This thesis addresses the possibility of applying the compressed sensing (CS) framework to Functional Magnetic Resonance Imaging (fMRI) acquisition. The fMRI is one of the non-invasive neuroimaging technique that allows the brain activity to be captured and analysed in a living body. One disadvantage of fMRI is the trade-off between the spatial and temporal resolution of the data. To keep the experiments within a reasonable length of time, the current acquisition technique sacrifices the spatial resolution in favour of the temporal resolution. It is possible to improve this trade-off using compressed sensing.
The main contribution of this thesis is to propose a novel reconstruction method, named Referenced Compressed Sensing, which exploits the redundancy between a signal and a correlated reference by using their distance as an objective function. The compressed video sequences reconstructed using Referenced CS have at least 50% higher in terms of Peak Signal-to-Noise Ratio (PSNR) compared to state-of-the-art conventional reconstruction methods. This thesis also addresses two issues related to Referenced CS. Firstly, the relationship between the reference and the reconstruction performance is studied. To maintain the high-quality references, the Running Gaussian Average (RGA) reference estimator is proposed. The reconstructed results have at least 3dB better PSNR performance with the use of RGA references. Secondly, the Referenced CS with Least Squares is proposed. This study shows that by incorporating the correlated reference, it is possible to perform a linear reconstruction as opposed to the iterative reconstruction commonly used in CS. This approach gives at least 19% improvement in PSNR compared to the state of the art, while reduces the computation time by at most 1200 times.
The proposed method is applied to the fMRI data. This study shows that, using the same amount of samples, the data reconstructed using Referenced CS has higher resolution than the conventional acquisition technique and has on average 50% higher PSNR than state-of-the-art reconstructions. Lastly, to enhance the feature of interest in the fMRI data, the baseline independent (BI) analysis is proposed. Using the BI analysis shows up to 25% improvement in the accuracy of the Referenced CS feature.
