A 3D/2D Coronary Artery Registration Method Based on Deep Reinforcement Learning

At present, vascular interventional surgery mainly depends on the conventional coronary angiography (CCA) and coronary computed tomography angiography (CCTA). However, due to various reasons such as different imaging principles and different operation methods, both CCA and CCTA imaging technologies can only be used separately. In order to integrate the two kinds of medical image information, medical image registration technologies based on artificial intelligence have emerged. In this paper, we propose a 3D/2D coronary artery registration method based on the combination of feature and deep reinforcement learning. Based on the centerline feature, we perform a projection transformation to unify the dimensions between two modes, then design a deep reinforcement learning algorithm, complete the mapping of the state space by convolutional neural network (CNN), discretize the actions of the agent, and use the Euclidean distance to define the reward function for fine coronary artery registration. The method is used for experiments with clinical CCA and CCTA data sets, and compared with other registration algorithms. The results show that the registration error rate of the proposed method is lower than that of other algorithms, it has strong accuracy and robustness, and can handle the registration problems of complex vascular segments.