Autoencoder Features for Differentiation of Leukocytes Based on Digital Holographic Microscopy (DHM)

Stefan Röhrl, Matthias Ugele, Christian Klenk, Dominik Heim, Oliver Hayden, Klaus Diepold

Abstract: The differentiation and counting of leukocytes is essential for the diagnosis of leukemia. This work investigates the suitability of Deep Convolutional Autoencoders and Principal Component Analysis (PCA) to generate robust features from the 3D image data of a digital holographic microscope (DHM). The results show that the feature space is not trivially separable in both cases. A terminal classification by a Support Vector Machine (SVM) favors the uncorrelated PCA features.