Lowering the thermal noise barrier in functional brain mapping with magnetic resonance imaging - PubMed (original) (raw)

Lowering the thermal noise barrier in functional brain mapping with magnetic resonance imaging

Luca Vizioli et al. Nat Commun. 2021.

Abstract

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for investigating the human brain. However, the inherently poor signal-to-noise-ratio (SNR) of the fMRI measurement represents a major barrier to expanding its spatiotemporal scale as well as its utility and ultimate impact. Here we introduce a denoising technique that selectively suppresses the thermal noise contribution to the fMRI experiment. Using 7-Tesla, high-resolution human brain data, we demonstrate improvements in key metrics of functional mapping (temporal-SNR, the detection and reproducibility of stimulus-induced signal changes, and accuracy of functional maps) while leaving the amplitude of the stimulus-induced signal changes, spatial precision, and functional point-spread-function unaltered. We demonstrate that the method enables the acquisition of ultrahigh resolution (0.5 mm isotropic) functional maps but is also equally beneficial for a large variety of fMRI applications, including supra-millimeter resolution 3- and 7-Tesla data obtained over different cortical regions with different stimulation/task paradigms and acquisition strategies.

© 2021. The Author(s).

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing interests.

Figures

Fig. 1

Fig. 1. Stimuli and paradigm, epi images, and tSNR.

Panel A depicts the visual stimuli (left) used and a schematic of the visual presentation paradigm (right). Panel B shows an example slice from a single volume extracted from an fMRI time series for Standard (left column) and NORDIC (right column) reconstructions before any preprocessing, for two subjects S1 and S2. Panel C shows average (across all 8 runs) brain temporal signal-to-noise ratio (tSNR) maps of 2 exemplar slices in 2 representative subjects (S1 and S2) for NORDIC (left) and Standard (center) reconstructions and the normalized difference between the 2. The last was computed by performing (tSNRNORDIC -tSNRSTANDARD)/tSNRNORDIC). The slices chosen represent one of the anterior-most slices in the covered volume, and an occipital slice that includes a portion of the target ROI in V1.

Fig. 2

Fig. 2. NORDIC vs. Standard _t_-Maps.

Leftmost panel shows functional images as _t_-maps (target > surround) thresholded at |t| ≥ 5.7 for a single NORDIC processed run, and for 1, 3, and 5 Standard processed runs combined, for subject 1 (S1), subject 2 (S2), subject 3 (S3) and subject 4 (S4).

Fig. 3

Fig. 3. Voxel responses within target ROI.

Panel A shows the single-run (arranged over the _x_-axis) _t_-values (activity elicited by the target >0) induced by the target stimulus for Standard (red) and NORDIC (blue) data. Panel B is the same as panel A, but for beta weights (transformed into percent signal change). Panel C shows the single-run standard deviation computed across single-trial PSC beta estimates elicited by the target condition. For these three panels, gray dots represent responses to single voxels with the target ROI (497 for S1 and 461 for S2). The box-and-whisker plots, computed across all ROI voxels, represent the interquartile range (IQR—with box limits being the upper and lower quartiles), with the whiskers extending 1.5 times the IQR or to the largest value. The horizontal lines within the boxplot represent the median, while the diamond the mean across voxels. Panel D shows the target ROI, representing the left retinotopic representation of the target in V1 for 2 exemplar subjects in all three planes. Panel E shows the single runs, single-voxel scatterplots for _t_-values (activity elicited by the target > 0), for Standard (_x_-axis), and NORDIC (_y_-axis). Panel F: same as panel E for the beta percent signal change responses to the target condition. Source data are provided as source Data files.

Fig. 4

Fig. 4. PSC maps and cross-validated prediction accuracy.

Panel A: Average percent signal change (PSC) maps and cross-validated _R_2 for both subjects S1 and S2. The top 2 rows in Panel A show the average (across runs) PSC maps elicited by the target (left), surround (middle) and their contrast target > surround (right), for NORDIC and Standard reconstructions, respectively. As evident by these images, PSC amplitude and the extent of stimulus-induced signal change is comparable across reconstructions (see also Supplementary Fig. 5). The 3rd row in panel A shows the average (across folds) cross-validated _R_2 maps for NORDIC (left), Standard (middle), and their difference (right). Only in the relevant portion of the cortex for the stimulus used (i.e. areas where stimulus-induced BOLD activity is expected, as indicated by the PSC maps) do the _R_2 maps show higher precision of PSC estimates for NORDIC images. Panel B: For each subject, the left column shows a single run of NORDIC and Standard BOLD time-courses for a target-selective voxel (lighter lines) and its prediction estimated on a separate run (darker lines). These plots highlight a closer correspondence between model prediction and empirical time-courses for NORDIC data, as summarized by the significantly larger (2-sided paired sample _t_-test; S1: t(14) = 12.8, ci(11.83;16.6), Cohen’s d: 3.31 p < 0.01 Bonferroni corrected; S2: t(14) = 20.1, ci(13.33;16.25), Cohen’s d: 5.41 p < .01 Bonferroni corrected) target ROI average cross-validated _R_2 (bar plots represent the mean and error bars indicate standard errors of the mean across 15 cross-validation folds shown as gray dots). Source data are provided as source Data file.

Fig. 5

Fig. 5. Functional point spread function (PSF) and global image smoothness.

Panel A, top row shows NORDIC normalized beta percent signal change (PSC) maps for differential mapping target (in red) > surround (in blue) (left), and the target only (right) single-condition image for subjects 1 and 2. The white dotted line is determined in the differential image as the “boundary” between the two stimulations. The same white dotted line is also superimposed on the target-only PSC map where PSC values are greater than zero but decreasing in magnitude progressively away from this “boundary” posteriorly. The functional PSF is calculated from this spread in PSC beyond the “boundary”. Panel A, the lower row is identical to the upper row but obtained from Standard reconstruction data. Panel B left panel for each subject: the PSC magnitude changes (normalized to the highest value) along traces perpendicular to the “boundary” are displayed as the average (across traces and runs). The model fits (solid line) and data (dotted line) are shown for both the NORDIC (blue) and Standard (red) reconstructions. The vertical gray dotted line represents the “boundary” as derived from the differential maps. Panel B, right panel for each subject portrays the full-width at half maximum (FWHM) standard deviation of the gaussian kernel that was convolved with a step function to model functional PSF (see the “Methods” section). Panel C: Mean global smoothness of images used for the fMRI time series for Standard (red) and NORDIC (blue) in four subjects, before (left panel) and after preprocessing related interpolations. Error bars represent the standard error of the mean across 6 independent runs (shown as gray dots). Source data are provided as a source Data file.

Fig. 6

Fig. 6. 3D GE EPI images and fMRI data obtained with 0.5 mm isotropic voxels.

Panel A shows a single slice from a single time point in the consecutively acquired volumes forming the fMRI time series for Standard (left) and NORDIC (middle) images. The right panel shows the average of 10 images of the same slice for the Standard reconstruction. Panel B shows _t_-thresholded (|t| ≥ 2.9) functional maps (for the contrast target > surround on a _T_1 weighted anatomical image for standard (left) and NORDIC (right) reconstructions for a saggital and axial slice (with related zoom-ins on the sagittal (blue) and axial (red) planes). Panel C shows the same _t_-maps as in panel C on the inflated cortical space and at different _t_-thresholds. No spatial smoothing or masking was applied to the data.

Fig. 7

Fig. 7. Flowchart of the NORDIC algorithm for a series m(r,τ).

First, to ensure i.i.d. noise the series is normalized with the calculated _g_-factor kernels as m(r,τ)/g(r). From the normalized series, the Casorati matrix Y=[y1,⋯,yj,⋯,yN] is formed, where yj is a column vector that contains the voxel values in each patch. The low-rank estimate of Y is calculated as YL=U⋅Sλthr⋅VT, where the singular values in S, λ(i) are set to 0 if λ(i)<λthr. After re-forming the series mLLR(r,τ) with patch averaging, the normalization with the calculated _g_-factor is reversed as mNORDIC(r,τ)=mLLR(r,τ)⋅g(r).

Similar articles

Cited by

References

    1. Ogawa S, et al. Intrinsic signal changes accompanying sensory stimulation: functional brain mapping with magnetic resonance imaging. Proc. Natl Acad. Sci. USA. 1992;89:5951–5955. doi: 10.1073/pnas.89.13.5951. - DOI - PMC - PubMed
    1. Kwong KK, et al. Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc. Natl Acad. Sci. USA. 1992;89:5675–5679. doi: 10.1073/pnas.89.12.5675. - DOI - PMC - PubMed
    1. Bandettini PA, Wong EC, Hinks RS, Tikofsky RS, Hyde JS. Time course EPI of human brain function during task activation. Magn. Reson. Med. 1992;25:390–397. doi: 10.1002/mrm.1910250220. - DOI - PubMed
    1. Van Essen DC, Ugurbil K. The future of the human connectome. Neuroimage. 2012;62:1299–1310. doi: 10.1016/j.neuroimage.2012.01.032. - DOI - PMC - PubMed
    1. Sudlow C, et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 2015;12:e1001779. doi: 10.1371/journal.pmed.1001779. - DOI - PMC - PubMed

Publication types

MeSH terms

Grants and funding

LinkOut - more resources