Multi-Resolution Medical Image Registration with Dynamic Convolution

Recent deep learning medical image registration (DLMIR) methods based on static convolutional neural network (CNN) have achieved advanced registration performance with feature representation and learning ability of CNN. To further improve the registration accuracy, the common practice is to increase the depth or width of the network, which increases the computational overhead. To address this problem, in this paper, we utilize dynamic convolution instead of static convolution to perform the registration task, where Dynamic convolution kernels are formed by the nonlinear aggregation of several parallel and input-dependent convolution kernels. We evaluate the proposed method on a public Magnetic Resonance (MR) brain scan dataset. The results demonstrate that the proposed method outperforms existing methods in terms of registration accuracy without increasing the depth and width.