U-Net cascaded with dilated convolution for medical image registration

Registration is a critical step in many biomedical image analysis tasks. In this work, we propose a novel cascaded 3D U-Net with dilated convolution for volumetric image registration. We implement a coarse-to-fine registration strategy that commonly used in traditional registration methods by cascading two 3D U-Net based networks. In the first registration stage, a 3D U-Net network is employed to predict the large and coarse displacements. In order to capture more precise pixel-level displacement field, in the second stage, we propose to substitute the down-sampling operations in the encoding stage of 3D U-Net by dilated convolutions. As result, the resolution of feature map can be maintained as well as its field-of-view and semantic information. In addition, to obtain a more parameter-efficient network that is essential for volumetric image analysis, we replace the 3D convolution kernel by 3D dense separated convolution (3DDSC) module. We evaluated the proposed method on two commonly used datasets, and the results demonstrated that our method could obtain better performance than state-of-the-art image registration algorithms with short execution time.