Self-supervised Learning via Inter-modal Reconstruction and Feature Projection Networks for Label-Efficient 3D-to-2D Segmentation.