Extending Local Canonical Correlation Analysis to Handle General Linear Contrasts for fMRI Data

Table 1

Mean square errors (MSEs) of estimated coefficients for different methods. To define different spatial patterns of activations, 100000 randomly shaped activations within a 3 × 3 grid of pixels having a size of 2 to 9 pixels were generated. The corresponding time courses for the activated voxels were simulated to be linear combinations of the 4 random temporal regressors with random amplitudes. Different levels of noise were introduced by resampling 3 × 3 patches of resting-state fMRI data. The mean square errors between the originally simulated amplitudes of regressors and estimated ones are shown. The cCCA-RG method achieves more than 25% less MSE compared to GLM-NS. The GLM-GS method is worse than GLM-NS due to the small and irregularly defined activation patterns.

	Δ𝛽^21	Δ𝛽^22	Δ𝛽^23	Δ𝛽^24
GLM-NS	0.1331	0.1837	0.2200	0.1488
GLM-GS	0.1627	0.2047	0.2336	0.1769
cCCA-RG	0.0979	0.1339	0.1538	0.1091