基于双通道级联注意力网络的医学图像配准

Medical image registration based on dual-stream cascaded attention network

为提高卷积神经网络在医学图像上的配准性能,提出一种双通道级联注意力网络用于医学图像配准。针对浮动图像和固定图像,用两个卷积神经网络对配准场进行估计;用配准场级联策略提高配准场变形估计性能;在配准场估计过程中引入注意力机制用于自动学习和优化注意力特征并分配特征权重,进一步加强特征表达能力,提高配准性能。通过对脑部图像和肺部图像的配准实验分析,验证了该方法的有效性和准确性。

To improve the registration performance of convolutional neural networks on medical images,a dual-stream cascaded attention network was proposed for medical image registration.For the moving image and the fixed image,two convolutional neural networks were used to estimate the registration field.The registration field cascade strategy was used to improve the estimation performance of the registration field deformation.The attention mechanism was introduced in the registration field estimation process to automatically learn and optimize attention features,feature weights were assigned to further strengthen feature expression capabilities and improve registration performance.Through the registration experiment analysis of the brain images and lung images,the validity and accuracy of the method are verified.