Currently, non-rigid image registration algorithms are too computationally intensive to use in time-critical applications. Existing implementations that focus on speed typically address this by either parallelization on GPU-hardware, or by introducing methodically novel techniques into CPU-oriented algorithms. Stochastic gradient descent (SGD) optimization and variations thereof have proven to drastically reduce the computational burden for CPU-based image registration, but have not been successfully applied in GPU hardware due to its stochastic nature. This paper proposes 1) NiftyRegSGD, a SGD optimization for the GPU-based image registration tool NiftyReg, 2) random chunk sampler, a new random sampling strategy that better utilizes the memory bandwidth of GPU hardware. Experiments have been performed on 3D lung CT data of 19 patients, which compared NiftyRegSGD (with and without random chunk sampler) with CPU-based elastix Fast Adaptive SGD (FASGD) and NiftyReg. The registration runtime was 21.5s, 4.4s and 2.8s for elastix-FASGD, NiftyRegSGD without, and NiftyRegSGD with random chunk sampling, respectively, while similar accuracy was obtained. Our method is publicly available at https://github.com/SuperElastix/NiftyRegSGD.

@en

Currently, non-rigid image registration algorithms are too time intensive to use in time-critical applications. To solve this problem, stochastic gradient descent (SGD) has been implemented in image registration. But, SGD depends on manual step size selection which is dicult and time consuming. To avoid such manual selection, SGD has been improved further by using adaptive stochastic gradient descent (ASGD) and fast adaptive stochastic gradient descent (FASGD) to select an optimal step size automatically. Although FASGD has reduced the computation time drastically, non-rigid registration still cannot be used in time critical applications. So far, a serial implementation of FASGD has been tested on CPU
architecture in elastix toolbox. Thus, a parallel implementation of SGD can be a possible solution to this problem.
The work proposed in this thesis implemented a NiftyReg toolbox extension to graphic processing units (GPUs), divided into two methods. First, NiftyReg2, a possible optimization of the current NiftyReg. Second, NiftyRegSGD, a high performance implementation of SGD on the GPU framework of NiftyReg. A novel sampling strategy, random chunk sampling is also proposed which is tailored to the GPU architecture. Random chunk sampling is an optimization to utilize memory bandwidth of GPU eectively to increase the throughput of CUDA kernels.
Experiments have been performed on 3D lung CT data of 19 patients, which compared NiftyRegSGD (with and without random chunk sampler) with CPU-based elastix FASGD and NiftyReg. The registration runtime was 21.5s, 13.02s, 4.4s and 2.8s for elastix-FASGD, NiftyReg2, NiftyRegSGD without, and NiftyRegSGD with random chunk sampling, respectively, while similar accuracy was obtained. Thus, proposed GPU based non-rigid registration can be used for a time critical application with further extensions. The abstract which discusses the work done during this thesis has been accepted for publication in the medical
imaging conference of the Society of Photographic Instrumentation Engineers (SPIE).