A subspace method for reconstruction of time-series fMRI images from sparse data

Abstract

Functional magnetic resonance imaging (fMRI) is a powerful imaging modality commonly used to study brain functions. It utilizes the difference in the oxygen content of brain tissues over time to produce functional connectivity maps, or visualizations of brain regions activated when a subject performs a task. As such, it provides invaluable insight into the inner workings of the brain and how disease changes its functionality.

Unfortunately, fMRI sees limited use outside of research settings due to its long data acquisition time and the large amount of data required to generate useful results. Methods which reduce the amount of required data while maintaining acceptable results become necessary to allow the availability of fMRI in clinical settings.

This thesis presents a novel method to reconstruct high-resolution spatiotemporal fMRI image sequences given highly undersampled data. It introduces a model which combines low-rank subspaces with prior information to produce results which outperform other state-of-the-art reconstruction techniques such as SENSE. A comparison of image quality and fMRI analyses over a wide variety of datasets shows the superiority of the proposed method.