Mitochondria Detection in Electron Microscopy Images
click on the images to jump to some results.
Electron microscopy (EM) is key to mapping the
morphology of neural structures. Recent techniques, such as Focused Ion
Beam Scanning Electron Microscopy (FIB-SEM), can now deliver image
stacks at the nanometer resolution in all three dimensions. Such stacks
show very fine structures that are critical to unlocking new insights
into brain function but are still mostly analyzed by hand, which can
require months of tedious labor. As a result, the vast majority of this
very high quality data goes unused. Furthermore, although they contain
tens of millions of voxels, these stacks span volumes smaller than
10x10x10 μm3, which presents less than a billionth of the
volume of the entire brain. If it is ever to be mapped in its entirety,
automation will be required.
Our goal is to propose a fully automated approach that can segment large
EM datasets by using sophisticated cues that capture global shape and texture.
Automatic segmentation
of mitochondria also called "cellular power plants" from a FIB-SEM image stack.
The blue overlay indicates areas segmented as mitochondria by grouping
individual voxels into supervoxels, computing shape and intensity
features from these supervoxels, and feeding them to a graph-cut algorithm.
The voxels that were painted blue in the above example
are carved out of the volume to produce a 3D reconstruction of the mitochondria.
Note how elongated some of them are.
Interactive results
This Flash application shows the results of our proposed 2D segmentation method,
the results of competing methods, and walks through the intermediate
steps of our approach.
