Optimal basis representations for the analysis of functional magnetic resonance imaging data

The research presented here focuses on the use of various basis representations to optimally analyze functional magnetic resonance imaging (fMRI) data. The nature of fMRI data is such that known properties of the signal and noise structure for a given experimental setting are very limited. A basis transformation converts data from one coordinate system to another. The impetus for applying different basis functions is that these various representations may emphasize different natural signal structures occurring in the data. The scope of the research includes denoising (improving signal to noise), classification (supervised learning), and data driven analysis (non-supervised learning) as specific applications of optimal basis selection. Evaluation of techniques utilizes receiver operator characteristic (ROC) analysis, which is a common tool in fMRI for comparing data analysis techniques. In addition, recently proposed model performance metrics such as model reproducibility and prediction accuracy are demonstrated for assessing our supervised learning implementations.