Artifactual time‐course correlations in echo‐planar fMRI with implications for studies of brain function

Nikolaus Kriegeskorte, Jerzy Bodurka, Peter A. BandettiniInternational Journal of Imaging Systems and Technology 2008 · DOI

TL;DR

This paper reveals that correlations between fMRI voxel time courses, often interpreted as neuronal interactions, are significantly influenced by technical artifacts inherent in echo-planar imaging (EPI). By comparing a human brain scan with a water-filled phantom, the researchers demonstrate that these artifactual correlations are surprisingly similar to those observed in brain data, suggesting that many observed "neuronal" connections might be imaging-related noise.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone for understanding brain activity by measuring changes in blood oxygen levels, which are assumed to reflect neuronal activity. A common analytical approach involves examining correlations between the time series of different brain regions (voxels). These correlations are frequently interpreted as evidence of functional or effective connectivity, implying direct or indirect neuronal interactions. However, the authors highlight a critical issue: the fMRI acquisition process itself, particularly Echo-Planar Imaging (EPI), can introduce spurious correlations between voxel signals. This means that observed correlations might not accurately represent actual brain communication, potentially leading to misinterpretations of brain function. The paper aims to empirically demonstrate the extent and nature of these artifactual correlations.

Methods

The core of the study involves comparing voxel time-course correlations in two distinct scenarios:

  1. Human Brain (Resting State): A human subject was scanned in a resting state with eyes closed using EPI.
  2. MRI Phantom: A water-filled sphere, devoid of any biological or neuronal activity, was scanned using EPI under similar conditions.

The researchers then calculated the linear correlation coefficient between the time courses of pairs of voxels. This was done as a function of:

  • Voxel Separation: The distance between the centers of the voxel pairs.
  • Orientation: Whether voxels were within the same EPI slice or in different slices, and their relative orientation along readout and phase-encoding directions.

To analyze these correlations, they plotted the average correlation as a function of voxel distance and also generated "neighbor-correlation maps" to visualize the spatial structure of these correlations.

Datasets

The study utilized two primary "datasets":

  1. Human Brain Data: Resting-state fMRI data from a single human subject.
  2. Phantom Data: fMRI data from a water-filled sphere (MRI phantom).

These datasets were used to empirically demonstrate the phenomenon of artifactual correlations. The phantom data serves as a crucial control, as any observed correlations within it must be attributed to the imaging process itself, not biological activity.

Key Insights and Ideas

  • Phantom Mimics Brain Correlations: The most striking finding is the remarkable similarity between the spatial patterns of voxel time-course correlations observed in the human brain and the water-filled phantom. This strongly suggests that a significant portion of observed correlations in fMRI data are not due to neuronal interactions but are instead artifacts of the EPI acquisition.
  • Artifacts Span Long Distances: These artifactual correlations are not confined to immediately adjacent voxels. They extend over several centimeters, both within and between imaging slices, challenging the assumption that long-range correlations necessarily imply widespread neuronal communication.
  • Short-Range Correlations are Strong: Neighboring voxels exhibit particularly strong positive correlations, which rapidly decrease with distance but remain detectable even at significant separations.
  • Implications for Connectivity Studies: The findings necessitate a re-evaluation of how functional and effective connectivity are interpreted. Researchers must be aware that observed correlations could be partially or wholly explained by instrumental factors, potentially leading to overestimation of neuronal interactions.

Key Results

  • Similar Correlation Decay: The plots of average voxel time-course correlation as a function of voxel separation for both the human brain and the phantom showed a similar pattern: strong correlations at short distances that gradually decrease but remain present even at separations of several centimeters.
  • Within-Slice vs. Between-Slice: Correlations were generally stronger for voxels within the same EPI slice compared to those in different slices.
  • Directional Differences: In the phantom data, correlations tended to be higher in the readout direction than in the phase-encode direction or between slices.
  • Phantom Correlations Vanish at ~8 cm: While brain correlations remained positive even at longer distances, phantom correlations tended to fall off to near zero within approximately 8 cm of voxel separation.

Limitations and Future Work

  • Artifactual vs. Genuine Signals: The paper acknowledges that while artifactual correlations are present, they do not imply that all observed correlations in the brain are artifactual. Genuine physiological and cognitive signals also contribute. The challenge is to disentangle these.
  • Characterizing Artifacts: The authors emphasize the need for better characterization of these artifactual components. This could lead to improved fMRI pulse-sequence design, more sophisticated statistical analysis techniques that account for these artifacts, or methods to correct for them.
  • Relative Contributions: Future work should focus on assessing the relative contributions of artifactual versus genuinely physiological and cognitive covarying processes to observed time-course correlations.
  • Mechanical Vibration: While mechanical vibration and head motion are considered potential causes of artifactual correlations, the authors note that the differing mechanical properties of the brain and phantom suggest these might not be the sole or primary drivers of the observed similarities.

The paper concludes by urging neuroscientists to consider these artifactual correlations when interpreting fMRI results, as they can significantly influence conclusions about brain function.

Summary updated

Abstract

Abstract Brain function is widely investigated with functional magnetic resonance imaging (fMRI) in humans and animals. In fMRI, the time courses of voxels typically reflect the local blood‐oxygen level, which is taken as an indicator of neuronal activity. Voxel time‐course correlations are often explicitly modeled and interpreted in terms of neuronal interactions. They also affect standard analyses that do not explicitly target neuronal interactions. As a consequence, time‐course correlations between voxels influence conclusions about cognitive and physiological brain processes in many studies. However, voxel correlations are known to arise not only from cognitive and physiological processes, but also as artifacts of fMRI techniques such as the commonly used echoplanar imaging. We empirically demonstrate this phenomenon by plotting time‐course correlation as a function of voxel separation for a human brain in resting state and for a water‐filled sphere (called a “phantom”). The phantom served as a test object known not to contain any interacting neuronal systems. The plots for brain and phantom are surprisingly similar. The correlational structure found in the phantom must be artifactual. Artifactual correlations spanning many centimeters occur within and between different imaging slices. Correlations between voxel time courses do not necessarily reflect brain processes. Instead, fMRI is affected by artifactual correlations, which are very strong for neighboring voxels and clearly present even at large distances. This needs to be taken into account in the neuroscientific interpretation of voxel correlations. © 2008 Wiley Periodicals, Inc. Int J Imaging Syst Technol, 18, 345–349, 2008
BibTeX
@article{kriegeskorte2008artifactual,
  author = {Nikolaus Kriegeskorte and Jerzy Bodurka and Peter A. Bandettini},
  title = {{Artifactual time‐course correlations in echo‐planar fMRI with implications for studies of brain function}},
  journal = {International Journal of Imaging Systems and Technology},
  year = {2008},
  doi = {10.1002/ima.20166},
}
Figure