Accelerations of structural and functional brain connectivity using heterogeneous computing

In this thesis, the main aims are to accelerate algorithms in diffusion tractography and functional MRI connectivity analysis, by mapping them on parallel architectures. Diffusion Tractography is an algorithm for studying micro-structure of brain white matter (WM), and functional MRI is a neuroimaging procedure to explore the time series of brain activities. Both of these algorithms are widely applied in neuroscience researches. Diffusion-weighted (DW) magnetic resonance (MR) images can be processed to yield orientation information on the underlying anisotropic distribution of elongated fibers, such as in the white matter of the brain. As MR field strengths increase and scan protocols are improved, the spatial and angular resolution that can be achieved have reached the point where traditional diffusion tensor imaging (DTI) methods are being replaced in favour of non-tensor methods, so as to allow for multiple fiber directions per image voxel. Once the fiber distribution is known, the generation of whole streamlines - stochastic representations of white matter fiber bundles - while computationally dense, is intrinsically parallel and is eminently suited to acceleration with the contemporary graphical processing unit~(GPU). Here, we report the design, implementation, validation and performance analysis of two different parallel mappings of the standard probabilistic tracking algorithm that is applied to DW MR images represented in spherical harmonics. We achieve a 10x speedup on a commodity GPU, compared to the standard multi-core CPU implementation, while recovering the expected distribution of the streamlines. Our parallel implementation scales well across different hardware and problem sizes. The best rate achieved is one million streamlines computed in less than 20 seconds. The work on the accelerated sliding-window-based spatial ICA components tracking on functional MRI time-series data is also main part of our project, which is a more flexible approach of studying brain functional dynamics than traditional Region-of-Interests (ROI)-based analysis. In order to perform real-time ICA processes, we are investigating into the parallel mapping and implementations of the FastICA algorithm on GPU(s) to achieve less than 2 seconds (i.e. the sampling speed of the current scanning for fMRI sequences). This may open new possibilities of performing real-time neurofeedback studies and intraoperative image-guided neurosurgery.