Dealing with Lack of Training Data for Convolutional Neural Networks: The Case of Digital Pathology