Voxelmorph Learns to Pay Attention

This paper proposes a novel attention-based 3D image registration architecture to use in medical image registration. Previous registration methods have employed deep neural networks such as convolutional neural networks (CNNs) to extract the direct relationship between two pairs of image volumes. They do not pay extra attention to the particular regions of the images, as the medical image contexts might need. Our proposed architecture uses a Spatial Attention Module (SAM) and a deep VoxelMorph-based U-net convolutional neural network to guarantee the necessary attention to the particular regions of medical images in the registration process. The experimental results on a 4D cardiac Computed Tomography scans (Cardiac CT scans) dataset show that our registration method has outperformed state-of-the-art registration algorithms and reached better accuracy scores in both unsupervised and semi-supervised training approaches.