Parallel Feature Extraction - High-Throughput Image Reconstruction and Analysis - page 157

Figure 6.13 (a) The slice data visualized with a 3D graphics package. A coarse wire-frame model
is used. (b) The slice data visualized with a fine wire-frame model. (c) The slice data visualized with
a continuous surface.
Jain et al. [20] use an alternate approach to process brain slice data, and
segment the image pixels into two types of classes: those that fall within a cell and
those that fall outside. Human experts generate suitable training data by labeling
different image regions according to these classes. A nonlinear filtering step is first
applied to reduce noise. A convolutional network is trained using back-propagation
to perform such a classification automatically.
Fiala [15] has created a tool for facilitating the viewing and processing of
serial slice data. Features provided include alignment, region growing, and volume
visualization.
Smith [36] reviews current techniques for neural circuit reconstruction. Smith
observes that ''progress in electron microscopy-based circuit analysis will depend
heavily on the development of schemes for robust and efficient automated segmen-
tation.'' Promising emerging directions are the use of machine learning techniques,
and the use of complementary imaging techniques such as immunofluorescence.
6.6 Conclusion
The use of large 3D image datasets is growing rapidly in the biological sciences,
driven by problems such as neural circuit reconstruction through serial slice data.
In this chapter, we reviewed several basic 3D image processing algorithms that
are useful in handling such serial slice data. Some of the algorithms considered
were median and morphological filtering, and 3D connected component anal-
ysis. We also examined the problem of contour extraction between successive
slices.