The effect of artifacts on dependence measurement in fMRI

Arthur Gretton, Andrei Belitski, Yusuke Murayama, Bernhard Schölkopf, Nikos K. LogothetisMagnetic 2006 · DOI

TL;DR

This paper investigates how common artifacts in fMRI data, specifically those from image reconstruction and physiological noise (breathing), can artificially inflate measures of functional connectivity. By carefully analyzing these artifacts in macaque visual cortex, the authors demonstrate that their removal is crucial for accurately assessing true neural dependencies and understanding brain networks.

What Problem Does This Paper Address?

The study of effective connectivity in neuroscience aims to understand how different brain regions interact. Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for this, measuring brain activity indirectly through blood oxygenation changes (BOLD signal). However, fMRI data is susceptible to various artifacts that can mimic or distort genuine neural signals. This paper focuses on two specific types of artifacts:

  1. Image Reconstruction Artifacts: These arise during the process of converting raw MRI signals into images. A common example is a "Gibbs ringing" effect, which creates spurious correlations between nearby voxels, particularly in the readout direction of the scan.
  2. Physiological Noise (Breathing): Respiration causes fluctuations in the magnetic field within the scanner, leading to oscillations in the BOLD signal. These breathing-induced signals can create artificial dependencies between voxels, especially those close to large blood vessels.

The core problem is that these artifacts can lead to misleading interpretations of functional connectivity, making it appear as though brain regions are more interconnected than they actually are, or obscuring real connections. This is particularly problematic when aiming for model-free measures of dependence, which don't rely on pre-defined assumptions about neural interactions.

Methods

The researchers employed a combination of experimental techniques and statistical measures to address their research question:

  • High-Field fMRI: They used a 4.7 Tesla fMRI scanner to acquire high-resolution data from the primary visual cortex (V1) of an anesthetized macaque monkey. This high field strength allows for better signal-to-noise ratio and spatial resolution.
  • Stimulus Presentation: Monkeys were presented with rotating polar-transformed checkerboard gratings to elicit a visual response.
  • Dependence Measures: Three statistical measures were used to quantify the similarity between activity in different voxels:
  • Correlation: A standard measure of linear association, sensitive to second-order interactions.
  • Mutual Information (MI): A more general measure that captures non-linear dependencies and information shared between variables.
  • Constrained Functional Covariance (COCO): A novel, model-free measure designed to detect general dependencies, particularly useful for complex signals.
  • Artifact Isolation and Removal:
    • Phantom Scans: To isolate image reconstruction artifacts, scans were performed on a water-filled phantom, which has no biological signal. This allowed them to observe the "rippling" effect in isolation.
  • Breathing Artifact Removal: A projection method was used to remove breathing artifacts. This involved identifying the breathing frequency and phase in the fMRI signal and projecting out the corresponding sinusoidal component. This was done by analyzing the spectral peaks at the breathing frequency and its harmonics.
  • Analysis of Dependence vs. Distance: The researchers systematically analyzed how dependence measures changed with the spatial distance between voxels, both in the horizontal (readout) and vertical (phase-encode) directions.

Datasets

  • Macaque Visual Cortex (V1) fMRI Data: The primary dataset consisted of fMRI scans acquired from the primary visual cortex of an anesthetized macaque monkey during visual stimulation.
  • Phantom Data: A water-filled phantom was used to specifically study image reconstruction artifacts.

No new benchmarks were introduced; the focus was on analyzing existing fMRI data with novel artifact removal techniques.

Key Insights and Ideas

  • Artifacts are Not Trivial: The study powerfully demonstrates that seemingly minor artifacts in fMRI data can significantly distort measures of functional connectivity, leading to potentially incorrect conclusions about brain networks.
  • Directional Dependence: Image reconstruction artifacts (rippling) primarily affected dependence in the horizontal (readout) direction, while breathing artifacts impacted both horizontal and vertical directions. This highlights the importance of considering scan acquisition parameters.
  • Breathing's Impact on Stimulus Correlation: Breathing artifacts not only affected inter-voxel dependence but also reduced the correlation between individual voxels and the visual stimulus. This suggests that removing breathing artifacts can improve the sensitivity to neural responses.
  • Averaging vs. Removal: While averaging fMRI responses over many stimulus repetitions can mitigate the effect of breathing, it can also obscure subtle inter-voxel dependencies that might be trial-specific. Therefore, direct artifact removal is often preferable.
  • Model-Free Dependence is Sensitive: The use of general dependence measures like MI and COCO, while powerful for uncovering complex interactions, also makes them more susceptible to structured noise if not properly addressed.

Key Results

  • Phantom Scans: Showed strong, short-range correlations in the horizontal direction due to image reconstruction artifacts, while the vertical direction remained largely unaffected.
  • Macaque V1 Scans:
    • Breathing artifacts were found to artificially boost short-range dependence and cause dependence to decay more slowly with increasing distance.
  • Removing breathing artifacts led to a more accurate representation of dependence vs. distance, particularly in the vertical direction.
  • Removal of breathing artifacts increased the number of voxels significantly correlated with the visual stimulus.
  • The dependence in the horizontal direction was generally greater and over a longer range than in the vertical direction, partly due to the inherent visual topography of V1 and partly due to the rippling artifact.

Limitations and Future Work

  • Anaesthetized Subject: The study was conducted on an anesthetized macaque. The authors acknowledge that respiration in awake humans is more variable in frequency and amplitude, making artifact removal more challenging.
  • Cardiac Artifacts: The paper briefly mentions cardiac activity as another source of structured noise, which is more complex to model and remove due to its non-sinusoidal nature and aliasing at typical fMRI sampling rates.
  • Future Directions:
    • Investigating the source of the rippling artifact in image reconstruction to enable its direct removal or compensation.
  • Developing more robust methods for removing physiological noise, particularly cardiac artifacts, which are more challenging to handle.
  • Extending these artifact analysis and removal techniques to human fMRI studies.

Summary updated

BibTeX
@inproceedings{gretton2006effect,
  author = {Arthur Gretton and Andrei Belitski and Yusuke Murayama and Bernhard Schölkopf and Nikos K. Logothetis},
  title = {{The effect of artifacts on dependence measurement in fMRI}},
  booktitle = {Magnetic},
  year = {2006},
  doi = {10.1016/j.mri.2005.12.036},
}

Loading related content...

Figure