Data Analysis Pipeline for High Content Screening in Drug Discovery - High-Throughput Image Reconstruction and Analysis - page 214

plate handler can usually be integrated with the imager for a large screening ef-
fort. Several instruments also provide live imaging and liquid handling capabilities.
In Table 9.1 we provide a partial list summarizing their main features. Technical
details can easily be obtained by contacting the specific vendors. One has to con-
sider many factors before choosing a high content imager that fits a specific need.
Obviously, price, capability, IP status, and maintenance are all issues that need to
be considered carefully. The most important thing is to test several instruments
with real samples and questions in mind. For example, if counting responsive cells
is the only desired application, one can choose a high speed instrument such as
Explorer with low magnification and totally avoid image analysis. But if detailed
intracellular structures such as endosomes are the primary interest, one has to
go for an instrument with a higher magnification and a more sensitive camera,
perhaps with confocal capability. High content imagers generate a large volume
of raw data, especially when fitted with an automated plate loader. If large scale
screening is the main application, one has to consider data archiving solutions.
A database built in with the instrument is good for handling a large, multiuser
environment but sufficient IT support is often required. All data generated for this
chapter was from an upgraded version of Cellomics ArrayScan VI or MetaXpress
Ultra Confocal imagers.
9.6 Image Analysis
Image analysis is a complex topic, and comprehensive coverage of the topic is
beyond the scope of this chapter. At this moment most of the HCS instrument
vendors provide some level of image analysis capabilities which are developed for
specific biological applications, such as nuclear translocation, object content and
texture measurements, vesicle counting, and neurite length/branch scoring. Vendor
software usually gets the most use, largely because the proprietary format used by
different vendors makes it hard to export and analyze in other software packages.
But most vendor software is sufficient in generating wellular summary of image
data for simple analysis. Generally speaking, most of the HCS image analysis
starts with nuclei recognition, because nuclei are spatially well separated with
sharp boundaries. A nuclear counter stain such as Hoechst33342 is most often
used to outline the nuclear boundary. Thus, they can be easily recognized and
accurately separated with almost any boundary detection algorithm. For the same
reason, bright spots such as endosomes and Golgi can be detected and segmented
easily [21].
Accurate cell boundary detection is a complex task. Because cells are irregu-
larly shaped and frequently overlapped, efforts to accurately identify cell bound-
aries often suffer from serious under or over-segmentation problems. In practice,
a proportional ring expansion or watershed expansion from the nuclear area is
often used to approximate the cellular area. More complex iterative decision mak-
ing processes have been shown to be able to detect cell boundary accurately. First,
over-segmentation of cell bodies and then joining at a later stage with complex,
environment-sensitive criteria were employed in these exercises [22, 23]. Once
boundary detection is achieved, it is rather straightforward to score a multitude of