Algebraic Approach to Medical Image Restoration on the GPU
Todor Mitev · Trinity College Digital Repository (Trinity College, Hartford CT) · 2012
Computed tomography (CT) scans have recently become of paramount importance in medicine. They allow doctors to diagnose patients with internal body problems without ever having to touch scalpels or use expensive probes. Furthermore, they drastically decrease the amount of time necessary to diagnose many illnesses, and do not require any recovery time. The importance of CT scans has led numerous researchers to consider the algebraic approach to medical image restoration, but none have attempted to examine the benefits of utilizing a GPU’s raw computing power. Tackling this problem on the GPU is reasonable because the ART algorithm allows for a high amount of parallel computations, which GPUs are particularly well-suited for. This project builds on Peter Toft’s efforts into creating a sequential implementation of the ART algorithm. It modifies Toft’s original implementation as to exploit its parallelizability and execute it on the GPU. This approach allows for the simultaneous execution of multiple instructions, instead of relying on a sequential computation. It requires some modifications to the original algorithm, which necessitate more computations overall, but this drawback is offset by the fact those computations are executed simultaneously. Moreover, the parallel implementation utilizes sparse matrices to conserve space and execution time. The outcome of the project is a parallel version of the ART algorithm and a precise measurement of how much time can be saved by running the algorithm with various configurations on a large dataset.