High Performance OpenCL Implementation of Medical Image Processing Algorithms [CP]

Current X-ray machines use lower radiation doses which introduces noise to the output images. Therefore such systems need to enhance the image and reduce the noise via different algorithms to provide the best possible output. In addition, it is crucial to accelerate these image processing algorithms as the output is intended to be a real time video (uoroscopy). Such systems are used for example in surgeries for implants or other medical examinations and there is a need to provide constant performance, otherwise they may lead to injuries or fatalities due to latency issues. Currently, such systems often rely on server PCs to implement the image processing chains. Since PC hardware needs to be replaced regularly during the lifetime of an X-ray machine, this increases the maintenance cost as well as the overall cost of the machine significantly. Therefore, we need to provide a framework that would allow us to develop the algorithm only once and then enable us to port it to a new platform, while the performance is ensured. In order to do so, a high performance framework solution was investigated. A number of alternative solutions were investigated and the most attractive framework was selected to be the Open Computing Language (OpenCL). OpenCL provides the means to develop the image processing algorithm once and port it to different platforms, changing only the target platform from the OpenCL API. During this thesis exploration we were able to redevelop a high quality algorithm provided by Philips Healthcare from a Matlab model to OpenCL in an optimal time period, while we investigated portability and performance. We first developed a tool chain that enables transformation from Matlab to OpenCL. Furthermore, 11 image processing kernels which constitute the algorithm were developed in OpenCL performing a speedup of up to 150x in some cases. We were able to run the algorithm on three different hardware platforms using the same OpenCL kernels and achieve a speedup up to of 36x compared to the baseline implementation in C.Electrical Engineering, Mathematics and Computer ScienceQuantum Engineerin