A FCN-based Unsupervised Learning Model for Deformable Chest CT Image Registration

Image registration is a fundamental technique for many automatic medical image analysis tasks, but it can be time-consuming, especially for deformable three-dimensional image registration. In this paper we propose a fast unsupervised learning method for deformable image registration using a fully convolutional network (FCN). The network directly learns to estimate a dense displacement vector field (DVF) from a pair of input images. A spatial transform layer then uses the DVF to warp the moving image to the fixed image. Different from supervised learning based image registration methods, the network is trained by maximization of a similarity metric between the fixed image and the warped moving image. Thus training does not require supervised information such as manually annotated or synthetic ground truth. We evaluate the proposed model on publicly available datasets of inspiration-expiration chest CT image pairs. The results demonstrate that the accuracy of the model is comparable to that of the conventional image registration while executing orders of magnitude faster.