies needs to be further investigated.

Keywords: imaging; neurofeedback; signal processing; spectroscopy.

Figure 1

Spectral processing of the single voxel spectroscopic data. (a) Raw (blue) and processed (red) magnitude free induction decay functions (FIDs); (b) Raw (blue) and processed (red) magnitude spectra. Note that graphics and spectra were normalized to their maximum values for comparison purposes and the panels are scaled differently.

Figure 2

Linear regression of FID data. The linear regression fit (red) is exemplarily shown for a VC fSVPS single voxel ln(FID) data (blue) and its optimal linearization length (red dot; l_optim = 0.078 s, t = 19.12, p < 0.001).

Figure 3

Functional single-voxel proton spectroscopy (fSVPS)-estimated T2* time series. (a) The primary motor cortex (PMC); (b) PMC in real time (PMC NF); and (c) Visual cortex (VC) functional single-voxel proton spectroscopy (fSVPS) time series (processed) are displayed for four control subjects. Acquired time series show significant variability between subjects and regions of interest (ROIs) and overall higher than expected T2* values.

Figure 4

Time-locked T2* changes. Group event-related averages (blue) and their SD across subjects (red bars) for (a) PMC; (b) PMC NF; and (c) VC fSVPS time series are shown. Note that PMC NF time series showed much lower event-related standard deviations but also lower average signal change (b; see also Figure 5c).

Figure 5

fSVPS quality measures. Average (a) t-value; (b) contrast to noise ratio(CNR); and (c) percent signal change (blue bars ± SD in read) show similar patterns across the PMC, PMC NF, and VC conditions. The SVPS approach provides robust statistics based on its high CNR and percent signal change. Note that the panels are scaled differently.

Figure 6

Optimal linear regression length. Experimental statistic (t-value) depends on the applied linear regression length and reveals an optimum duration, l_optim, here shown in a typical example for (a) the PMC (blue dot; l_optim = 200 ms, t = 17.0, p < 0.001) and (b) the VC condition (blue dot; l_optim = 78 ms, t = 19.12, p < 0.001). Note that panels are scaled differently.