Presenting CODS (Cell Organelle Dynamic Simulation)

Abstract

The interactions between organelles within living cells have an effect on the health of the organism [1]. One way to study this behavior is to develop algorithms which can automatically detect and track objects in microscopy video data. A large amount of data is needed to train algorithms to identify particular behaviors in subcellular machinery. If the case of segmentation is considered, videos of these organelles are difficult to label because of the inability to locate objects with precision in microscopy videos. This thesis introduces CODS (Cell Organelle Dynamic Simulation), a solution for generating as much automatically labeled data as needed for use in several varieties of machine learning. CODS is able to work in parallel to produce dynamic complex simulations of mitochondria performing behaviors such as fission, fusion, and kiss-and-run in various behavior profiles. This lays the groundwork to easily build further complex interactions and to add organelle types to perform further development and study. CODS offers the ability to easily manipulate dozens of parameters relating to the type, number, and behavior of organelles desired. This includes the parameters of the microscope being simulated, providing the opportunity to produce data which works to mimic a variety of situations. I discuss the myriad uses of the CODS simulation engine and show that output from CODS is good enough to train a video segmentation algorithm to detect and track mitochondria through time in real microscopy data.