Henrik Marschner^{1}, Laurentius Huber^{2}, André Pampel^{1}, and Harald E. Möller^{1}

In this study we investigate possible benefits of an application of ‘AWESOME’ de-noising on fMRI. The application in a high-SNR finger tapping experiment showed a reduction of the already low thermal noise contribution and therefore improvement of tSNR and reduction of false positives; no adverse effects in the form of smoothing or suppression of ‘true’ activation was observed. A second investigation of the scalability of tSNR improvement on a resting state experiment with variable slice thickness / SNR showed that thermal noise can be reliably reduced and the tSNR proportionally improved without visible reduction of detail sharpness / resolution.

Acquisition: Two data sets acquired at 7T
(MAGNETOM 7T,
Siemens Healthcare, Germany) with a 32-channel head coil (Nova
Medical, Wilmington, MA, USA) were analyzed.
(i)
resting-state BOLD-based fMRI data consisting of
3D GRE-EPI scans in a healthy subject employing the following
parameters: 1.5x1.5mm^{2} in-plane resolution, varying slice
thickness of 0.25/0.5/1/2 mm, 12 slices, TR=3s, TE=22ms,
in-plane GRAPPA 2, 50 repetitions.^{3}
(ii) task-based fMRI data: CBV and BOLD contrasts were recorded using
a 3D slice-saturation
slab-inversion VASO (SS-SI-VASO) sequence: 0.75x0.75mm^{2}
in-plane resolution, 1.8mm slice thickness, 8 slices, TR=3s, TE=22ms,
in-plane GRAPPA 2.^{3}
The paradigm was a unilateral finger-tapping task (block design;
12-min total acquisition time).

Preprocessing: Data were corrected for motion. Subsequently,
the background phase in the data was corrected using an algorithm
based on total variation de-noising.^{7} The AWESOME
algorithm requires a normalized noise-level map that was calculated
from the highest frequency band of the 1D wavelet transformation of
the time line for each voxel.

AWESOME de-noising: AWESOME^{1,2} operates in the
wavelet space of complex MR images. In the wavelet space, signals
(i.e. brain structures) and noise are well separated in multiple
frequency bands, with thermal noise mostly represented in high
frequency bands. From the complex noise distribution, a filter is
derived to separate noisy and mostly meaningful data in the wavelet
space. A mean wavelet data set is calculated over the image series.
From this high SNR mean data set, the original signal contributions
are estimated based on phase-weighted rescaling using the fraction of
total signal energy per voxel over the series and each single voxel
signal in the wavelet space. The thereby estimated signals replace
noisy signals in the original wavelet data. The inverse wavelet
transformation of the new data results in a de-noised MR image
series.

The calculation of the mean wavelet data set was performed
differently in both time series, because data containing *alternating*
contrasts (i.e. CBV-weighted VASO and BOLD contrasts acquired with
the SS-SI-VASO sequence) can suffer from cancellation of meaningful
data in the complex mean. Thus, (i) the mean of rsMRI data is
calculated from the complex wavelet coefficients; (ii) the mean of
the task-based-fMRI data is calculated from the mean imaginary part
and the mean absolute real part, which is then bias-corrected for the
background noise.

FSL was used for statistical fMRI analysis and estimation of smoothness in the VASO data set.

rs-fMRI: In Figure 1, the result of AWESOME-based de-noising is demonstrated for the acquisition with the minimal slice thickness (0.25 mm). De-noising of rsMRI data resulted in an SNR improvement per volume by up to four times as compared to the original SNR. As a consequence, the time-series SNR (tSNR) is almost doubled (Figure 2). Most notably, this is obtained without visible loss of image detail. The results for all rs-fMRI data are summarized in Figure 2.

Task-based fMRI: Although the voxel size was rather small, the SNR was already relatively high. Therefore, the effect of de-noising is not readily visible in Figure 3. Yet, fMRI analysis showed a reduction of false-positives in the statistical maps of the VASO signal changes (Figure 4). On the other hand, the areas of activation appear unchanged (cf. Figure 4). This can be seen as well in the VASO signal time course (Figure 5). Also, the corresponding cortical proﬁles of VASO signal change are not significantly altered.

1. Marschner H, Pampel A, Möller HE. Adaptive Averaging of Non-Identical Image Series in the Wavelet Space. Proc Intl Soc Mag Reson Med 2015;23:3721.

2. Marschner H, Eichner C, Anwander A, Pampel A, Möller HE. De-noising of diffusion-weighted MRI data by averaging of inconsistent input data in wavelet space. Proc Intl Soc Mag Reson Med 2016;24:2071.

3. Huber L, Ivanov D, Marrett S, Panwar P, Uludag K, Bandettini PA, Poser BA. Blood volume fMRI with 3D-EPI-VASO: any benefits over SMS-VASO? Proc Intl Soc Mag Reson Med 2016;24:0944.

4. Huber L, Ivanov D, Krieger SN, Streicher MN, Mildner T, Poser BA, Moller HE, Turner R. Slab-selective, BOLD-corrected VASO at 7 Tesla provides measures of cerebral blood volume reactivity with high signal-to-noise ratio. Magn Reson Med 2014;72(1):137-148.

5. Huber L, Goense J, Kennerley AJ, Trampel R, Guidi M, Reimer E, Ivanov D, Neef N, Gauthier CJ, Turner R, Moller HE. Cortical lamina-dependent blood volume changes in human brain at 7 T. Neuroimage 2015;107:23-33.

6. Krieger SN, Huber L, Poser BA, Turner R, Egan GF. Simultaneous acquisition of cerebral blood volume-, blood flow-, and blood oxygenation-weighted MRI signals at ultra-high magnetic field. Magn Reson Med 2015;74(2):513-517.

7. Eichner C, Cauley SF, Cohen-Adad J, Möller HE, Turner R, Setsompop K, Wald LL. Real diffusion-weighted MRI enabling true signal averaging and increased diffusion contrast. NeuroImage 2015;122:373-384.

Figure 1: The central time step/central slice for 0.25 mm slice
thickness of the original, de-noised and difference rs-fMRI data. The
removed noise is mostly free of structure, except the pattern of
non-uniform noise amplitude over the volume.

Figure 2: tSNR ratio as obtained after de-noising with identical
pre-processing per slice thickness. tSNR ratio is the ratio of tSNR
after de-noising to the original. A subtle checkerboard-like pattern
in the tSNR maps appears after de-noising due to limitations in the
precision of the estimation of original noise-free signals in the
wavelet space.

Figure 3: Magnitude images of both VASO and BOLD contrast show
already little contribution of thermal noise in original images.
Removal of this contribution (difference data) by de-noising enabled
subtle improvements in quantification as shown in Figure 4.

Figure 4: Activation maps of VASO ((S_{act}-S_{rest})/S_{rest}),
z-score and tSNR maps of a finger-tapping experiment with alternating
acquisition of VASO and BOLD contrasts; Besides the improved tSNR,
false positive activations are reduced while keeping the activation
pattern.

Figure 5: (Left) Trail-averaged, BOLD-corrected VASO time course
(S_{act}/S_{rest}) of a ROI in the left primary motor
cortex during right-hand finger tapping. Stimulus onset was at 0s and
stimulus offset was at 30s. (Right) Tapping-induced fMRI signal changes
as a function of cortical depth. The time course was not altered by
de-noising nor were the activation profiles across cortical layers
broadened or distorted.