![](bg2.png)
tion within the biological domain.
1.1 DIC Microscopy
There have been many advances in light microscopy
over the years, particularly the discovery of Green
Fluorescent Protein (GFP) which fluoresces under
blue light. As such it can be used to tag proteins
or genes, enabling the detection and localisation of
their expression (Tsien, 1998). However, the illumi-
nation can cause cell damage which affects both the
movement and lifespan of a cell (Stephens and Allan,
2003).
Using transmitted light for imaging live cells can
provide additional information and detail relating to
the cell shape. One such method is phase contrast
microscopy, which can outline the cells and highlight
some organelles. This, however, surrounds the im-
aged cell by a bright halo making it difficult to iden-
tify distinct edges (Lane and Stebbings, 2006).
Another method is DIC microscopy, in which a
beam splitter is applied to the light and half passed
through the cell. By measuring the difference in the
lengths of the optical paths the thickness of the cell
can be estimated (Murphy, 2001). The resulting im-
age appears as three-dimensional, and high contrast
images can be created showing more detail than previ-
ous methods. This is particularly useful for transpar-
ent objects, which would normally be difficult to see
without staining (Salmon and Tran, 2007; Schwartz
et al., 2003).
Although the method of obtaining DIC images can
provide many advantages, it can also cause problems
when the images are to be processed. The images
appear to be illuminated by a highly oblique light
source, which creates a shadow and light 3D effect,
but this also causes a large variation in the brightness
of the background (Schwartz et al., 2003; Kuijper and
Heise, 2008) and the object being studied, which can
be seen in Figures 2(a) and 2(b).
2 RELATED WORK
When studying normal and abnormal cell movement
it is not only important to find the location of the
cell, but to segment it in such a way that the size
and shape of the cell can also be recorded for anal-
ysis. Although a lot of previous work on cells has
involved counting or tracking the cells, far less has
been produced on shape analysis. Pincus and The-
riot compared methods for cell shape analysis by in-
vestigating methods which provided interpretable and
accurate numerical representations of the cell shape
(Pincus and Theriot, 2007). They found that principal
component analysis was the method which can most
accurately capture modes of shape variation, and this
has been very successful on keratocyte cells, which
can be modelled with only a few modes of shape vari-
ability (Keren et al., 2008). However, this approach
is not suitable for amorphous structures such as my-
oblast cells where the features cannot be aligned.
Much of the previous work on cell shape analy-
sis has involved manual segmentation, which is very
time consuming. Wu et al. used a two step procedure
to reduce this cost, by manually selecting each region
of an image which contained a cell (Wu et al., 1995).
They found that this reduced the variation in inten-
sity levels which occurs across images and allowed
for local thresholding. Further work to automate the
process of segmentation has often been for the pur-
poses of cell counting or tracking, in which obtaining
the cell boundary is not necessary. Jiang et al. used
SIFT to compare key-points and track cells in DIC
videos, without the necessity to find cell boundaries
(Jiang et al., 2010). Bise et al. also looked at tracking
and intentionally excluded portions of the cells from
the segmentation, such as the long thin parts which
deform significantly as these can confuse the tracking
(Bise et al., 2009).
Level sets have used to automate segmentation.
This was found to be successful on cells which
showed symmetry and did not contain too many vis-
ible sub-structures (Kuijper and Heise, 2008). Young
and Gray also relied on similarly shaped elliptical
cells applying a curvature constraint and segmenting
cells using edge contours (Young and Gray, 1996).
Segmentation of complex shaped cells was investi-
gated by Simon et al. but the method was found to
be unsuitable for images with a large background to
cell ratio, and for cells with a thin membrane (Simon
et al., 1998), such as ours. These cells (with poor in-
tensity contrast) were excluded from the analysis.
Li and Kanade developed a method for precondi-
tioning DIC images to assist with segmentation, but
when tested on our images the areas of most contrast
appear as if affected by a very bright “shadow” which
distorts the cell shape and there is no increase in the
definition of the thinner, less obvious parts of the cell
(Li and Kanade, 2009), as can be seen in Figure 1(a).
ImageJ (Abramoff et al., 2004) is an imaging program
which provides a method (pseudo-flat-field) for inten-
sity correction of images. When tested on our images
using a smaller filter this produced a “glow” around
the cell and slight blurring, as well as a small amount
of “shadow” in one corner. As the size of the filter
was increased, so did the area of shadow. The result
of using the default size is shown in Figure 1(b).
AutomatedSegmentationofCellStructureinMicroscopyImages
99