Optimal spatial regularisation of autocorrelation estimates in fMRI analysis.

Area of Science:

• Neuroimaging
• Statistical Modeling
• Signal Processing

Background:

• Temporal autocorrelations in fMRI residuals bias statistical testing within the General Linear Model (GLM).
• Prewhitening strategies aim to correct this bias by whitening the residual signal.
• Accurate estimation of autocorrelation structure is crucial but challenging due to noisy fMRI data.

Purpose of the Study:

• To systematically investigate the effect of smoothing kernel width on prewhitening accuracy in fMRI.
• To introduce an optimal method for selecting the bandwidth for spatial smoothing of autocorrelation estimates.
• To evaluate various aspects of the prewhitening strategy, including autocorrelation estimation and spatial regularization.

Main Methods:

• Investigated the impact of smoothing kernel width on prewhitening performance.
• Developed and evaluated a method for optimal bandwidth selection for spatial smoothing.
• Analyzed the influence of autocorrelation estimate type (biased/unbiased), accuracy, spatial regularization, and autoregressive model order.
• Validated the proposed method using both synthetic and real fMRI datasets.

Main Results:

• Spatial smoothing of autocorrelation estimates significantly reduces their variance, leading to improved prewhitening.
• The proposed method provides an "optimal" bandwidth selection, enhancing correction accuracy.
• The study systematically characterized the trade-offs associated with different prewhitening strategy parameters.

Conclusions:

• Optimizing spatial smoothing and bandwidth selection is critical for accurate prewhitening in fMRI GLM analysis.
• The developed method offers a robust approach to improve statistical inference in neuroimaging.
• This work provides valuable insights for refining fMRI data analysis techniques.