Impact of Censoring on the Quality of Cortical Parcellations and Personalized TMS Targets

This study demonstrates that using lenient censoring strategies to retain high-motion data in resting-state fMRI yields higher-quality individualized cortical parcellations and personalized TMS targets than strict censoring, suggesting that patients with motion artifacts may not require re-scanning for precision neuroimaging applications.

The Gist

The Big Picture: The "Shaky Hand" Problem

Imagine you are trying to take a high-resolution photo of a beautiful landscape, but your hand is shaking. In the world of brain imaging (fMRI), this "shaky hand" is called head motion. Even tiny movements can blur the picture and create "noise" that makes it look like different parts of the brain are talking to each other when they aren't.

For years, scientists have had a standard rule for dealing with this: If the photo is too shaky, throw it away. They would delete the blurry parts of the scan or even cancel the whole session if the person moved too much. They thought this was the only way to get a clean picture.

This paper asks a simple question: Is throwing away the shaky data actually making our brain maps worse?

The Experiment: Building a "Perfect" Map

To answer this, the researchers needed a "Gold Standard" or a "Perfect Map" to compare against. Since they couldn't get a perfect photo in real life, they used a clever trick:

1. The "Ground Truth": They took data from 19 people who had spent hours in the MRI scanner (a lot of data). They took the cleanest, most still parts of those hours and combined them to create a "Perfect Map" for each person. This is their reference point.
2. The Simulation: They then took the messy parts of the data (the shaky parts) and chopped them up into tiny 1-minute pieces. They mixed and matched these pieces to create fake 10-minute scans that ranged from "slightly shaky" to "very shaky."
3. The Test: They tried to build brain maps from these fake, shaky scans using two different strategies:
   • The "Strict" Strategy: Delete any part of the scan that is shaky. If the whole scan is too shaky, throw the whole thing away.
   • The "Lenient" Strategy: Keep the shaky parts, but try to fix the blur mathematically. Don't throw anything away.

The Results: The "Salad" Analogy

Here is where the results get surprising.

Imagine you are making a salad. You have a bowl of fresh lettuce (good data) and some wilted leaves (shaky data).

• The Strict Strategy says: "If there is even one wilted leaf, throw the whole bowl away and start over."
• The Lenient Strategy says: "Keep the bowl, pick out the really bad leaves, but keep the slightly wilted ones because they still have nutrients."

What the study found:

1. Throwing data away hurts you: When they used the "Strict" strategy, the brain maps they created were actually further away from the "Perfect Map" than when they kept the data. By being too strict, they threw away useful information along with the noise. It's like throwing away the whole salad because one leaf is a little brown; you lose the nutrients in the rest of the leaves.
2. Keeping the data helps: The "Lenient" strategy (keeping the slightly shaky data and just fixing it) produced brain maps that were much closer to the "Perfect Map."
3. The "High Motion" Surprise: Even when they took two very shaky scans and kept them all (Lenient), the resulting brain map was often better than taking one clean scan and one shaky scan, then throwing the shaky one away (Strict).

Why Does This Happen?

Think of your brain's activity like a unique fingerprint. Even when you move your head, your brain is still doing its unique thing.

• Strict Censoring is like trying to clean a fingerprint by scrubbing it so hard that you erase the unique ridges. You remove the "noise" (the shake), but you also accidentally remove the "signal" (the unique brain pattern).
• Lenient Censoring is like gently wiping the glass. You remove the smudges, but you keep the fingerprint intact.

What This Means for Patients (The TMS Connection)

This isn't just about science; it's about real people needing treatment.

• The Current Dilemma: A patient comes in for a brain scan to plan a treatment called TMS (Transcranial Magnetic Stimulation) for depression. If the patient fidgets or moves too much, the clinic might say, "Sorry, your scan is ruined. Come back next week for a new one." This costs money, causes stress, and delays treatment.
• The New Solution: This paper suggests that clinics don't need to send the patient home. Even if the patient moved a lot, the doctors can use "Lenient Censoring" to fix the data. They can still get a highly accurate map of the brain and a precise target for the TMS treatment without needing a second scan.

The Bottom Line

For a long time, we thought "less motion = better data" meant "throw away the motion." This paper shows that "more data = better data," even if some of it is a little messy.

By being a little more forgiving (lenient) with the data and keeping the "shaky" parts, we can actually see the brain more clearly than by being too strict and throwing everything away. It's a win for accuracy, a win for patients, and a win for saving time and money.

Technical Summary

1. Problem Statement

Head motion in resting-state functional MRI (rs-fMRI) systematically biases functional connectivity (FC) estimates. The standard mitigation strategy involves censoring (removing) high-motion volumes and discarding entire high-motion scanning runs. However, this approach presents two critical issues:

1. Sample Bias: Discarding high-motion runs disproportionately excludes specific demographics (e.g., males, non-White individuals, lower socioeconomic backgrounds), reducing the representativeness of large-scale studies.
2. Signal Loss: Motion exists on a continuum. Stringent censoring may discard volumes containing meaningful signal alongside noise, potentially degrading the quality of individual-specific brain maps.
3. Clinical Dilemma: In personalized Transcranial Magnetic Stimulation (TMS), if a patient's scan contains high motion, clinicians often face the choice of discarding data (requiring costly re-scans) or using degraded data. The trade-off between strict data cleaning and retaining data for individualized targets remains unclear.

2. Methodology

The study utilized a "ground-truth" simulation framework leveraging precision-fMRI datasets (datasets with hours of scanning per participant).

• Datasets: 50 densely sampled healthy adults from 8 public datasets (e.g., MSC, Kirby, MyConnectome, NSD). After filtering for sufficient low-motion data to define a ground truth, 19 participants were used for the main two-run analysis.
• Ground-Truth Definition:
  • For each participant, the cleanest data (≥1 hour of post-censored data across ≥3 sessions) were selected using strict criteria (FD > 0.08 mm, DVARS > 50).
  • These clean data served as the "silver-standard" ground truth for individual-specific cortical parcellations and personalized TMS targets.
• Simulation of Motion:
  • Remaining non-ground-truth data were divided into 1-minute windows.
  • These windows were randomly recombined to create simulated 10-minute runs with varying levels of motion contamination (categorized by the percentage of frames that would be censored under strict thresholds: 0–25%, 25–50%, 50–75%, 75–100%).
  • Simulated sessions consisted of either one or two 10-minute runs.
• Censoring Strategies Tested:
  • Strict Censoring: FD threshold = 0.08 mm, DVARS = 50. High-motion runs (<5 mins remaining) were discarded.
  • Lenient/No Censoring: No volume removal or high thresholds (FD = 0.5 mm, DVARS = 100), retaining all runs.
  • Intermediate: Various combinations of FD and DVARS thresholds.
• Downstream Metrics:
  • Parcellation Quality: Measured by the Dice coefficient between the simulated parcellation and the ground-truth parcellation (using the Multi-Session Hierarchical Bayesian Model, MS-HBM).
  • TMS Target Quality: Measured by the Euclidean distance between the simulated TMS target (using a tree-based algorithm for depression/anxiety) and the ground-truth target.
• Statistical Analysis: Linear Mixed-Effects (LME) models and paired t-tests were used to compare strategies across motion bins, accounting for participant identity and overlapping data windows.

3. Key Contributions

• Re-evaluation of Censoring: The study challenges the dogma that strict censoring is always superior for individual-level analyses, demonstrating that aggressive removal of data can degrade individual-specific signal.
• Ground-Truth Simulation: A robust framework for evaluating preprocessing pipelines by simulating motion levels against a high-quality individual baseline, moving beyond group-level QC-FC correlations.
• Clinical Protocol Optimization: Provides evidence-based guidelines for handling high-motion scans in clinical TMS settings, potentially eliminating the need for re-scanning.

4. Key Results

• Motion Degrades Quality: As expected, higher motion levels generally led to parcellations and TMS targets that deviated further from the ground truth.
• Lenient > Strict: At any given motion level, lenient censoring (or no censoring) consistently produced higher quality parcellations and TMS targets than strict censoring.
  • Parcellation: No censoring yielded statistically better Dice coefficients ($p = 1.4 \times 10^{-5}$) than strict censoring.
  • TMS Targets: No censoring yielded significantly lower Euclidean distances (better accuracy) than strict censoring ($p = 8.2 \times 10^{-4}$).
• High-Motion vs. Mixed-Motion:
  • Parcellations and TMS targets derived from two high-motion runs with no censoring were comparable to or superior to those derived from mixed-motion sessions where the high-motion run was discarded and the low-motion run was strictly censored.
• Optimal Strategy:
  • Lenient censoring (e.g., FD = 0.5 mm, DVARS = 100) without discarding runs achieved performance close to the "noise ceiling" (the best possible performance for that specific motion level).
  • Strict censoring was consistently the worst-performing strategy for individualized targets.
• Robustness: Results held across different TMS algorithms (tree-based vs. cone), different motion metrics (FD vs. FDrms), and different session durations (1-run vs. 2-run).

5. Significance and Implications

• Clinical Practice: For patients undergoing personalized TMS, clinicians should retain high-motion runs and apply lenient censoring rather than discarding data and re-scanning. This reduces costs, patient burden, and delays in treatment.
• Research Integrity: In precision neuroimaging, retaining data via lenient censoring preserves individual-specific information that strict censoring removes. This is crucial for studies focusing on trait-level individual differences rather than group averages.
• Reconciling Conflicts: The authors reconcile these findings with prior literature (which advocates strict censoring) by noting that previous studies focused on removing across-individual correlations between motion and FC (QC-FC). However, for individual-specific targets, strict censoring removes meaningful trait-related variance, thereby degrading the specific signal needed for personalized medicine.
• Limitations: The study was conducted on healthy adults with relatively low average motion. While simulations covered high-motion ranges (up to ~1.14 mm FD), caution is advised for populations with extreme motion exceeding these simulated bounds.

Conclusion: The study concludes that lenient censoring is the optimal strategy for generating high-quality individual-specific cortical parcellations and personalized TMS targets, offering a practical solution to the motion artifact problem in precision neuroimaging.