TOP-GAN: Label-Free Cancer Cell Classification Using Deep Learning with a Small Training Set

•We classified optical path delay maps of healthy and cancer cells without staining.•Simple machine learning classifiers gave limited accuracy.•We applied many deep learning methods in spite of the very small training sets.•One of them is TOP-GAN, transferring of pre-trained generative adversarial network.•Applied for the first time for medical imaging, and yielded more than 90% accuracy.