A cascaded registration network RCINet with segmentation mask

Traditional deformable registration methods achieve brilliant results and show strong theoretical support but are computational intensive since they optimize each image pair’s objective function. Recently, supervised learning methods have facilitated fast registration. However, it requires ground truth and does not guarantee a diffeomorphism registration. This paper proposes a new unsupervised learning method Recursive Cascaded Network with a segmentation mask, for two-dimensional medical image registration. Different from the original cascaded network, the network framework into two parts. The first section obtains a pair of image rolls and uses the registration sub-network to predict the deformation vector field from the moving image to the fixed image. The second part introduces anatomical segmentation into the network during training, makes full use of the auxiliary information of the volume, adds an autoencoder to encode the anatomical segmentation, and incorporates it into the learning process of the model in the form of constraints. The local and global ideas are combined to ensure the deformation field’s rationality and improve the distribution. The most important thing is that we propose a formula for calculating the cascaded network’s deformation field used in the test stage to evaluate the relationship between the registration accuracy and the deformation field’s effectiveness. Our experiments show that the system has a better registration effect and less information loss than the current state-of-the-art method. Simultaneously, the cascade method’s accuracy is an improvement at a certain number of layers, and the increase in accuracy needs to sacrifice the effectiveness of the deformation field.