Fig. 2.

Machine learning was performed on 1 to 17 features derived from phase maps reconstructed from adherent cells’ holograms. Derived features from epithelial (GIE) and mesenchymal (HGF) cell types were used for training. For transfer learning, six PCs representing most of the variation in cell phase maps from two untrained cancer cell lines, MCF-7 and MDA-MB-231, were used for testing and to generate machine learning prediction scores as candidates for an EM score.