Amat led the effort to develop an efficient solution. His first priority was to reduce the complexity of the data. His strategy was to first cluster the voxels (essentially three-dimensional pixels) that make up each image into larger units called supervoxels. Using a supervoxel as the smallest unit reduces an image's complexity a thousand-fold, Keller says.
Next, the program searches for ellipsoid shapes among groups of connected supervoxels, which it recognizes as cell nuclei. Once a cluster of supervoxels is identified as a cell nucleus, the computer uses that information to find the nucleus again in subsequent images. High-speed microscopy captures the images quickly enough that a single cell can't migrate very far from frame to frame. "We take advantage of that situation and use the solution from one time point as the starting point for the next point," Keller says.
"With this fairly fast, simple approach, we can solve easy cases fairly efficiently," Keller says. Those cases make up about 95 percent of the data. "In harder cases, where we might have mistakes, we use heavier machinery."
He explains that in instances where cells are harder to track because image quality is poor or cells are crowded, for example the computer draws on additional information. "We look at what all the cells in that neighborhood do a little bit into the future and a little bit into the past," Keller explains. Informative patterns usually emerge from that contextual information. The strategy takes more computing power than the initial tactics. "We don't want to do it for all the cells," Keller says. "But we try to crack these hard cases by gathering more information and making better informed decisions."
All of these steps can be carried out as quickly as images are acquir
|Contact: Jim Keeley|
Howard Hughes Medical Institute