Knife-Edge Scanning Microscopy: High-Throughput Imaging and Analysis of Massive Volumes of Biological Microstructures - High-Throughput Image Reconstruction and Analysis - page 14

immediately available (according to the developers of ATLUM, imaging such a vol-
ume would take hundreds of years). Thus, a more reasonable approach is to have
the ultrathin section library containing the chips to serve as the database itself.
In the future, robotic random access SEM imaging will be used to do on-demand
imaging of objects of interest.

Summary and Comparison
Even though the methods above are unified under the common theme of phys-
ical sectioning, the resolution and the typical volume they can handle all differ,
putting them in a relatively complementary role with respect to each other. In other
words, these methods cannot be ranked on an absolute scale since there are rela-
tive advantages and disadvantages to each method. Table 2.1 provides a summary
comparison of these microscopy methods.
Finally, we wish to emphasize that no matter what physical volume these
methods deal with (ranging from 100 3 mm 3 upto1cm 3 ), the resulting volume data
can exceed several TBs, thus posing serious challenges to computational analysis,
and high-performance computing methods could help address these challenges.

2.1.2 Volumetric Data Analysis Methods
Once the volume data are obtained from physical sectioning microscopy, the next
task is to extract objects of interest from the data set. This is a nontrivial task due
to distortions, noise, and artifacts resulting from the cutting process (often due
to vibrations known as chatter ). Furthermore, the difficulty in automation for this
kind of data is dramatically increased by the density and the huge number of objects
in the microscopy data. This is unlike volume data from medical imaging (magnetic
resonance imaging or computer-aided tomography) where object count and density
are both low, for which various automated methods exist (e.g., National Library
of Medicine's Insight Toolkit, http://www.itk.org/).
In this

section, we will review three main approaches taken in 3D

reconstruction.

Image Registration, Segmentation, and Reconstruction
A traditional approach for 3D reconstruction is to consider the volume data as
a stack of images, while processing one image at a time. This is often necessary
because the z -axis resolution is so much lower than the lateral resolution for most

Table 2.1 Summary Comparison
Method Resol. (x,y) Resol. (z) Volume Modality Time

1 cm 3 All-optical hist. 0.5 mm 1 mm Fluorescence ∼ 900 hours

0.3--0.6 mm . --1mm1 c 3 KESM Bright field, ∼ 100 hours

fluorescence*

∼ 100 3 mm 3

Array ∼ 0.2 mm 0.05--0.2 mm Fluorescence, N/A

tomography EM

500 3 mm 3 SBF-SEM ∼ 0.01 mm ∼ 0.03 mm ∼ EM N/A

∼ 2.15 3 mm 3 ATLUM ∼ 0.01 mm0.05 mm EM N/A

*Expected in the near future.