Visualizing Cell Motility Patterns from Time Lapse Videos with Interactive 2D Maps Generated with Deep Autoencoders

Abstract

Cell motility, the ability of cells to move, is crucial in a wide range of biological processes; for instance, in cancer, it is directly related to metastasis. However, it is a complex phenomenon which is not well-understood yet, and studies are mainly done by human observation, which is subjective and error-prone. We intend to provide an automated mechanism to analyze the movement patterns that occur in in-vitro cell cultures, which can be registered by time lapse microscopy. Our approach, which is still a work in progress, utilizes an interactive 2D map that organizes motility patterns based on their similarity, enabling exploratory analysis. We extract the velocity fields that represent the cell displacements between consecutive frames and use a deep convolutional autoencoder to project a characterization of short video sequences of smaller parts of the original videos into a 10D latent space. The samples (small videos) are visualized in a 2D map using the Uniform Manifold Approximation and Projection (UMAP). The possibilities and extent of our method are showcased through a small interactive application that allows to explore all the types of cell motility patterns present in the training videos on a 2D map.