Hybrid supervised and self-supervised deep learning for quantitative mapping from weighted images using low-resolution labels