While image-based motion estimation with Deep Learning has the advantage of an easier comprehension by a human observer, there are benefits to address the issue in k-space, as the distortion only affects echo trains locally; furthermore, Neural Networks can be designed to rely on the intrinsic k-space structure instead of image features. To our knowledge, these advantages have not been exploited so far. We show that fundamental Deep Neural Network techniques can be used for motion estimation in k-space, by examining different networks and hyperparameters on a simplified problem. We find suitable architectures for extracting 2D transformation parameters from under-sampled k-spaces for slice registration. This leads to a minimum residual of around 1.2 px/deg.