Time-segmented and contrast-neutral motion correction using an end-to-end deep learning approach