Research On Applying GPU Parallel Computation To Medical Image Processing

Medical image processing and analyzing techniques play an important role inclinical diagnosis and medical researching, which mainly contain medical imageenhancement and image segmentation as well as image registration and reconstruction.With massive data and complex algorithms, the speed of medical image processing inpersonal computers can not meet the demands of practical applications, so how toprocess massive data accurately and efficiently is a problem to be resolved urgently.As a high-level parallel stream processor, GPU (Graphics Processing Unit) has amuch more powerful floating-point computing ability compared to CPU. With thespread of GPU and CUDA (Compute Unified Device Architecture), GPU is no longerrestricted to the traditional image rendering, but plays a more and more important rolein general computing. Using the heterogeneous programming of CPU and GPUprovides strong compute ability and improves the speed of medical image processing.Based on the studies of medical image enhancement and image segmentationhome and abroad, accelerated algorithms using GPU are realized. Firstly, find out dataparallelism of the time-consuming homogeneity similarity based filter algorithm,optimize the memory access and speed up the data access, compared with the Matlabresults using an interface file, at last achieve a speed-up of hundreds while maintainsdata accuracy. Secondly, among all the algorithms of image segmentation, Cannyoperator is accelerated with GPU as an example. Find out data parallelism of eachstep, choose different optimization strategy to obtain the largest speed-up. Lastly, usecompute profiler to balance the GPU resource. Using compute profiler providespotential optimize paths and resource limits. According to different applications,choose different optimizing directions and balance the GPU resource to achieve thelargest speed-up.Compared to the CPU results, accelerated homogeneity similarity based filterachieves approximately77x speed-up, while accelerated Canny algorithm achievesapproximately9x speed-up. CUDA programming shortens the executing time ofthree-dimensional image processing largely, solves the speed bottleneck of medicalimage processing, and it can be applied to applications which demand high speed.