3D Medical Image Registration Based on Spatial Attention

Learning-based image registration methods have confirmed the ability of neural network and achieved impressive performance. Recently, the convolutional neural network model based on encoding and decoding structure has shown great advantages in the task of medical image registration. In this work, an unsupervised end-to-end registration method based on spatial attention is proposed for deformable 3D medical images. Specifically, we use multi-scale feature fusion to reduce the difference between low-level features and high-level features. And spatial attention is introduced to enhance the salient areas of feature map and the interpretability of convolutional neural network. At the same time, we use dilated convolution to solve the problem of information loss in the process of increasing the receptive field, which facilitates the accuracy improvement. We evaluated the proposed method on four datasets OASIS, ADNI, ABIDE and ADHD200. The experimental results show that our method achieves promising results compared with the popular method SyN.