Meeting Banner
Abstract #2667

Fast Diffusion-Guided QSM Using Graphical Processing Units

Owen L. Kaluza1, Amanda C. L. Ng2, 3, David K. Wright4, 5, Leigh A. Johnston, 56, John Grundy7, David G. Barnes2

1Monash e-Research Centre, Monash University, Clayton, Victoria, Australia; 2Monash Biomedical Imaging, Monash University, Clayton, Victoria, Australia; 3Department of Electrical & Electronic Engineering, The University of Melbourne, Parkville, Victoria, Australia; 4Centre for Neuroscience, The University of Melbourne, Parkville, Victoria, Australia; 5Florey Institute of Neuroscience and Mental Health, Parkville, Victoria, Australia; 6NeuroEngineering Laboratory, Dept. Electrical & Electronic Engineering, The University of Melbourne, Parkville, Victoria, Australia; 7Centre for Complex Software Systems and Services, Swinburne University of Technology, Hawthorn, Victoria, Australia

Diffusion-guided quantitative susceptibility mapping (QSM) is a new technique that promises improved mapping without the need for multiple-orientation (COSMOS) image acquisitions. However, the computation time for realistic image sizes on central-processing unit (CPU)-based supercomputers is prohibitively expensive. We have analysed the dQSM algorithm and developed an OpenCL-based implementation that runs on graphics processing unit (GPU)-based compute clusters. Our implementation yields identical results to the parallel CPU code, in drastically less time. Dual-GPU cluster nodes can compute the dQSM map 8 - 10 times faster when their GPUs are used compared to their multi-core CPUs. With this work, use of dQSM in research imaging facilities becomes practicable on quite modest computational facilities.

Keywords

acceleration access according achieve acquisition adaptation allocated application arranged audience beyond biomedical blue brain calculations capability central clinical clinically cluster coarse code completion complex complexity comprises computation computational compute computed computes conditions configurations constraints contemporary contrasts core cores council density deployed deployment derived despite diffusion discovery dual efficient electrical electronic elements engineering entirely environment every exact execution expect expected expensive extent facilities fast formal formulation funding future gene graphical guided hardware health human ideal immediately implementation implemented improved initial input introducing iteration iterations john just kernel kernels leaving loop loops mapping maps massive matrix measured medicine memory mental minutes modeled modest morphology mouse necessary necessitate node nodes numerical operates output overhead parallelism patterns performance post precision predisposing problem processing property provision quantitative realistic relatively reporting scalability sciences servers services society software solution speed spheres spherically statement studied sufficient supercomputer support susceptibility systems target theoretical timing today unit units utilizing validation water