基于深度强化学习的多模态医学图像配准  

Robust Multimodal Medical Image Registration Using Deep Recurrent Reinforcement Learning

作　　者：姚明青 胡靖 Yao Mingqing;Hu Jing(College of Computer Sciences,Chengdu University of Information Technology,Chengdu 610225)

机构地区：[1]成都信息工程大学计算机学院,成都610225

出　　处：《计算机辅助设计与图形学学报》2020年第8期1236-1247,共12页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61602065);成都信息工程大学中青年学术带头人科研基金(J201608);成都信息工程大学科研基金(KYTZ201610)。

摘　　要：传统图像配准方法中,图像特征的表示和相似性测度的选择易受到人为因素的影响,不能准确地表征图像特征和配准图像的相似度,从而对配准结果产生较大误差.针对此问题,借助端到端的强化学习方法,对这2个部分进行隐式的表达,从而避免人工设计的缺陷.具体而言,设计了一个人工智能体模型,由策略网络和价值网络2部分组成,用以指导浮动图像朝着参考图像的方向正确移动,进而实现图像配准.提出使用异步表演者-评论家方法进行模型训练,以避免经验回放操作,降低模型训练对存储容量的要求并加快模型的收敛;同时提出一种奖赏函数,能够给予每个时间步上图像配准动作估计更为准确的奖励.此外,在测试阶段,使用了蒙特卡罗前向推理策略,进一步提高配准参数的准确性.在MR和CT的临床医学图像配准数据集上进行实验,与传统基于尺度不变性配准算法、基于深度学习配准算法等进行对照分析,实验结果表明,所提出的方法目标配准误差可以减少30%左右,同时能够更好地处理具有大幅度形变的配准问题.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.

关 键 词：图像配准 强化学习 表演者-评论家 奖励函数 前向推理 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]