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.

In this paper we propose a method to solve nonrigid image registration through a learning approach, instead of via iterative optimization of a predefined dissimilarity metric. We design a Convolutional Neural Network (CNN) architecture that, in contrast to all other work, directly estimates the displacement vector field (DVF) from a pair of input images. The proposed RegNet is trained using a large set of artificially generated DVFs, does not explicitly define a dissimilarity metric, and integrates image content at multiple scales to equip the network with contextual information. At testing time nonrigid registration is performed in a single shot, in contrast to current iterative methods. We tested RegNet on 3D chest CT follow-up data. The results show that the accuracy of RegNet is on par with a conventional B-spline registration, for anatomy within the capture range. Training RegNet with artificially generated DVFs is therefore a promising approach for obtaining good results on real clinical data, thereby greatly simplifying the training problem. Deformable image registration can therefore be successfully casted as a learning problem.