Impact of multi-echo ICA modeling decisions on motor-task fMRI analysis

This study demonstrates that a conservative multi-echo ICA modeling approach, which orthogonalizes regressors to preserve task-related signals, significantly outperforms aggressive denoising methods in motor-task fMRI analyses, particularly when dealing with high task-correlated head motion and low signal-to-noise ratios.

The Gist

The Big Picture: Taking a Clear Photo of a Moving Brain

Imagine you are trying to take a photograph of a specific flower in a garden (the brain). You want to see exactly how the flower reacts when the sun hits it (the motor task, like moving a hand or foot).

However, there are two problems:

1. The Wind: The wind is blowing the flower around (this is head motion).
2. The Camera Shake: Because the flower is moving, you are shaking the camera to try to keep it in the frame. This makes the whole picture blurry (this is task-correlated motion, where moving your hand makes your head wiggle).

For years, scientists have used a special "Multi-Echo" camera that takes five photos at slightly different times to help fix the blur. They also use a smart filter called ME-ICA to separate the "real flower" (brain signal) from the "wind and shake" (noise).

The Problem: The scientists realized that their smart filter was sometimes too smart. In its eagerness to remove the wind, it accidentally cut off parts of the flower. If you are trying to photograph a tiny, delicate flower (like a foot movement), the filter might delete the flower entirely, thinking it's just noise.

The Three Strategies (The "Denoising" Models)

The researchers tested three different ways to use this filter to see which one kept the flower intact while still removing the wind.

1. The "Aggressive" Sweeper (The Over-zealous Janitor)

• How it works: This method looks at every piece of "junk" (noise) the camera found and throws it all in the trash, including anything that looks even slightly like the flower.
• The Analogy: Imagine a janitor sweeping a room. If they see a leaf that looks a little bit like a petal, they sweep it up just to be safe.
• The Result: The room is very clean (low noise), but you might have accidentally swept away the flower you were trying to study. This worked okay for big, strong flowers (hand grasps), but it wiped out the tiny foot and shoulder movements.

2. The "Moderate" Filter (The Selective Sorter)

• How it works: This method tries to be smarter. It looks at the junk and says, "If this piece of junk is moving exactly at the same time as the flower, I won't throw it away, because it might be part of the flower."
• The Analogy: The janitor stops and thinks, "Wait, that leaf is moving with the flower. I'll leave it alone."
• The Result: Better than the aggressive sweeper, but sometimes it still leaves too much wind in the picture, or it's not quite gentle enough for the tiniest flowers.

3. The "Conservative" Gardener (The Careful Pruner)

• How it works: This is the new method the paper champions. Instead of just throwing things away, it carefully separates the wind from the flower before deciding what to keep. It uses a mathematical trick called orthogonalization.
• The Analogy: Imagine the wind and the flower are tangled together in a knot. The Aggressive method cuts the whole knot. The Conservative method carefully unties the knot, separates the wind from the flower, and only throws the wind away, leaving the flower perfectly intact.
• The Result: The picture might still have a little bit of wind in the background (slightly noisier), but the flower is 100% visible and clear.

What They Found

The researchers tested these methods on two groups of people:

• Healthy people doing hand, shoulder, and foot tasks.
• People with Multiple Sclerosis (MS) doing similar tasks (who often move their heads more because of the condition).

The Key Discoveries:

1. The "Tiny Flower" Problem: When the brain signal was weak (like moving a shoulder or a foot) or when the person moved their head a lot, the Aggressive method failed. It deleted the signal. The Conservative method saved the day, showing clear brain activity where the others saw nothing.
2. The Trade-off: The Conservative method did leave a little bit more "wind" (noise) in the final picture compared to the Aggressive method. However, having a little bit of wind is better than having no flower at all.
3. Reliability: When people did the task twice, the Conservative method gave the most consistent results. It was the most reliable way to see the brain working, especially for difficult tasks.

The Bottom Line

If you are studying how the brain controls movement, especially if the movement is small (like a foot) or if the person has trouble staying still (like someone with MS), don't be too aggressive with your noise removal.

The paper suggests using the Conservative approach. Think of it as being a careful gardener: it's better to leave a little bit of dirt on the petal than to accidentally cut the petal off while trying to clean it. This ensures you don't miss the brain activity you are trying to find.

Technical Summary

1. Problem Statement

Multi-echo Independent Component Analysis (ME-ICA) is a robust denoising technique for functional MRI (fMRI) that improves sensitivity and reliability by distinguishing BOLD (neural) signals from non-BOLD artifacts (e.g., motion, physiological noise) using multi-echo acquisition. However, a critical trade-off exists in motor-task fMRI:

• The Conflict: Motor tasks often induce task-correlated head motion. Standard "aggressive" ME-ICA denoising (removing all components classified as noise) risks removing task-correlated noise components that share variance with the neural signal of interest, leading to signal loss (false negatives).
• The Gap: While previous studies (e.g., Moia et al., 2021) tested modeling decisions on breath-hold tasks (which evoke systemic vascular responses), these strategies had not been rigorously evaluated on motor tasks, which evoke localized neural activation and often involve more complex, task-correlated motion patterns.
• The Question: How do different implementations of ME-ICA nuisance regressors (Aggressive vs. Moderate vs. Conservative) affect activation metrics, test-retest reliability, and signal retention in motor-task fMRI, particularly in populations with high motion (e.g., Multiple Sclerosis) or low-signal tasks (e.g., shoulder/ankle movements)?

2. Methodology

Datasets and Participants

The study analyzed 56 fMRI runs from 21 participants across four distinct datasets:

1. HealthyHand: Healthy participants performing hand-grasp tasks (Low motion, High signal).
2. HealthyShoulder: Healthy participants performing shoulder-abduction tasks (Variable motion, Low signal).
3. MSHand: Multiple Sclerosis (MS) patients performing hand-grasp tasks (High task-correlated motion).
4. MSFoot: MS patients performing ankle-flexion tasks (Low signal, variable motion).
   Note: The HealthyHand dataset included a "Limited" (still) and "Amplified" (intentional motion) condition to simulate varying motion levels.

Three Modeling Approaches

The researchers compared three methods for incorporating rejected ME-ICA components (classified as noise) as nuisance regressors in subject-level General Linear Models (GLM):

1. Aggressive: All rejected ME-ICA components are included as regressors. (Standard approach).
2. Moderate: Rejected components are included unless they are significantly correlated with the task regressor (collinearity check: $P < 0.05, R^2 > 0.5$). If correlated, they are excluded to prevent removing task signal.
3. Conservative: All rejected components are orthogonalized to the core model (task, motion, polynomials, CO2) and to all accepted (BOLD-weighted) ME-ICA components before inclusion. This ensures rejected components do not share variance with the task or accepted signal, theoretically preserving the task signal while removing orthogonal noise.

Analysis Pipeline

• Preprocessing: Multi-echo data was optimally combined (ME-OC) using tedana. No spatial smoothing was applied.
• Denoising: Nuisance regressors included motion parameters, derivatives, end-tidal CO2, and the three ME-ICA strategies.
• Metrics:
  • Data Quality: Temporal Signal-to-Noise Ratio (tSNR) in motor regions.
  • Activation: Median t-statistics and percent of activated voxels in the contralateral Primary Motor Cortex (M1) and ipsilateral cerebellum.
  • Reliability: Spatial correlation between repeated runs (test-retest similarity).
• Grouping: Scans were stratified by Signal Intensity (High: Hand; Low: Shoulder/Foot) and Motion Characteristics (Low/High/Extreme FD and Task-Motion correlation $R^2$).

3. Key Results

Data Quality (tSNR)

• The Conservative model resulted in the lowest tSNR across all datasets compared to Aggressive and Moderate models.
• Reasoning: Orthogonalization prevents the model from explaining as much total variance (leaving more residual noise), whereas Aggressive models explain more variance (including task-related variance) by removing it.

Activation Metrics (t-statistics and Extent)

• High Motion / Low Signal: The Conservative model significantly outperformed the Aggressive model.
  • In low-signal tasks (shoulder, ankle) with high motion, the Aggressive model often resulted in zero or minimal activation (signal loss).
  • The Conservative model preserved significant activation and higher t-statistics in these difficult scenarios.
• High Signal / Low Motion: Differences were less pronounced, though Conservative still showed trends toward higher activation extent.
• Extreme Motion: In scans with extreme motion (FD > 1mm), all models struggled to detect significant activation, suggesting a threshold where denoising cannot recover the signal.

Test-Retest Reliability

• Low Signal Group: The Conservative model demonstrated significantly higher spatial correlation between repeated runs compared to Aggressive and Moderate models across all motion categories.
• High Signal Group: Results were mixed; the Moderate model sometimes showed better correlation in high-motion/low-correlation scenarios, but the Conservative model excelled in extreme motion groups.

Visual Evidence

• Group-level maps showed that the Aggressive model failed to detect precentral gyrus activation in the HealthyShoulder dataset, whereas the Conservative model successfully visualized these regions.
• However, the Conservative model retained more global noise and "striping" artifacts in extra-motor regions compared to the Aggressive model.

4. Key Contributions

1. Validation of Orthogonalization: The study provides empirical evidence that orthogonalizing rejected ME-ICA components (the Conservative approach) is superior for preserving motor-task signal in scenarios with high task-correlated motion and low expected neural activation.
2. Context-Specific Recommendations: It challenges the "one-size-fits-all" Aggressive denoising approach, demonstrating that while Aggressive works well for high-signal tasks (hand grasp) in healthy populations, it is detrimental for low-signal tasks (shoulder/ankle) or clinical populations (MS).
3. Nuanced Interpretation of tSNR: The authors highlight that higher tSNR (achieved by Aggressive models) does not equate to better detection of the task signal; in fact, high tSNR in this context often indicates the removal of the task signal itself.
4. Stratified Analysis Framework: The study introduces a framework for categorizing motor fMRI data by signal intensity and motion characteristics to guide denoising strategy selection.

5. Significance and Implications

• Clinical Applications: For studies involving patient populations (e.g., MS, stroke, children) who exhibit task-correlated motion, the Conservative model is recommended to prevent the loss of clinically relevant motor activation data.
• Task Selection: Researchers studying proximal limb movements (shoulder) or distal movements with lower cortical representation (ankle) should avoid aggressive denoising, as these tasks are more susceptible to signal loss.
• Methodological Guidance:
  • Use Conservative modeling for subject-level analysis in low-signal/high-motion datasets to maximize signal retention.
  • Use Moderate modeling if the analysis requires strict noise control in extra-motor regions, accepting a potential trade-off in motor signal sensitivity.
  • Use Aggressive modeling only for high-signal, low-motion datasets where noise removal is the primary concern.
• Future Directions: The paper suggests combining ME-ICA with other physiological denoising (RETROICOR) or anatomical correction methods to further mitigate the noise retention inherent in the Conservative approach.

Conclusion: The paper concludes that the Conservative ME-ICA modeling approach (orthogonalization) is the most robust strategy for motor-task fMRI when dealing with high task-correlated motion and low expected signal, effectively balancing the need to mitigate motion artifacts without unintentionally removing the neural signal of interest.