Adversarial Learning for Deformable Image Registration: Application to 3D Ultrasound Image Fusion

We present an adversarial learning algorithm for deep-learning-based deformable image registration (DIR) and apply to 3D liver ultrasound image fusion. We consider DIR as a parametric optimization model that aims to find displacement field of deformation. We propose an adversarial learning framework inspired by generative adversarial network (GAN) to predict the displacement field without ground-truth spatial transformation. We use convolutional neural network (CNN) and a spatial transform layer as registration network to generate the registered image. Similarity metrics of image intensity and vessel masks are used as loss function for the training. We also optimize a discrimination network to measure the divergence between the registered image and the fixed image. Feedback from the discrimination network can guide the registration network for more accurate and realistic deformation. Moreover, we incorporate an autoencoder network to extract anatomical features from vessel masks as shape regularization. Our approach is end-to-end, only requires image pair as input in registration tasks. Experiments show that the proposed method outperforms state-of-the-art deep-learning-based methods.