Hierarchical Latent Space Clustering for Synthetic Cell Images

We propose an interactive learning approach to generate synthetic cell microscopy images. Synthetic data can play an important role in augmenting unbalanced or rare datasets when training neural networks for classification or feature detection. Recent advances in Variational Autoencoders and Generative Adversarial Networks enable clustering in the latent space to generate samples from different classes with distinct features. We demonstrate that repeated binary clustering of input images can be employed to derive a hierarchical tree where the leaves represent classes of images with distinct features. Instead of experts annotating individual cells in raw image data, they can review and classify a larger collection of synthetic cell images with similar appearance, also learning about the variation observed in different classes. Using images of dividing cells, we show it is possible to assign biologically meaningful labels to automatically derived classes. Our method helped to resolve distinct sub-classes of stages during mitotic cell division that are almost impossible to label unambiguously in the raw data. Our interactive learning approach not only promises reducing the time required to accurately label training data, but also augmenting and balancing training data sets in a principled way.