Adaptive smoothing based on Gaussian processes regression increases the sensitivity and specificity of fMRI data

Area of Science:

• Neuroimaging
• Machine Learning
• Statistical Analysis

Background:

• fMRI data preprocessing commonly uses spatial and temporal filtering to enhance statistical power.
• Conventional smoothing methods, like Gaussian filters, can reduce sensitivity by losing fine-scale neural activity structures.
• This tradeoff between sensitivity and specificity limits the precision of fMRI analyses.

Purpose of the Study:

• To introduce a novel Gaussian process (GP) regression-based smoothing method for single-subject fMRI data.
• To address the computational challenges of GP regression for high-dimensional fMRI datasets.
• To demonstrate GP regression as a viable alternative to conventional smoothing techniques in fMRI preprocessing pipelines.

Main Methods:

• Developed an efficient implementation of Gaussian process (GP) regression for fMRI data.
• Applied GP regression for adaptive, voxel-wise smoothing based on local neural activity patterns.
• Integrated GP smoothing as a direct replacement for traditional temporal and spatial filters in fMRI analysis.

Main Results:

• The proposed GP-based smoothing method demonstrated improved sensitivity and specificity compared to conventional smoothing strategies.
• Simulated and experimental fMRI data showed enhanced detection of neural activity patterns.
• The new GP implementation effectively handles the large dimensionality of fMRI data.

Conclusions:

• Gaussian process (GP) regression offers a superior approach to fMRI data smoothing, balancing sensitivity and specificity effectively.
• This method provides a more accurate representation of fine-scale neural activity.
• GP regression is a practical and powerful tool for advancing fMRI analysis.