Improving delay and strength maps derived from resting-state fMRI using PCA-based denoising and group data from the HCP dataset.

Resting-state functional magnetic resonance imaging (rs-fMRI) analyses use correlations in low-frequency "noise" to infer neuronal connectivity. A significant fraction of this oscillatory signal is non-neuronal, and is therefore a confound for rs-fMRI; however, we have shown that these signals carry valuable information, which can aid in clinical diagnosis and tracking recovery in stroke and moyamoya patients. Specifically, we have developed a method (RIPTiDe) that extracts blood arrival time delay (blood flow) and signal strength maps (perfusion) from BOLD data, yielding critical insight into vascular structure and function. In this study, we demonstrate a principal component analysis (PCA)-based method to denoise these rs-fMRI derived delay and strength maps to enhance signal-to-noise ratio without requiring prior knowledge of the noise percentage. We used group data from the Human Connectome Project (HCP) dataset, and conducted spectral analysis on the BOLD derived maps to identify the structural components' locations using both a naïve, and an optimized approach; we removed noise components by back-projecting only a subset of images to the original space. To assess signal reliability, we calculated the intraclass correlation coefficients (ICC) of the voxelwise parameters before and after noise removal within each subject. Mean ICC values were calculated for each projection dimension. The dimension achieving the highest ICC was selected as the signal-to-noise separation threshold for denoising. This optimized method for selecting the number of PCA components to retain increases the average ICC values of the delay and strength maps by 250 % and 108 %, respectively.