Improving reliability of subject-level resting-state fMRI parcellation with shrinkage estimators

Amanda F. Mejia, Mary Beth Nebel, Haochang Shou, Ciprian M. Crainiceanu, James J. Pekar, Stewart H. Mostofsky, Brian Caffo, Martin A. LindquistNeuroImage 2015 · DOI

TL;DR

This paper introduces a novel "shrinkage" method to improve the reliability of brain parcellations derived from resting-state fMRI (rsfMRI) data. By borrowing information across subjects, this approach significantly enhances the consistency of individual brain maps, even with shorter scan times, making rsfMRI analysis more robust and reproducible.

What Problem Does This Paper Address?

Resting-state fMRI (rsfMRI) is a powerful tool for understanding brain function by analyzing spontaneous neural activity. A key application is "brain parcellation," which involves dividing the brain into distinct functional regions. These parcellations are crucial for defining "nodes" in brain network studies and for accurately comparing brain activity across individuals. Traditionally, parcellations were based on group-level data, averaging across many subjects to overcome the inherent noise in individual scans. However, there's a growing need for reliable subject-level parcellations, which are challenging to obtain due to the low signal-to-noise ratio (SNR) of rsfMRI and often limited scan durations (e.g., 5-10 minutes). This low SNR leads to noisy estimates of functional connectivity between brain regions, making it difficult to consistently identify the same functional areas for a given individual across different scanning sessions. The paper aims to address this by improving the reliability of the underlying connectivity estimates used for parcellation.

Methods

The core of the paper's contribution lies in applying "shrinkage estimators" to improve the reliability of functional connectivity measures, specifically inter-voxel correlation coefficients. Shrinkage estimators are a statistical technique that combines a noisy, subject-specific estimate with a more stable estimate derived from a larger group of subjects (the "group mean"). This "borrowing of strength" helps to reduce the impact of noise and improve the accuracy of the individual estimates.

The authors propose several shrinkage methods, differing primarily in how they estimate the "noise variance" (within-subject variability). These include:

  • Common Noise Variance: Assumes the noise variance is the same across all subjects.
  • Individual Noise Variance: Estimates noise variance separately for each subject.
  • Scaled Noise Variance: Starts with a common noise variance estimate and adjusts it with a subject-specific scaling factor.
  • Global Noise Variance: Assumes noise variance is constant across all subjects and voxel-pairs.

These shrinkage estimators are then used as input for a standard spectral clustering algorithm to generate the brain parcellations. The paper also explores performing shrinkage directly on correlation coefficients versus performing it on Fisher-transformed correlation coefficients, which are more statistically well-behaved.

Datasets

The study utilizes two types of datasets:

  1. Simulated Data: A synthetic dataset was generated to have a known "ground truth" parcellation. This allowed for precise evaluation of the accuracy of the shrinkage methods. The simulation varied parameters like the number of subjects, time series length, correlation strength, and between-subject variance.
  2. Real fMRI Data: A test-retest dataset of resting-state fMRI scans from 20 healthy adult volunteers was used. Two 7-minute scans were acquired from each subject on the same day. This data was crucial for evaluating the reproducibility of subject-level parcellations. The precentral gyrus (motor cortex) was chosen as the region of interest.

Key Insights and Ideas

  • Shrinkage is powerful for noisy data: The fundamental idea that borrowing information from a group can stabilize noisy individual measurements is powerfully demonstrated. This is particularly relevant for rsfMRI, where individual scans are inherently noisy.
  • Subject-level reliability is achievable with shorter scans: The paper shows that shrinkage can significantly improve the reliability of subject-level parcellations, making it feasible to obtain meaningful individual brain maps even with relatively short scan times (e.g., 7 minutes). This is a major step towards more practical and accessible rsfMRI analysis.
  • Global noise variance estimator is robust and practical: The "global noise variance" estimator, which assumes a constant noise variance across all subjects and voxel-pairs, performed very well, especially in settings where test-retest data is not available. This makes it a highly practical choice for many research scenarios.
  • Shrinking raw correlations can be better: Surprisingly, applying shrinkage directly to the raw correlation coefficients, rather than to their Fisher-transformed counterparts (which are statistically more appropriate), often led to better results in terms of reducing Mean Squared Error (MSE) and improving parcellation reliability. This suggests that the evaluation metric (e.g., MSE on the correlation scale) plays a crucial role.
  • Shrinkage improves parcellation consistency: The improved connectivity estimates directly translate to more consistent and reproducible brain maps for individual subjects, even across different scanning sessions.

Key Results

  • Improved Test-Retest Reliability: Parcellations of the motor cortex generated using shrinkage estimates showed up to a 30% increase in test-retest reliability compared to those derived from raw correlation estimates.
  • Reduced MSE: Shrinkage significantly reduced the Mean Squared Error (MSE) of correlation estimates. For example, with single-session data, the median MSE reduction ranged from 69.9% to 73.9% depending on the noise variance estimation method.
  • Enhanced Dice Similarity: The Dice similarity coefficient, a measure of overlap between parcellations, increased substantially. For instance, with single-session data, shrinkage-based parcellations showed a 23.2% to 28.1% higher Dice coefficient compared to raw parcellations.
  • Performance across methods: While all shrinkage methods showed improvements, the global noise variance estimator often yielded the best results, particularly in single-session scenarios.
  • Simulation insights: Simulations confirmed that shrinkage benefits are more pronounced with shorter time series, lower within-cluster correlations, and higher between-subject variance.

Limitations and Future Work

  • Generalizability: The study primarily focused on the motor cortex (precentral gyrus) and a population of healthy adults. The authors acknowledge that the benefits of shrinkage might vary for different brain regions, different similarity metrics, and more diverse populations (e.g., children, elderly, or clinical groups). Future work should explore these variations.
  • Choice of Clustering Algorithm: The paper used spectral clustering but noted that the shrinkage approach is agnostic to the specific clustering method. Future research could investigate the impact of shrinkage on other clustering algorithms.
  • Subject-Specific Shrinkage Nuances: While subject-specific shrinkage (individual and scaled noise variance) was explored, the authors noted potential issues with preserving subject rankings when shrinkage is highly individualized. This warrants further investigation.
  • Longer Scans: While shrinkage is beneficial for short scans, the paper suggests that longer scans might still offer advantages, and future research should explore the interplay between shrinkage and longer scan durations.
  • Other Similarity Metrics: The methods were demonstrated with correlation coefficients but are applicable to other distance or similarity metrics. Further exploration with metrics like those used in Nebel et al. (2014) is suggested.
  • Implementation: The authors have made their methods available as MATLAB and R packages, facilitating future research.

Summary updated

BibTeX
@article{mejia2015improving,
  author = {Amanda F. Mejia and Mary Beth Nebel and Haochang Shou and Ciprian M. Crainiceanu and James J. Pekar and Stewart H. Mostofsky and Brian Caffo and Martin A. Lindquist},
  title = {{Improving reliability of subject-level resting-state fMRI parcellation with shrinkage estimators}},
  journal = {NeuroImage},
  year = {2015},
  doi = {10.1016/j.neuroimage.2015.02.042},
}
Figure