Highly accelerated MR parametric mapping by undersampling the k-space and reducing the contrast number simultaneously with deep learning.