Improving the accuracy of single-trial fMRI response estimates using GLMsingle

Jacob S. Prince, Ian Charest, Jan W. Kurzawski, John A. Pyles, Michael J. Tarr, Kendrick KayeLife 2022 · DOI

TL;DR

This paper introduces GLMsingle, a new MATLAB/Python toolbox that significantly improves the accuracy of estimating brain responses to individual fMRI trials. By combining custom HRF fitting, data-driven denoising, and voxel-wise regularization, GLMsingle enhances the reliability of fMRI data, leading to better downstream analyses like representational similarity and decoding.

What Problem Does This Paper Address?

Modern neuroscience is generating massive fMRI datasets with thousands of stimuli per subject. A major challenge with these "condition-rich" designs is achieving a sufficient signal-to-noise ratio (SNR) because each stimulus is presented only a few times. This low SNR makes it difficult to accurately estimate the brain's response to each individual trial, which in turn hinders our ability to understand neural representations and computations. Traditional General Linear Model (GLM) approaches often struggle with this, especially when trials are close together in time, leading to noisy and correlated response estimates.

Methods

GLMsingle tackles this problem by integrating three key techniques into a single, automated pipeline:

  1. Custom Hemodynamic Response Function (HRF) Fitting: Instead of using a single, generic HRF for all voxels, GLMsingle uses a library of 20 candidate HRFs. For each voxel, it identifies the HRF that best fits the observed data, allowing for more accurate modeling of individual voxel responses.
  2. Data-Driven Denoising (GLMdenoise adaptation): GLMsingle adapts the GLMdenoise technique to identify and remove noise components from the fMRI signal. It uses principal component analysis on voxels deemed "noise" (low experimental signal) to derive nuisance regressors, which are then incorporated into the GLM. Cross-validation is used to determine the optimal number of these regressors.
  3. Voxel-wise Ridge Regression: To stabilize estimates for closely spaced trials, GLMsingle applies fractional ridge regression. This technique regularizes the beta estimates on a voxel-by-voxel basis, with the degree of regularization determined by cross-validation. This helps to reduce noise inflation caused by correlated predictors.

The entire process is automated and designed to be efficient, even for large datasets.

Datasets

The authors evaluated GLMsingle using three datasets:

  • Natural Scenes Dataset (NSD): A large-scale fMRI dataset with thousands of natural images presented to subjects. This dataset was used for developing and testing the HRF library and for evaluating GLMsingle's performance.
  • BOLD5000: Another large-scale fMRI dataset featuring scene images, used to test the generalizability of GLMsingle across different subjects, scanners, and experimental designs.
  • StudyForrest (music-listening component): A smaller-scale, non-visual (auditory) dataset used to demonstrate GLMsingle's applicability to diverse experimental paradigms.

Key Insights and Ideas

  • Individual voxel HRFs matter: The assumption of a single HRF for all voxels is a simplification. GLMsingle's approach of finding the best-fit HRF for each voxel significantly improves data quality.
  • Denoising is crucial for condition-rich designs: With many trials and limited repetitions, separating signal from noise is paramount. GLMsingle's data-driven denoising effectively removes unwanted variability.
  • Regularization is key for temporal overlap: When trials are close in time, BOLD responses overlap, making estimation difficult. Ridge regression helps disentangle these overlapping responses.
  • GLMsingle benefits downstream analyses: The improvements in single-trial estimates aren't just statistical curiosities; they translate into tangible gains for higher-level analyses like representational similarity and decoding.

Key Results

  • Improved Test-Retest Reliability: GLMsingle (b4) substantially improved voxel test-retest reliability compared to a baseline GLM (b1) across both NSD and BOLD5000 datasets. For example, in NSD, the mean difference in reliability between b4 and b1 was positive across subjects, indicating better consistency.
  • Reduced Temporal Autocorrelation: GLMsingle significantly reduced temporal autocorrelation between beta maps of nearby trials, especially in rapid event-related designs. This indicates better disentanglement of responses to consecutive stimuli.
  • Enhanced Between-Subject Similarity: Representational Similarity Analysis (RSA) showed a consistent strengthening of representational correspondence between subjects (within and across datasets) when using GLMsingle betas. This suggests that GLMsingle reveals more shared neural representations.
  • Boosted MVPA Decoding Accuracy: Image-level MVPA decoding accuracy (predicting image identity from brain activity) approximately tripled in NSD and doubled in BOLD5000 when using GLMsingle betas compared to the baseline GLM.

Limitations and Future Work

  • Requirement for Repeated Conditions: GLMsingle relies on cross-validation to determine hyperparameters, which necessitates at least some repeated trials in the dataset. It is not suitable for resting-state data or experiments with purely continuous events without discrete repetitions.
  • Hyperparameter Estimation Noise: While data-driven, hyperparameter estimation is still subject to noise, and the minimum data requirements for accurate estimation are an open question.
  • Potential for Bias: The authors acknowledge that denoising and regularization can introduce bias, particularly in scenarios where hypotheses specifically depend on subtle differences in responses to very close trials. However, they argue that for most use cases, the benefits outweigh this risk.
  • Future Work: The authors suggest that future studies could refine or expand the HRF library, potentially by deriving it from larger subject pools or restricting estimation to individual subjects. Further investigation into minimum data requirements for hyperparameter estimation is also warranted.

Summary updated

Abstract

Advances in artificial intelligence have inspired a paradigm shift in human neuroscience, yielding large-scale functional magnetic resonance imaging (fMRI) datasets that provide high-resolution brain responses to thousands of naturalistic visual stimuli. Because such experiments necessarily involve brief stimulus durations and few repetitions of each stimulus, achieving sufficient signal-to-noise ratio can be a major challenge. We address this challenge by introducing <i>GLMsingle</i>, a scalable, user-friendly toolbox available in MATLAB and Python that enables accurate estimation of single-trial fMRI responses (glmsingle.org). Requiring only fMRI time-series data and a design matrix as inputs, GLMsingle integrates three techniques for improving the accuracy of trial-wise general linear model (GLM) beta estimates. First, for each voxel, a custom hemodynamic response function (HRF) is identified from a library of candidate functions. Second, cross-validation is used to derive a set of noise regressors from voxels unrelated to the experiment. Third, to improve the stability of beta estimates for closely spaced trials, betas are regularized on a voxel-wise basis using ridge regression. Applying GLMsingle to the Natural Scenes Dataset and BOLD5000, we find that GLMsingle substantially improves the reliability of beta estimates across visually-responsive cortex in all subjects. Comparable improvements in reliability are also observed in a smaller-scale auditory dataset from the StudyForrest experiment. These improvements translate into tangible benefits for higher-level analyses relevant to systems and cognitive neuroscience. We demonstrate that GLMsingle: (i) helps decorrelate response estimates between trials nearby in time; (ii) enhances representational similarity between subjects within and across datasets; and (iii) boosts one-versus-many decoding of visual stimuli. GLMsingle is a publicly available tool that can significantly improve the quality of past, present, and future neuroimaging datasets sampling brain activity across many experimental conditions.
BibTeX
@inproceedings{prince2022improving,
  author = {Jacob S. Prince and Ian Charest and Jan W. Kurzawski and John A. Pyles and Michael J. Tarr and Kendrick Kay},
  title = {{Improving the accuracy of single-trial fMRI response estimates using GLMsingle}},
  booktitle = {eLife},
  year = {2022},
  doi = {10.7554/elife.77599},
}
Figure