Deformable Medical Image Registration Using a Randomly-Initialized CNN as Regularization Prior

We present deformable unsupervised medical image registration using a\nrandomly-initialized deep convolutional neural network (CNN) as regularization\nprior. Conventional registration methods predict a transformation by minimizing\ndissimilarities between an image pair. The minimization is usually regularized\nwith manually engineered priors, which limits the potential of the\nregistration. By learning transformation priors from a large dataset, CNNs have\nachieved great success in deformable registration. However, learned methods are\nrestricted to domain-specific data and the required amounts of medical data are\ndifficult to obtain. Our approach uses the idea of deep image priors to combine\nconvolutional networks with conventional registration methods based on manually\nengineered priors. The proposed method is applied to brain MRI scans. We show\nthat our approach registers image pairs with state-of-the-art accuracy by\nproviding dense, pixel-wise correspondence maps. It does not rely on prior\ntraining and is therefore not limited to a specific image domain.\n