A component based noise correction method (CompCor) for BOLD and perfusion based fMRI

Abstract

A component based method (CompCor) for the reduction of noise in both blood oxygenation level-dependent (BOLD) and perfusion-based functional magnetic resonance imaging (fMRI) data is presented. In the proposed method, significant principal components are derived from noise regions-of-interest (ROI) in which the time series data are unlikely to be modulated by neural activity. These components are then included as nuisance parameters within general linear models for BOLD and perfusion-based fMRI time series data. Two approaches for the determination of the noise ROI are considered. The first method uses high-resolution anatomical data to define a region of interest composed primarily of white matter and cerebrospinal fluid, while the second method defines a region based upon the temporal standard deviation of the time series data. With the application of CompCor, the temporal standard deviation of resting-state perfusion and BOLD data in gray matter regions was significantly reduced as compared to either no correction or the application of a previously described retrospective image based correction scheme (RETROICOR). For both functional perfusion and BOLD data, the application of CompCor significantly increased the number of activated voxels as compared to no correction. In addition, for functional BOLD data, there were significantly more activated voxels detected with CompCor as compared to RETROICOR. In comparison to RETROICOR, CompCor has the advantage of not requiring external monitoring of physiological fluctuations.

Figures

Figure 1

Schematic of the CompCor algorithm in which significant principal components derived from time-series data within noise regions-of-interest (nROI) are used to form an estimate Pest of the physiological noise matrix P. Incorporation of Pest into the general linear model for the signal in gray matter allows for estimation and removal of physiological fluctuations.

Figure 2

Areas with a high fraction of white matter and cerebrospinal fluid (CSF), as denoted by the magenta voxels, overlaid on their respective partial volume maps from a representative slice from Subject 1. White matter-only areas (panel a) were determined by first thresholding the white matter partial volume fraction map at 0.99 and then performing a map erosion by two pixels to minimize the effect of partial voluming with other tissue types. Panel b) displays CSF-only areas with a partial volume fraction greater than 0.99 with application of a nearest neighbor clustering criteria.

Figure 3

Panel (a) shows a spatial map of the fractional variance of physiological noise for the resting BOLD scan from a representative slice in subject 1. Panel (b) shows a spatial map of the temporal standard deviation (tSTD). Areas of high fractional variance of physiological noise correspond to areas of high tSTD. Panel c) compares the tSTD to the fractional variance of physiological noise on per voxel basis. Data points in red represent the 2% of voxels in the slice with the highest tSTD.

Figure 4

Mean fractional variance of physiological noise across subjects as a function of the fraction of voxels (sorted by tSTD) that is included in the noise ROI for BOLD (red) and ASL (blue) resting-state data. The black dotted line denotes the 2% threshold that is used in the present study.

Figure 5

Average normalized power spectra of components estimated with the application of various correction schemes to the resting BOLD run from subject 1. As shown in panel (a), cardiac and respiratory elements estimated by RETROICOR are located at 1.2 and 0.2 Hz, respectively. Application of aCompCor (panel b) or tCompCor (panel c) estimates components similar to the cardiac and respiratory elements identified by RETROICOR.

Figure 6

Average normalized power spectra of components estimated by the various correction schemes for the resting ASL run from subject 1. As shown in panel (a), cardiac (red) and respiratory (green) elements identified by RETROICOR are aliased due to the long TR. The power spectrum of components estimated by either aCompCor (panel b) or tCompCor (panel c) are similar to the sum of the cardiac and respiratory elements identified by RETROICOR.

Figure 7

Percent temporal standard deviation across gray matter voxels (partial volume >0.9) for uncorrected data (denoted as None) and data after application of RETROICOR (denoted as Phys) and CompCor for (a) resting-state BOLD data, (b) downsampled resting-state BOLD data, (c) first-echo resting state ASL perfusion data, and (d) second-echo resting-state BOLD data . Values are normalized on a per subject basis by the mean temporal standard deviation for the uncorrected data, so that the values for the uncorrected data are 100%. Diamonds represent a significant difference (p

Figure 8

Percent of significantly activated voxels across subjects (N=10) for uncorrected data (denoted as None) and data with application of RETROICOR (denoted as Phys) and CompCor for (a) periodic design BOLD data, (b) downsampled periodic design BOLD data, (c) first-echo block design ASL perfusion data, and (d) second-echo block design BOLD data. Values are normalized on a per subject basis by the number of activated voxels for the uncorrected data, so that the values for the uncorrected data are 100%. Diamonds represent a significant difference (p

Figure 9

Representative receiver operating characteristic curve showing the true positive rate versus false positive rate for uncorrected data (blue), RETROICOR (green), aCompCor (red), and tCompCor (cyan).

Figure 10

Example of the application of tCompCor in the presence of motion-related signal changes. The top row shows the original time series (red), the time series after the application of tCompCor (blue), and the periodic design reference function (black dashed). The middle row shows estimates of the roll (blue) and left-right displacement (green) time courses. The top three principal components as identified by tCompCor are shown in the bottom row.