Multi-echo fMRI: A review of applications in fMRI denoising and analysis of BOLD signals

TL;DR: Evidence is shown that the multi‐echo approach expands the range of experiments that is practicable using fMRI, and a compelling future role of the multi-echo approach in subject‐level and clinical fMRI is suggested.

About: This article is published in NeuroImage.The article was published on 2017-07-01 and is currently open access. It has received 218 citations till now. The article focuses on the topics: Resting state fMRI.

Summary

Introduction

• Recent studies show that fMRI data can be severely affected by artifacts (Power et al., 2012).
• These artifacts relate to subject head motion, cardiac and respiratory effects, and hardware (Glover et al., 2000; Jo et al., 2010).
• Studies on the effects of fMRI artifacts have brought into question many of the compelling findings on brain function based on fMRI, for example, relating to human brain development (Fair et al., 2009).

Available online 29 March 2017

• Thus, while in recent years the field of fMRI research has both enjoyed advanced technology and expanded use, it also deals with a deep discomfort related to many known and unknown effects of artifact.
• The authors discuss an emerging fMRI approach, called multi-echo (ME)-fMRI, which focuses on improving the fidelity of fMRI signals through a physically-driven determination of the origins of fMRI signals as arising from either BOLD contrast or artifact.
• To date, the challenges in controlling fMRI artifacts have been met with generic time series signal processing methods such as regression and frequency-restricting bandpass filters (Satterthwaite et al., 2012).
• Spontaneous activity is band pass filtered to retain a narrow range of frequencies in order to exclude hardware related signal drifts and high-frequency noise (Carp, 2013).
• This altogether means that artifact signals are not characterized well enough by modeling artifact time courses and regressing them out of data.

ME-fMRI and fMRI relaxometry

• After excitation, standard fMRI uses 2-D echo planar imaging to acquire slice images at a single TE, one slice at a time.
• S0 and T2* may be estimated from just a single volume of fMRI data.
• A voxel's T2* value determines not only its signal intensity scaling with TE, but also indicates the TE at which the largest amplitude signal change due to BOLD contrast is detected (Fig. 5).
• This suggests that signal changes due to R2* (i.e. T2*) and S0 may be separable from each other.

A general model for fMRI time series

• ME-ICA showed how brain network components could be grouped apart from artifacts of many kinds without spatial or temporal templates of expected functional effects.
• Last, the number of signal PCs is estimated as the rank where the variance of a data PC is most likely different than the variance of the corresponding noise PC.
• The general model for ME-fMRI data proposed here involves the separation of all data into MR signals and Gaussian noise, and then MR signals into BOLD and non-BOLD sources.
• These results showed that ME-ICA and its model for BOLD, non-BOLD signals and noise could improve resting state studies of human, but could also be used to study novel brain function, such as drug manipulations in animals.
• From ME-ICA, these artifacts tend to manifest with: higher κ than conventionally TE-independent artifacts, lower κ values than functional networks, higher ρ than components of neurally-related BOLD signals (Kundu et al., 2011, Kundu et al., 2013), and typically high levels of variances explained.

Motion artifacts

• While motion artifacts have long been known to affect fMRI, only recently has the severity of their impact on studies of connectivity become widely accepted (Power et al., 2012; Satterthwaite et al., 2012).
• Based on the general ME-fMRI model, ME-ICA has been shown to separate out motion artifacts as non-BOLD signals.
• In some conditions, standard denoising seems to overfit noise models and increase DVARS values.
• While data of these two groups led ME-ICA to explain similar percentages of data variance, the number of BOLD components in high motion data was lower (Kundu et al., 2013).
• The comparison of expected to observed errors in group-level connectivity maps is shown for two group-level connectivity contrasts, one based on ME-ICR subject-level seed connectivity maps, the other based on standard connectivity maps (Fig. 15).

ME-fMRI using MB EPI and 7 T MRI

• On the one hand, high spatial resolution is important for localizing brain function precisely.
• One motivation was to enable MB-fMRI on the GE platform that could be compatible with the sequence on the Siemens platform used to acquire data compatible with the Human Connectome Project (HCP) (Van Essen et al., 2012).
• These results together suggest that fMRI using the combined MEMB approach could be superior to using standard single-echo multi-band EPI at 7 T.
• In summary, to achieve faster volume imaging speed for multi-echo data, several hardware and pulse sequence approaches can be used, with multi-band EPI being highly promising.
• Studies involving homogeneous samples of participants where functional and artifact components would be expected to be repeatable across subjects could benefit more from maximizing imaging speed and not spending time acquiring multiple echoes.

ME-fMRI of non-normative neuroanatomy

• The use of ME-ICA at 7 T could make high-resolution fMRI studies possible for anatomies that are non-standard since ME-ICA needs no comparisons to standard anatomy to identify networks and artifacts.
• TEs using 7 T MRI acquired from controls and epilepsy patients, analyzed with ME-ICA.
• T1 and T2 images show a dysembrogenic neuroepithelial tumor (DNET) in the right superior temporal lobe, affected by edema (dark area in Fig. 18).
• By playing high-κ BOLD time series as a dynamic movie, transient BOLD signals can be seen arising near the lesion, suggesting functional cortex.
• Further study can help to determine to what extent MEICA and MEMB-EPI at 7 T could be used for studying lesional cortex when both high spatial resolution and functional sensitivity is needed.

Making ME-fMRI practical

• It might be surprising that with all the information to be gained from ME-fMRI over standard fMRI, ME-fMRI is still not standard.
• ME-ICA writes three different time series datasets that vary by how strictly BOLD versus non-BOLD signals are retained, with a rationale based on end use (Fig. 19).
• The removal of drifts in multi-echo fMRI can be done by removing non-BOLD signals, which makes possible the detection of low-frequency activity related to BOLD contrast.
• It is key to note that, at group-level analysis, this reduction of non-BOLD noise also reduced between-subject variance.
• Moreover, subcortical areas link to each other and to the cerebral cortex in complex functional relationships.

Conclusion

• Over the history of the development of fMRI methods, ME-fMRI has been important in the process of validating BOLD-related origins of novel fMRI signal observations, such as resting state time series correlation.
• The past limitations of ME-fMRI in regards to whole-brain coverage or standard resolutions are mostly gone, enabling its use in a wide range of fMRI studies.
• At present, MEfMRI is available in research sequences from the authors covered in this review, on many current and upcoming scanners from major vendors (Olafsson et al., 2015; Poser et al., 2006).
• From using ME-ICA the authors find that higher estimates of data dimensionality may help in separating complex BOLD signals and artifact while keeping a stable ICA.
• These features suggest that the ME approach may be beneficial in tackling emerging issues of statistical power in task-related as well as resting state data.

Figures

Fig. 4. A representative case of using T2*maps to drive anatomical-functional coregistration. (A) T2* map generated from multi-echo fMRI data. The T2* map is thresholded to mask out CSF, by modeling CSF as having twice the T2* of gray matter. This masked T2* map is segmented, to generate a T2* weight mask for driving anatomical-functional segmentation. (B) T2* weighted coregistration uses local Pearson correlation to estimate an affine warp. The weight maps show high intensity in gray matter, including hippocampus. Native coregistation (i.e. rigid-body) mis-registers subcortical areas. Standard affine warp based on global estimate overscales the anatomical image.

Fig. 11. Components from ME-ICA of a single 10 min. run of resting state ME-fMRI acquired at 3 T using 3.75 mm isotropic voxel size, using ME-PCA dimensionality estimation. Components show: (a) limb representations of motor cortex, (b) foot area, (c) left-lateralized hand area, (d) language network, (e) right-lateralized hand area, (f) primary visual cortex (g) thalamus.

Fig. 15. (A) T-tests of seed-connectivity maps in various permuted subject groups based on level of subject head motion, for seven seeds. Over an average of 50 random groupings, both ME-ICR (blue) and conventional (yellow) connectivity have type I error that match the expected error (black). As groupings are biased by the level of head motion (50 permutations), conventional analysis shows doubled type I error. In contrast, ME-ICR shows good type I error control (50 permutations). For the single test of high versus low motion subjects based on median split, ME-ICR shows good type I error control, while conventional connectivity shows high type I error. (B) Regions that show connectivity differences due to type I error, after cluster correction. Adapted from Kundu et al. (2013).

Fig. 5. TE-independence of non-BOLD versus TE-dependence of BOLD fMRI signal changes as underlies artifact and activation, respectively. Exponential decays of signal with TE that differ (a) in intercept S0 (spin-density), and (b) in decay rate (relaxation rate, due to magnetic susceptibility). Computed difference in signal at different TEs due to (c) artifact-related change in S0, ΔS0 and (d) BOLD-related change R2*, ΔR2*. Computed difference in percent signal change from mean (e) ΔS0 and (f) Δ R2* such as due to BOLD contrast.

Fig. 6. Combination of echo images from a multi-echo fMRI experiment to give optimally combined image. Signal dropout artifact is due to short T2* in high susceptibility areas, and echo combination weights towards early TE signals in short T2* areas. (A) For images at 3 T (B) and 7 T.

Fig. 10. Κ and ⍴ spectra from ME-PCA (left), and from ME-ICA given ME-PCA dimensionality reduction (right). (A) ME-PCA shows that PCA components of fMRI data have coincidentally high Κ and ⍴ weights, indicating mixed TE-dependence and TE-independence. This pattern in turn suggests that PCA components of fMRI data have both BOLD and nonBOLD contributions. ME-PCA also shows that PCA components, past a low-dimensional estimate, have detectable MR signals. A higher dimensionality estimate is obtained based on TEdependence and independence weights (B) ME-ICA unmixing of signals after ME-PCA dimensionality reduction. Two distinct regimes appear, one with high Κ and low ⍴, and the other with vice-versa. These regimes indicate a specifically R2* weighted and a non-R2* weighted regimes, found in terms of ICA component groups.

Fig. 9. Comparison of signal processing models for single-echo and multi-echo fMRI data. (A) Single-echo processing steps indicate nuisance types to be modeled. (B) Multiecho processing based on two steps, ME-PCA and ME-ICA, that separate thermal noise, BOLD and non-BOLD signals.

Fig. 14. Multi-echo independent coefficients regression (ME-ICR). At subject-level, after finding BOLD components, vectors of IC coefficients can be used to computed seed-based connectivity analysis. IC weights are independent, meaning ordinary least squares fits are suited to conventional statistical significance of connectivity estimates. Pearson correlations of IC vectors can be Fisher transformed to Z-values that factor in degrees of freedom.

Fig. 1. Multi-echo fMRI images from 3 T and 7 T. (A) 3 T multi-echo images with inplane and multi-band acceleration with 2.5 mm isotropic resolution and < 1 s TR. (B) 7 T multi-echo images with GRAPPA factor 3 and 2.5 mm isotropic resolution, no multiband acceleration, giving TR=2.5 s.

Fig. 18. ME-ICA of data from 7 T ME-fMRI of focal lesional epilepsy patients. (A) Default mode network of patient with dysembryonic neuroepithelial tumor (DNET), rendered on T1 and T2 MRI. (B) Other canonical functional networks of patient with DNET. (C) Seed-based functional connectivity of DNET lesion. Seed in parenchyma of lesion shows abnormal connectivity with putamen and cerebellum. (D) Seed connectivity for patient with venous cavernoma, with seed in hippocampus near abnormality. Connectivity is seen from seed region to ipsilateral superior temporal lobe and insula.

Fig. 13. Comparison of subject motion, signals before and after motion regression, and ME-ICA BOLD (High-K) and non-BOLD (Low-K) time series. The framewise displacement (second panel) matches the DVARS trace of raw data and of motion parameter regressed data (third panel). This pattern suggests that motion regression did not remove the main problematic variance due to motion, with possible indications of overfitting. In contrast, the BOLD time series DVARS is separated from that of the non-BOLD and raw (overlapping), indicating an effective denoising.

Fig. 8. Multi-echo independent components analysis (ME-ICA). (A) κ and ⍴ pseudo-F statistics representing component-level TE-dependence and independence, of BOLD and artifact, respectively. (B) In the first three columns, the percent signal changes across TEs, of network and artifact components in top and bottom rows, respectively. The last two columns show parameter maps from fitting percent signal changes to TE-dependence and TE-independence models, thresholded by goodness of fit FR2* and FS0, respectively. (C) Κ and ⍴ spectra, for 7 T and 3 T data.

Fig. 7. TE-dependent scaling of BOLD percent signal change during task and rest. (a) Scaling of checkerboard-task related time course from primary visual cortex, showing clear block structure. (b) TE-dependent scaling of resting state signals from precuneus. While resting state time course is not predictable beyond a frequency range, TEdependent scaling suggests a BOLD effect.

Fig. 2. T2* relaxometry. (A) T2* decay based on simple monoexponential decay model in Eq. (1), as function of intercept S0 and decay time T2*, inverse of rate R2*. (B) False color multiecho EPI images from a 3-TE multi-echo fMRI experiment. (C) Maps of parameters S0 and T2*. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

Fig. 21. Detection of sigmoidal BOLD effect after separating BOLD from non-BOLD signals. Raw fMRI from primary visual cortex during slow visual contrast change does not show slow BOLD change. Conventional time course filtering does not expose a slow varying effect. After ME-ICA, BOLD signal shows sigmoidal effect that localizes to visual cortex. Non-BOLD signal shows linear drift consistent with usual fMRI artifact. Adapted from Evans et al. (2015).

Fig. 19. Pipeline of component and time series outputs from ME-ICA. Key time series from multi-echo fMRI acquisition and ME-ICA analysis: (A) Time series of "middle" echo, can be used for standard fMRI analysis. (B) Time series after combining echoes, can be used for standard analysis. These time series have enhanced functional contrast and have compensated for dropout artifact. (C) Time series with non-BOLD artifact removed, but retaining Gaussian-like noise. (D) Time series from only BOLD components, suitable for studies of BOLD dynamics.

Fig. 20. Cortical dynamics captured in high-Κ time series. Using multi-echo multi-band fMRI sequence with in-plane acceleration factor (GRAPPA) 3, and multi-band factor 3, leading to 2.5 mm isotropic resolution sampled at TR=0.87 s. In evolution over about 10 s, the transition between two states of spontaneous activity states is seen.

Fig. 17. Components from ME-ICA at 7 T MRI of healthy individual. Canonical cortical networks, such as the default mode network are found with good spatial resolution. In addition, subcortex is parcellated into functionally distinct networks, even at the individual-subject level.

Fig. 12. ICA of spontaneous activity from individual anesthetized rat (under 1% inhaled isoflurane). (A) ME-ICA components show functional networks sharply localized to anatomical regions. These components are ranked based on BOLD weighting. (B) FSL MELODIC shows some comparable BOLD networks, but with ranking based on variance explaining mixing artifacts with networks. FSL MELODIC components show a high degree of mixing between network and artifact patterns.

Fig. 23. Parcellation of thalamus and midbrain gray matter nuclei based on seed-based connectivity. Adapted from Morris et al. (2016a, 2016b).

Fig. 22. Estimates of statistical power for an activation task of the mentalizing domain, sampled across several brain regions. Three methods are compared, including optimal combination and standard time series analysis, and GLMdenoise (Kay et al., 2013). (A) Mean activation is not different between different methods. (B) Inter-subject variability is considerably lower after ME-ICA denoising. (C) Power for detecting vmPFC activation is substantially higher for ME-ICA than other methods. Adapted from Lombardo et al. (2016).

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Multi-echo fmri_ a review of applications in fmri denoising and analysis of bold signals" ?

In this review the authors discuss an emerging fMRI technology, called multi-echo ( ME ) -fMRI, which focuses on improving the fidelity and interpretability of fMRI. This review covers recent multi-echo fMRI acquisition methods, and the analysis steps for this data to make fMRI at once more principled, straightforward, and powerful. These findings suggest a compelling future role of the multi-echo approach in subject-level and clinical 

Q2. What have the authors stated for future works in "Multi-echo fmri_ a review of applications in fmri denoising and analysis of bold signals" ?

Emerging fMRI methods using multi-band acceleration and ultra-high field MRI may further make the ME approach attractive, and help counter new artifacts in more demanding studies, such as at 7 T ( Boyacioğlu et al., 2014 ). With an ME-fMRI approach, there exists the possibility to guide statistics with ‘ physical ’ signal models, as the authors demonstrated in ME-ICA results. However, looming issues include how to study regions that are smaller than the smoothness factor of images, such as in subcortex, especially when data are smoothed outright to reduce noise as in standard fMRI. Much work does remain on further validating signals after ME-ICA and related processing.