Artificially-generated consolidations and balanced augmentation increase performance of U-net for lung parenchyma segmentation on MR images