Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning Based Registration

Abstract

This paper introduces an unsupervised adversarial similarity network for image registration. Unlike existing deep learning registration frameworks, our approach does not require ground-truth deformations and specific similarity metrics. We connect a registration network and a discrimination network with a deformable transformation layer. The registration network is trained with feedback from the discrimination network, which is designed to judge whether a pair of registered images are sufficiently similar. Using adversarial training, the registration network is trained to predict deformations that are accurate enough to fool the discrimination network. Experiments on four brain MRI datasets indicate that our method yields registration performance that is promising in both accuracy and efficiency compared with state-of-the-art registration methods, including those based on deep learning.

Metric Learning for Image Registration

“…These approaches typically have no guarantees on spatial regularity or may not straightforwardly extend to 3D image volumes due to memory constraints. Alternative approaches have been proposed which can register 3D images [35,38,12,23,2,15] and assure diffeomorphisms [47,48]. In these approaches, costly numerical optimization is only required during training of the regression model.…”

“…One of the main conceptual downsides of current regression approaches is that they either explicitly encode regularity when computing the registration parameters to obtain the training data [47,48,35], impose regularity as part of the loss [23,2,15] to avoid ill-posedness, or use lowdimensional parameterizations to assure regularity [38,12]. Consequentially, these models do not estimate a deformation model from data, but instead impose it by choosing a regularizer.…”

Probabilistic Multilayer Regularization Network for Unsupervised 3D Brain Image Registration

“…We conducted the preprocessing steps for each 3D brain image, where these steps include brain extraction, voxel spacing re-sampling (1mm), affine spatial normalization, "Whitening" operation, and intensity normalization; see [4] for detail. To evaluate the registration performance, we register each unaligned image as well as its segmentation mask, and measure the overlap between the registered segmentation mask and the segmentation mask of reference image using a widely-used Dice metric; see [4,6,13] for the details of the Dice definition. In general, a larger Dice indicates a better 3D brain registration result.…”

“…Image registration aims to transform different images into one system with the matched imaging contents, which has significant applications in brain image analysis, including brain atlas creation [3], tumor growth monitoring [7] and multi-modality image fusion [5]. When we analyze a pair of brain images that were acquired from different sensors and viewpoints at different times, we need to transform one image (unaligned image) to another image (reference image) by establishing the anatomical correspondences [4,6,13]. The correspondence between the unaligned image x and the reference image y is usually formulated by a transformation function φ z , which is parametrized by a latent variable z.…”

Neural Multi-Scale Self-Supervised Registration for Echocardiogram Dense Tracking

“…Recent unsupervised deep learning based registrations, such as VoxelMorph [6], are facilitated by the spatial transformer network [21,12], and the VoxelMorph is further extended to diffeomorphic transformation and Bayesian framework [10]. Adversarial similarity network adds an extra discriminator and uses adversarial training to improve the unsupervised registration [11]. These purely learning based methods cannot be directly applied to ultrasound images for velocity estimation especially for vortex detection in cardiac blood flow because of great noise in echocardiogram, large velocity variations of cardiac blood flow and large amounts of missing and new blood within the ultrasound plane.…”