Parallel Feature Extraction - High-Throughput Image Reconstruction and Analysis - page 143

CHAPTER 6
Parallel Feature Extraction
A. Ravishankar Rao and Guillermo A. Cecchi
6.1 Introduction
In order to motivate the need for parallel feature extraction techniques, we con-
sider the following problem, which has been receiving increasing attention. One of
the grand challenges identified by the National Academy of Engineering in 2008
(http://www.engineeringchallenges.org) was to reverse-engineer the brain. This is
indeed a complex problem, requiring insights from multiple disciplines such as
neuroscience, physics, electrical engineering, mathematics, and computer science
[1--41].
One approach suggested by Sporns et al. [38] and others is to build a detailed
structural model that captures the connectivity among the different neurons in
the brain. Having such a model would help in simulating interactions between
neural units, and understand signaling pathways such as feedback. As noted by
Kasthuri and Lichtman [23], a connectivity map provides an anatomical basis
for understanding how information is processed within a brain. An example of a
computational model of brain function built with such knowledge of anatomical
connectivity is the work of Seth et al. [31]. Shepherd et al. [35] have analyzed the
relationship between structure and function in cortical circuits.
We now consider the challenges in creating such a detailed structural model of
the brain. We examine a specific problem in order to make the discussion concrete.
However, many of the solution components are quite general, such as 3D image fil-
tering tasks, and can be applied to similar problems, irrespective of the data source.
6.2 Background
We consider the problem of imaging neural tissue. There are many imaging mech-
anisms are available, ranging from lower resolution optical imaging to high reso-
lution scanning electron microscopy (SEM). If SEM is used to image neural tissue,
1 petabyte of storage is required for imaging one cubic millimeter [23]. Due to
this large volume of data, researchers have to select a given resolution and then
determine the tissue mass that can be imaged. If a high resolution is used (e.g.,
SEM), then it is feasible to scan only certain neural structures (e.g., retina, hip-
pocampus, or the olfactory bulb). If lower resolution is used (e.g., the knife-edge
scanning microscope [26] or ultra-microscopy [11]), then it is possible to obtain
whole-brain maps. There are trade-offs to be made for working at each resolu-
tion level. With SEM techniques, it is possible to image individual synapses, thus
determining precise signaling connectivity at the neuronal level. With whole-brain
143