using a k-space acquisition property representing a k-space sampling pattern and its timing for generating a motion-corrupted image having a motion artifact as caused by a first initialising type of motion pattern such that the motion artifact is defined as function of a feature matrix representing motion induced corrupted encodings and the motion-corrupted image;

initializing at least one feature map representing a convolutional neural network's convolutional or multiplicative kernels with values of the feature matrix;

training the initialized convolutional neural network (CNN) with training images, the training images being corrupted by a second training motion pattern pattern to obtain motion artifacts;

obtaining a motion artifact in an input image using the trained CNN.