Prior image-guided cone-beam computed tomography augmentation from under-sampled projections using a convolutional neural network