Volumetric 3D Reconstruction of Tissues in C. elegans: Part 1
This is the first of a series of posts about a research project that I am undertaking at the Sloan Kettering Institute. First, a little background. I’m a Computer Vision and Graphics Engineer at the Sloan Kettering Institute, a research division of the Memorial Sloan Kettering Cancer Center located in the Rockefeller Research Laboratories on the Upper East Side of New York City. I work in the Zhirong Bao Developmental Biology Lab, which researches the genome and neural development using the model organism C. elegans. I focus primarily on computer vision software for segmenting and tracking cellular division and lineaging during embryogenesis. If you deal with similar data, or are interested in these tools, check out StarryNite and AceTree, our software for tracking and naming cells. In addition, we maintain 3D graphics software for presenting the volumetric data we generate, called WormGUIDES. This software is a 4D atlas of C. elegans embryogenesis, providing a spatiotemporal representation of this process and the data involved.
We are undertaking a project to develop a pipeline that will automate the volumetric 3D reconstruction of tissues in a C. elegans embryo as they develop. The tissues we are interested in are the pharynx, muscle and hypodermis. Some of these can be seen at left using GFP and DsRed (Green and Red Flourescent Protein)
Volumetric 3D Reconstruction consists of generating a three dimensional model of an entity given a set of two dimensional images of that entity. Our data is in the form of 3D stacks of 2D images that are generated at points during embryogenesis (approximately 400 time points during the development of about 600 cells). We’ve developed a pipeline for segmenting this data into nuclear positions and applying shape models to generate a complete set of nuclear positions and cell sizes throughout development (see A hybrid blob-slice model for accurate and efficient detection of fluorescence labeled nuclei in 3D).The next step is to take this data and generate models of the tissues that are formed from these cells. What this process looks like is illustrated below. Here, a series of 2D images, analogous to our 2D slice images of nuclei, are used to generate a 3D model, like a pharynx model in C. elegans.
In the realm of cellular microscopy data, tissue model generation is an open problem due to the variability of cell shapes within a single organism, volume of data in larger organisms, and the difficultly of generating clean, high quality images. As such, a variety of approaches must be attempted, and perhaps even different methods for different tissues employed. Approaches we intend to try include generating a convex hull from nuclear positions and other graph connectivity methods, looking more broadly at the dynamics of the tissues for segmentation as a whole unit, voronoi tessellations, and reimaging with a histone background to use nuclei as markers and their positions as seeds for masked membrane images to employ a watershed technique.
There are a number of heuristics which can be leveraged if the prospect of generating a tissue-agnostic pipeline becomes too lofty a goal. For example, the hypodermis establishes the basic body plan during embryogenesis, so active contours to find the boundaries of the embryo could prove effective. In addition, we might exploit the fact that hypodermal cells are only connected to their immediate neighbors, which could allow us to prune large sections of a tree of graph connectivity possibilities, or use a nearest neighbor algorithm (this would also address the challenge presented in the case of the hypodermis of a lack of information in the z direction since the tissue is so thin). In the case of the pharynx, it can be observed with the untrained human eye (if you know what cells to look at) that two parallel planes of cells elongate from their centroids in opposite directions, forming an ellipsoid-like shape. This suggests that we may be able to use a rough shape-model to generate this tissue model from nuclear positions, while using the same principle of proximal connectivity that is observed in the hypodermis. With the muscle, which forms along the walls of the inner tube of the organism, these same heuristics apply.
I’m starting this series as a way to track and remember a research project which my lab is very excited about. In addition, we have a deep commitment to open-source software and value the collaboration that we so often rely on in our line of work. I hope that those who follow this series (or even just read a single article) become inspired to embark on similar projects, give their advice and share their resources.
