Hierarchical learning approaches for improving label-free single-cell classification in holographic microscopy

Machine learning in combination with microscopy is a well-established paradigm for the identification of cells target (e.g. sick cells) or for the statistical study of cells’ populations. In general, the accuracy in classifying single cells depends on the selected imaging modality, i.e., the more informative it is, the more performant the classifier is. Here we show that the combination of machine learning and holographic microscopy is an effective tool to achieve the above goal, thus allowing higher classification performances if compared to other standard microscopies. Moreover, by exploiting a priori information about the samples to identify, the classification performance can be further increased. We demonstrate this paradigm for the differential diagnosis of hereditary anemias, in which RBCs, imaged by holographic microscopy, are used to predict firstly if an anemia occurs, then which type of anemia among five phenotypes.