Visualization in Connectomics - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 217

21.6 Microscale Connectivity
In contrast to light microscopy, which is limited in its resolution by the wavelength
of light, electron microscopy enables imaging of neuronal tissue at the nanometer
scale. Hence, electronmicroscopy is the only imagingmodality so far that can resolve
single synapses. However, the sample preparation and image acquisition in electron
microscopy is labor-intensive and time-consuming. As a consequence, the analysis
of the connectivity between single neurons has been limited to sparse analysis of sta-
tistical properties such as average synapse densities in different brain regions [ 20 ].
Little is known about the complete connectivity between single neurons. Information
about the individual strength of synapses or the number of connections between two
cells can have important implications for computational neuroanatomy and theoret-
ical analysis of neuronal networks [ 98 ].
Recently, significant progress has been made in the automation of ultra-thin serial
sectioning [ 36 ] and automatic image acquisition [ 21 , 52 ]. These techniques allow
neuroanatomists to acquire large datasets of multiple terabytes (TB) in size. With a
resolution of 5 nm per pixel, and a section thickness of 50 nm, one cubic millimeter of
brain tissue requires imaging of 20,000 sections with 40 gigapixels per image, leading
to an image volume of 800 TB. With data sets of this size new challenges emerge
for automatic computed analysis and visualization techniques. Important processing
tasks include demand-driven image stitching and alignment, cell segmentation and
3D reconstruction, as well as multi-scale visualization and multi-user interaction via
client server architectures.
Electron microscopy samples are typically densely stained. While in light
microscopy sparse staining is necessary to visually separate a cell of interest
from unstained background tissue (see Sect. 21.5 ), the fine resolution of electron
microscopy allows one to discriminate structures according to shape, size, and tex-
ture. Electron microscopy images are limited to gray scale and typically do not have
a uniform background. Instead, the background is noisy and highly variable, which
imposes an important challenge for the visualization of electron microscopy image
stacks. The image data cannot be visualized according to gray values alone, as the
densely stained tissue forms a nearly solid block. Instead, higher order features that
discriminate texture and shape, e.g., gradient histograms, are necessary to enhance
the visibility of different structures of interest in the visualization [ 42 ]. Ultimately,
full segmentation of the image data is necessary to allow the user visual inspection
of different biological structures, from small structures such as vesicles or mitochon-
dria to entire neuronal cells. Figure 21.4 shows example reconstructions of different
neuronal structures from electron microscopy images. A number of software pack-
ages have been developed to aid the user in manual segmentation of cell structures
in the images [ 14 , 29 , 37 ]. More recent semi-automatic methods greatly facilitate
this time-intensive process [ 16 , 76 , 77 , 91 ].
Progress has also been made on fully automatic segmentation of EM brain images
[ 39 , 44 , 47 , 48 , 101 , 103 ]. However, all methods developed so far require manual
interaction and inspection by users. Thus, visualization tools should not only provide
the ability to inspect the original EM data and the computed segmentations, but also