Robust Multimodal Medical Image Registration Using Deep Recurrent Reinforcement Learning

The key factors of a conventional image registration method lie in the choice of the suited feature representation and the similarity measure,and the inaccurate characterization of the similarity of image features and registered images will produce large errors in the registration results.Although elaborately designed,these two components are somewhat handcrafted using human knowledge.In this work,these two components are implicitly learned in an end-to-end manner via reinforcement learning.Specifically,we advocate an artificial agent model,which is composed of a combined policy and value network,to adjust the moving image toward the right direction.On one hand,we propose to train this model on asynchronous actor-critic to avoid memory replay,thereby reducing capacity requirement and accelerating model convergence.On the other hand,we propose a customized reward function to provide a more accurate rewarding measure for registration parameter prediction.Furthermore,we also propose a Monte Carlo look ahead inference in the testing stage to improve the registration capability.Quantitative evaluations on MR and CT image pairs from real clinical settings,compared with the traditional scale-based invariance registration algorithm and deep learning-based registration algorithm,demonstrate that the target registration error of the proposed method can be reduced by 30%,specifically in the case of large distortion.