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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Data Science

Background:

• Large-scale functional magnetic resonance imaging (fMRI) datasets offer high-resolution brain responses to naturalistic stimuli.
• Challenges in fMRI include brief stimulus durations and few repetitions, impacting signal-to-noise ratio and accurate estimation of brain responses.

Purpose of the Study:

• To introduce GLMsingle, a scalable and user-friendly toolbox for accurate estimation of single-trial fMRI responses.
• To address the challenge of low signal-to-noise ratio in fMRI experiments with brief stimuli and limited repetitions.

Main Methods:

• GLMsingle integrates three techniques: custom hemodynamic response function (HRF) identification, cross-validation for noise regressor derivation, and voxel-wise regularization using ridge regression.
• The toolbox requires only fMRI time-series data and a design matrix as input.

Main Results:

• GLMsingle substantially improves the reliability of beta estimates in visually-responsive cortex across subjects in large-scale datasets (Natural Scenes Dataset, BOLD5000).
• Comparable reliability improvements were observed in an auditory dataset (StudyForrest experiment).
• GLMsingle enhances representational similarity between subjects and improves decoding of visual stimuli.

Conclusions:

• GLMsingle significantly improves the quality of fMRI data analysis for past, present, and future neuroimaging studies.
• The tool benefits systems and cognitive neuroscience by enhancing higher-level analyses, decorrelating response estimates, and boosting stimulus decoding.