Baseline drift and physiological noise removal in high field fMRI data using kernel PCA

Baseline drift and physiological (cardiac and respiratory) fluctuations are among major sources contaminating blood oxygenation level dependent (BOLD) signals in high field functional magnetic resonance imaging (fMRI). Automatically detecting and removing them have been long-standing problems. We propose here a new method, utilizing kernel principal component analysis (KPCA) and frequency analysis, to detect and remove the noise from fMRI data. Differing from thermal noise, the main energy of baseline drift and physiological noise are characterized by the most significant kernel principal components that also contain information on brain structure. To maintain the details of brain anatomy, we filter the feature projections to the components that are found to contain significant baseline drift and physiological noise. This approach is different from most discriminant analysis-based denoising methods that remove insignificant or noisy components before the reconstruction. Experimental results show that the proposed method increases the BOLD contrast and the detection sensitivity of activated voxels.