Impact of automated ICA-based denoising of fMRI data in acute stroke patients

Abstract

Different strategies have been developed using Independent Component Analysis (ICA) to automatically de-noise fMRI data, either focusing on removing only certain components (e.g. motion-ICA-AROMA, Pruim et al., 2015a) or using more complex classifiers to remove multiple types of noise components (e.g. FIX, Salimi-Khorshidi et al., 2014 Griffanti et al., 2014). However, denoising data obtained in an acute setting might prove challenging: the presence of multiple noise sources may not allow focused strategies to clean the data enough and the heterogeneity in the data may be so great to critically undermine complex approaches. The purpose of this study was to explore what automated ICA based approach would better cope with these limitations when cleaning fMRI data obtained from acute stroke patients. The performance of a focused classifier (ICA-AROMA) and a complex classifier (FIX) approaches were compared using data obtained from twenty consecutive acute lacunar stroke patients using metrics determining RSN identification, RSN reproducibility, changes in the BOLD variance, differences in the estimation of functional connectivity and loss of temporal degrees of freedom. The use of generic-trained FIX resulted in misclassification of components and significant loss of signal (< 80%), and was not explored further. Both ICA-AROMA and patient-trained FIX based denoising approaches resulted in significantly improved RSN reproducibility (p < 0.001), localized reduction in BOLD variance consistent with noise removal, and significant changes in functional connectivity (p < 0.001). Patient-trained FIX resulted in higher RSN identifiability (p < 0.001) and wider changes both in the BOLD variance and in functional connectivity compared to ICA-AROMA. The success of ICA-AROMA suggests that by focusing on selected components the full automation can deliver meaningful data for analysis even in population with multiple sources of noise. However, the time invested to train FIX with appropriate patient data proved valuable, particularly in improving the signal-to-noise ratio.

Keywords: Acute stroke; BOLD; Denoising; Independent component analysis; Resting state; fMRI.

Fig. 1

RSN identifiability/reproducibility and lost temporal degrees of freedom (tDoF). RSN identifiability was defined using a Z-score ratio between the Z-scores within and outside a RSN. RSN reproducibility reflects the spatial correlation between group-level RSN maps for random splits of the samples normalized to pseudo Z-scores. The loss of tDoF is expressed as a percentage of the total tDoF, defined as the available number of time points. *** p < 0.001.

Fig. 2

Spatial pattern of change in BOLD signal standard deviation after using patient-trained FIX and ICA-AROMA. (A, C) Probability maps, representing areas where the BOLD variance was affected more frequently in patients. (B, D) Intensity maps, representing areas where voxels %ΔSTD was significantly more reduced in patients.

Fig. 3

Effects of ICA-AROMA and patient-trained FIX on resting state connectivity estimation. After performing a template-based dual regression analysis on the original data and on the data cleaned using ICA-AROMA and patient-trained FIX, differences in estimation of functional connectivity were evaluated using |%ΔZ |. *** p < 0.001.

Fig. 4

Group changes in functional connectivity in the 10 “well defined” networks. For each of the 10 “well defined” networks (red-yellow), significant group level changes (p < 0.05 TFCE) in functional connectivity are shown after the use of ICA-AROMA (green) and after the use of patient-trained FIX (blue). From top to bottom: visual-medial, visual-occipital, visual-lateral networks, the default mode network, the cerebellar network, the sensorimotor network, the auditory network, the executive-control and the left and right fronto parietal networks. (For interpretation of the references to color in this figure legend, the reader is referred to the online version of this chapter.)