Computer Vision News - May 2020

3 Summary Deep GANCS MRI 5 the relation map between an initial aliased image and the gold-standard image. An example is the Automap, by B.Zhu et al published in Nature in 2018. A Fully Connected network followed by a convolutional neural network (CNN) directly mapped the k-space data to the image domain for robust reconstruction. These approaches do significantly speed up the reconstruction, but they also suffer from aliasing and blurring effects. This happens because of the adoption of a pixel- wise l1/l2 cost for training. Structured artefacts and high-frequency texture details are unavoidable. Lately, Generative Adversarial Networks (GANs) seem to be successful in modelling low-dimensional manifolds and generating high-dimensional data such as natural images. The generated images look perceptually appealing and have been used with tasks such as image inpainting, style transfer and visual manipulation. Although successful in those areas, there has not been enough research on the local image restoration and for the correction of the aliasing artefacts in MRI reconstruction. The two main components of the approach are to ensure the trained manifold contains plausible MR images and that the points on the manifold are data consistent. The former is approached by GANs which have already been successful in estimating the prior distributions for images. The latter by performing back-and- forth mappings between the layer the projects onto a set of data-consistent images and the manifold.