IDC-Net: Multi-stage Registration Network Using Intensity Adjustment, Dual-Stream and Cost Volume

We propose a Multi-stage Registration Network Using Intensity Adjustment, Dual-Stream and Cost Volume (IDC-Net) for large deformation diffeomorphic image registration. Unlike recent deep learning-based registration approaches, such as VoxelMorph, computes a registration field with the same scale from a pair of images by using a single-stream encoder-decoder network, we design a dual-stream architecture with intensity adjustment able to compute multi-resolution deformation fields from convolutional feature pyramids. IDC-Net is composed of an intensity adjustment network (IAN) and a dual-stream based multi-stage registration network with cost volume (DC-Net). The cost volume embedded dual-stream registration module is proposed to capture the correlation between two images and predict multi-scale registration fields, having strong deep representation ability for deformation estimation. The intensity adjustment network is designed to obtain a pair of images with similar intensity distribution to reduce the influence of intensity differences on the registration. IAN and DC-Net promote each other through a cooperative mechanism, which refines the registration fields gradually in a coarse-to-fine manner via sequential warping, and enable IDC-Net with the capability for handling large deformations and keeping diffeomorphism between two images. We conduct experiments on 3D brain MRI and liver CT scans, and the results show that the proposed method outperforms other state-of-art methods by a significant margin.