A Research On Acceleration Of Non-Rigid Registration On Medical Image By Using GPU Hardware

Registration is a key factor in medical image processing. A registration procedure can joint CT or MRI images to ultrasound image in real time which could make up the defects of effect of different imaging devices. The registration could also help doctors to make up their mind on human bones orthodontic surgery preparation. This paper propose a method of Thin Plate Spline (TPS) registration by GPU acceleration on medical image, including the following parts:the usefulness of medical image registration, GPU hardware logical, the acceleration algorithm of matrix operation, the fundamental principles of TPS medical image registration, and the programming of CUDA acceleration.The main point of our experiment is accelerating matrix multiplication. Our experimental platform is INTEL XEON E1230V2,8G RAM, GTX 460 display card. Optimization procedure includes accelerate memory access, bank conflict, source matrix recombination, etc. The elapsed time of our first CUDA matrix multiplication version is 0.3286s, compare to the final version is 0.025s, acceleration ratio is 13. The quantity of control points is also important to the computation, because the size of the matrix is depended on the quantity of control points. The experiment shows that the elapsed time of 2D picture registration is 0.0106s with 65 control points and 0.0044s with 3 control points on MATLAB, while on VB.net 2010 elapsed time is 3214s in 1Gbyte CT volume data. Comparing to CUDA method is 70s, the accelerate ratio is 46.In the case of usage of multi-core processor in image processing, the processing speed is nearly linear increasing. Without major change to the source code, upgrading the hardware to NVIDIA Tesla K80, or multi-GPU, multi-CPU system could rise the speed-up ratio to 200, implying that the real time registration is very promising.