Automated Segmentation of Cell Structure in Microscopy Images

Nicole Kerrison and Andy Bulpitt

School of Computing, University of Leeds, Woodhouse Lane, Leeds, LS2 9JT, U.K.

Keywords:

Segmentation, Lighting Correction, Cell Motility, Microscopy, Lamellipodia, DIC.

Abstract:

Understanding cell movement is important in helping to prevent and cure damage and disease. Increasingly,

this study is performed by obtaining video footage of cells in vitro. However, as the number of images

obtained for cellular analysis increases, so does the need for automated segmentation of these images, since

this is difﬁcult and time consuming to perform manually. We propose to automate the process of segmenting

all parts of a cell visible in DIC microscopy video frames by providing an efﬁcient method for correcting the

lighting bias and a novel combination of techniques to detect different cell areas and isolate parts of the cell

vital to their movement. To the best of our knowledge we contribute the only method able to automatically

detect the thin cellular membranes in DIC images. We show that the method can be used to isolate features in

order to detect variations vital to motility in differently affected cells.

1 INTRODUCTION

The study of cells is a key ﬁeld in modern science in

order to help understand biological processes. Cur-

rently, much of the analysis is conducted by hand but

this can be time consuming and subject to human er-

ror. Automating this work would allow greater ac-

curacy, with reproducible results at a higher speed.

Automation would additionally allow for quantitative

analysis of the data, rather than just qualitative.

The movement of cells is vital to many biolog-

ical processes, as cells can move to infected areas or

wounds which need healing, and abnormal cell move-

ments can lead to disease and defects. In order to pre-

vent and cure disease we need to be able to understand

how cells migrate and morph, and to be able to mea-

sure their movement and change in shape (Dormann

and Weijer, 2006; Zimmer et al., 2006).

This is particularly important in ameoboid cells,

such as the myoblast cell, in which movement on sub-

strate is achieved through a crawling motion. The cell

extends large, but very thin protrusions called lamel-

lipodia which stretch out and adhere to the substrate,

providing anchor points for the cell to be pulled for-

ward (Becker et al., 2000; Karp, 2010; Middleton and

Sharp, 1984).

Biologists have taken videos of moving cells us-

ing time-lapse DIC (differential interference contrast)

microscopy, in which the light beam is split with part

passed through the cell and the difference calculated

to give a representation of the thickness of the cell,

discussed further in Section 1.1. Due to the nature

of DIC imaging, the thinly spread membranes of the

lamellipodia are often barely visible, with the texture

appearing the same as the background.

Researchers in cell motility wish to obtain infor-

mation about the cells and their movement, both in

unaffected (wildtype) cells, and in those in which

genes have been affected (knocked down or over ex-

pressed) or chemicals added. One particular chemi-

cal, blebbistatin, is an inhibitor which has been found

to affect the cells contractile forces and has an affect

on the lamellipodia and the cells’ motility (Limouze

et al., 2004; Straight et al., 2003; Kolega, 2006). Re-

searchers wish to know if brenk also has a similar

affect on the lamellipodia, and therefore the cells’

movement.

To ﬁnd out more about the lamellipodia they need

to be segmented separately from the cell in order to

calculate information about their size and shape, and

how these change over time. This work aims to seg-

ment not only the cell but also divide it into its compo-

nent parts to enable this information to be calculated.

This is done in a fully automated manner, in which

a video can be analysed with no need for manual se-

lection or segmentation, and no necessity for manual

adjustment of parameters. We ﬁrstly present results

conﬁrming our general segmentation method works

against a variety of DIC cell images, then focus on

the particular application of lamellipodia segmenta-

Kerrison N. and Bulpitt A..

Automated Segmentation of Cell Structure in Microscopy Images.

DOI: 10.5220/0004659800980105

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 98-105

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

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).

AutomatedSegmentationofCellStructureinMicroscopyImages

(a) Original (b) With

colour

background

(d) Corrected (e) Corrected

with colour

Figure 2: Correcting for DIC shadow effect.

The ﬁrst column shows the original frames and the second shows these with colour applied, to make the lighting variation

more visible. The middle column shows the brightness of pixels over the background when the cell is blurred out. The ﬁnal

two columns show the resulting images, with the shadow effect corrected.

(a) Preconditioned Image (b) ImageJ

Pseudo-Flatﬁeld

Figure 1: Alternative methods to account for lighting varia-

tions.

The ﬁrst image shows the result of preconditioning the im-

age using the method from (Li and Kanade, 2009). The

second image uses ImageJ (Abramoff et al., 2004) to cor-

rect the intensity but using a ﬁlter large enough not to affect

the cell results in a darker area still visible in the top right

of the image.

3 LIGHTING CORRECTION

DIC imaging causes lighting variations across the

frame, as can be seen in Figure 2(a). Two example

frames are shown, with the lighting and shadow ap-

pearing differently in both. The top image appears

to be lit from the bottom left, with shadow appear-

ing in the top right, and the bottom image appears to

be lit from roughly the middle with shadow appear-

ing in two corners, Figure 2(a). A colour map has

been added to the images, as in Figure 2(b), to make

the light variations more visible to the eye. The light-

ing variations occur smoothly across the whole image,

and affect the cells as well as the background.

Using a Gaussian blur, the cell detail can be re-

moved, shown in Figure 2(c). The intensity differ-

ences between the original and the blurred image are

stored for every pixel, which allows us to obtain the

detail of the cell and other objects in the image, with-

out need to model the background. The background

is then set to the mean of the image intensity of the

frame. The stored values are then re-added to the ad-

justed image to show the cell and other information

with no loss of detail, as seen in Figure 2(d), and again

in 2(e) with added colour to make the lack of variation

more apparent. This is repeated for each frame in the

video.

The pixel values were calculated across the diag-

onal with the greatest variation in intensity (e.g. bot-

tom left to top right in the examples shown in Figure

2) before and after the background lighting variation

was corrected. It can be seen from Figure 3 that this

efﬁcient method creates a uniform background level

across the frames.

This was also tested on images from an internet

search and was demonstrated to be equally effective

on all DIC images tested, including those where the

cells took up the majority of the frame. Some exam-

ple results are shown in Figure 3. The ﬁrst two rows

are the images from Figure 2, the third is an online

image of HeLa cells, in which there is very little back-

ground, and the fourth is an online image of a C. Ele-

gans worm tail. The original images all show a varia-

tion in background lighting as well as the detail of the

object being examined, as can be seen in Figure 3(a).

The uneven background without the detail (after the

Gaussian blur) is shown in 3(b). The corrected images

all show a more even background illumination with-

out loss of detail of the cell or worm, as seen in Figure

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

100

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 100 200 300 400 500 600 700

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0 100 200 300 400 500 600 700

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0 100 200 300 400 500 600 700

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0 50 100 150 200 250 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

(a) Original

Image

(b) After

Gaussian

Blur

Image

Figure 3: Pixel values before and after correction.

The pixel values were recorded along the diagonal with the

most lighting variation before and after lighting correction.

The ﬁrst two images are those shown in Figure 2 and the

second two were found using an online image search. All

show a more even background without loss of information

of the cell or worm.

3(c). This is successful even with only small amounts

of background visible as the lighting variations occur

not only over the background but also across the cells.

Although other there are other methods for correcting

lighting, this efﬁcient method proved suitable for all

DIC images tested.

4 THRESHOLDING AND EDGE

DETECTION

Due to the lighting variations in DIC images, thresh-

olding based on brightness is not normally possible,

as can be seen Figures 4(b) and 4(c). Our lighting

adjustment method not only corrects the background

visually, but also provides the mean to which the

background is adjusted on which thresholding can be

based. The image seen in Figure 4(d) shows in white

those pixels in the light-corrected image that are more

than 10% brighter or darker than the mean value, with

those that are nearest to the mean being set to black

as background.

The brightest and darkest areas of the image such

as the main body of the cell can be detected using

thresholding. In most cases, cell videos are taken of

one particular type of cell, and as such the main bod-

ies are of reasonably similar sizes. This stage allows

(a) Original (b) Threshold

at 50%

around mean

(d) Threshold after

lighting correction

Figure 4: Thresholding DIC images.

Prior to lighting correction, it is not possible to segment DIC

images using thresholding alone as the differences in the

background cover the same intensities as the cell. The ﬁnal

image shows that the main parts of the cell can be easily

thresholded after the lighting adjustment.

us to calculate the size and provides information for

removing non-cellular material from the frame. Al-

though this thresholding ﬁnds the main body of the

cell, the fainter lamellipodium still needs to be seg-

mented.

(a) Original (b)

Threshold

(c)

Combined

Figure 5: Combined results.

It can be seen that thresholding alone cannot detect large

area of the lamellipodia at the bottom, but when combined

with edge detection the full cell is detected. The images are

also shown in binary beneath the results to aid viewing.

Taking F(t) as a light corrected frame at time t,

we apply a Canny Edge detection generating a set of

edges E(t). We only keep the subset of edges e ⊆ E(t)

which are connected to the “on” pixels in the thresh-

olded image F(t), as shown in Figure 5(c). The top

AutomatedSegmentationofCellStructureinMicroscopyImages

101

images are shown in greyscale, and the lower images

are shown in binary with all non-zero pixels shown as

white for clarity.

The resulting image is then ﬁlled, smoothed and

the small non-cellular areas are removed, proportional

to the body size calculated in the thresholding step.

The result is a binary image in which the white areas

are the cell and the black are the background. This

white shape can be applied to the original image as

a template to provide the boundary of the cell and to

segment it from the background. We denote this cell

segmentation at time t as C(t).

(a) Outline of

one cell

(b) Outline of

two cells

many cells

Figure 6: Outline of cells from the template.

Using the template created from the combined thresholding

and edge detection method, the outline of the cells is traced

onto the original image. As can be seen, the detection is

very close to the actual edge, even when this is difﬁcult to

detect by eye. The bottom row of images have had the back-

ground manually adjusted so the outline of the cells can be

seen.

The resulting segmentation is very close to the cell

boundary (Figure 6), even where it is difﬁcult to de-

tect by eye. Manually adjusted images have been cre-

ated to show the cells with a paler background to as-

sist in visualisation. Additionally, it is shown to work

on varying sizes of cells and with different degrees of

background coverage. Tables 1 and 2 show the accu-

racy of the results which are discussed in section 7.1.

5 ITERATIVE CORRECTION

The resulting segmentation for the thresholding and

edge detection method was very close to the man-

ual segmentation for most frames (which will be dis-

cussed in Section 7.1), but in some frames the lamel-

lipodia were not entirely detected, an example of

which can be seen in Figure 7(a). In these cases it

was often also extremely difﬁcult or even impossible

to detect these areas by eye in the individual frame,

but the location could be inferred from the movement

in the surrounding frames. To detect these cases (of

incomplete segmentation) the cell size was compared

over consecutive frames, with those frames showing a

signiﬁcant increase or decrease being the most likely

to not be entirely segmented.

We denote a cell segmentation at time t as C(t).

At each time-step, we compare the area of C(t) with

that of C(t − 1). If the area has changed by at least

10%, we take the intersection of C(t − 1) and C(t +1)

to generate the basis of a new segmentation image,

C(t). As stated in section 4, we have a set of edges

E(t) generated from F(t). We now keep those edges

ˆe ⊆ E(t) that are connected to

C(t). We replace each

segmentation C(t) with the new version

C(t) + ˆe.

The entire process is then repeated over the entire

video using the updated frames. This is run iteratively

until such a time as no two consecutive frames have

an area difference of greater than 10%, or until the

method is creating no further changes, which is de-

tected by comparing the list of sizes in consecutive

runs.

The advantage of this iterative framework be-

comes evident where we have two or more consecu-

tive erroneous segmentations. Take as an example the

situation where C(t) and C(t + 1) are incorrect; when

taking the intersection of C(t − 1) and C(t + 1), we

still obtain an incomplete representation of the cell.

As such the initial pass of the method will not create

a full segmentation throughout the video. However,

if we repeat the full process as described above, seg-

mentation improves at each pass until we converge on

a more complete solution.

The resulting segmentation shows an improve-

ment on the previous results, as shown in Figure 7(b),

when compared with manually segmented cells (Ta-

bles 1 and 2).

6 SEGMENTATION OF

LAMELLIPODIA

As previously discussed (Section 1), it is important

to not only be able to segment the cell, but also to

identify the component parts. As the size and spread

of lamellipodia is vital to the cells movement, being

able to calculate this can assist in learning about the

cell’s motility.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

102

(a) Before Iterative

Correction

(b) After Iterative

Correction

Figure 7: Correcting the Segmentation.

In some frames the lamellipodia is not fully segmented, as

seen in (a). The previous and subsequent frames are used to

correct this, as shown in (b). In this instance, the boundary

from the segmentation is shown in yellow for clarity.

(a) Original (b) Lamellipodia

Figure 8: Cells segmented to show lamellipodia.

The ﬁrst columns show the original images, the second

shows the segmented lamellipodia.

Figure 8 shows the lamellipodia being segmented

separately from the cell body. As our method uses

thresholding to ﬁnd the main body of the cell this can

be removed from the ﬁnal segmentation to leave only

the lamellipodia. From this, the size can be calculated

over time and the individual regions tracked to assist

in building a pattern of the growth.

The segmentation of the lamellipodia was applied

to three groups of videos; those containing wildtype

cells, those containing cells which had been affected

by blebbistatin and those cells which had been af-

fected by brenk. The size of the lamellipodia was cal-

culated for each frame, as the absolute size and as rel-

ative to the size of the cell body to account for larger

and smaller cells. This was used to assess the affect

of these chemicals on the cells, and the results are dis-

cussed in Section 7.2.

7 RESULTS

The lighting correction method for preconditioning

and the segmentation method were applied to all im-

ages from a variety of test sets with visually good re-

sults. The segmentation of the cell and the lamellipo-

dia were then tested against a subgroup of 26 videos

with the results for each shown below.

7.1 Results for Segmenting the Cell

Although the results for this system were obtained us-

ing a fully automated system, no other automated sys-

tem could be found for comparison of results that was

able to detect and segment the cells. The segmenta-

tion was instead compared to results using CellTrack

(Sacan et al., 2008) by manually adjusting the param-

eters of the CellTrack algorithms to obtain the closest

possible result to the cell boundary. The automated

segmentations, both before and after the iterative cor-

rection, and the CellTrack segmentations were com-

pared with manual segmentations using DICE (Table

1) and Jaccard (Table 2), which compare the over-

lap in segmentations. The average results are shown

over each of three groups of videos; wildtype cells,

those affected by the chemical blebbistatin and those

affected by brenk.

Table 1: Comparison Of DICE Results (% accuracy).

Wildtype Blebb. Brenk

Orig. Segmentation 89.75 86.64 89.33

After Correction 90.13 87.35 90.95

CellTrack 88.81 85.52 90.89

Table 2: Comparison Of Jaccard Results (% accuracy).

Wildtype Blebb. Brenk

Orig. Segmentation 82.25 77.43 81.56

After Correction 82.65 78.29 83.69

CellTrack 80.20 75.49 83.44

The initial segmentation provided a very good re-

sult. From 910 frames tested, only 37 (roughly 4%)

AutomatedSegmentationofCellStructureinMicroscopyImages

103

had small parts of the lamellipodia missing. It can

be seen that this still provides a better result than the

best method involving manual adjustment in all but

one case. After using the iterative correction method

the resulting segmentation was more accurate in every

case.

7.2 Results for Segmenting the

Lamellipodia

The segmentation of the lamellipodia was also tested

on wildtype cells, cells affected by blebbistatin and

cells affected by brenk. Researchers wished to know

if both the blebbistatin and the brenk had an effect

on the lamellipodia, which would in turn affect the

cells’ motility, and if this affect was similar for both of

them. Initially, the size was calculated for the lamel-

lipodia in every frame as well as the size of the cell

body.

It was found that the mean size-ratio was greater

for both the brenk and blebbistatin (with an overall

mean ratio of 0.40 and 0.39 respectively) compared

to the wildtype cells (with a mean ratio of 0.18), as

shown in Table 3. Whilst the mean size-ratio was

larger in blebbistatin and brenk than in wildtype, we

see comparatively more variation in wildtype size-

ratios, as evidenced by a standard deviation of more

than 0.24 relative to a mean of 0.18. Conversely,

the spread of blebbistatin and brenk is smaller than

the mean size-ratio in both cases. This could imply

that although the lamellipodia is larger in both cases

it does not have the same ability to change size by

contracting and expanding as would be the case with

wildtype cells.

Table 3: Size of Lamellipodia Comparative to Cell Body.

Mean Standard Deviation

Wildtype 0.1842 0.2448

Blebbistatin 0.4043 0.3337

Brenk 0.3923 0.3195

The size-ratios for every frame in each group of

videos were compared with the Kolmogorov-Smirnov

test against the null hypothesis that they were from

the same distribution. Both the blebbistatin and the

brenk were tested against the wildtype cells and in

both cases were shown not to be from the same distri-

bution (for wildtype and blebbistatin p < 2.5 × 10

−18

and for wildtype and brenk p < 8.5 × 10

−20

). The

blebbistatin and brenk when tested against each other,

however, were not from different distributions (p =

0.5227).

8 CONCLUSIONS AND FURTHER

WORK

We have proposed a method for segmenting cells in

DIC images, even when there is very little contrast

difference between the cell and the background. This

involves a fast lighting correction technique which

has proved successful on not only the cellular footage

currently being assessed, but also on DIC images

found online.

We provide a fully automated segmentation tech-

nique which can segment all parts of the cell including

the thin membrane. To correct the images where the

boundary does not fully encompass the lamellipodia

we use an iterative method to improve the segmenta-

tion based on the surrounding frames. This is a fully

automated system which provides a more accurate re-

sult than currently available tools which require man-

ual input.

The results show that the cell can be accurately

segmented, and the information gained in the process

can then be used to identify lamellipodia separately.

This allows us to obtain further information about the

size of the components affecting the cells’ movement.

We have shown that when affected by brenk or bleb-

bistatin the lamellipodia in cells grow differently to

the lamellipodia in wildtype cells, and that overall the

size is similar.

REFERENCES

Abramoff, M. D., Magelhaes, P. J., and Ram, S. J. (2004).

Image processing with ImageJ. Biophotonics Int,

11(7):36–42.

Becker, W., Kleinsmith, L., and Hardin, J. (2000). The

world of the cell. Benjamin/Cummings series in the

life sciences. Benjamin Cummings.

Bise, R., Li, K., Eom, S., and Kanade, T. (2009). Reli-

ably tracking partially overlapping neural stem cells in

dic microscopy image sequences. In MICCAI Work-

shop on Optical Tissue Image Analysis in Microscopy,

Histopathology and Endoscopy (OPTMHisE), pages

67–77.

Dormann, D. and Weijer, C. J. (2006). Review: Imaging Of

Cell Migration. European Molecular Biology Organi-

zation Journal, 25(15):3480 – 3493.

Jiang, R., Crookes, D., Luo, N., and Davidson, M. (2010).

Live-Cell Tracking Using SIFT Features in DIC Mi-

croscopic Videos. Biomedical Engineering, IEEE

Transactions on, 57(9):2219–2228.

Karp, G. (2010). Cell Biology. John Wiley and Sons, Asia,

sixth edition.

Keren, K., Pincus, Z., Allen, G., Barnhart, E., Marriott,

G., Mogilner, A., and Theriot, J. (2008). Mecha-

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

104

nism of shape determination in motile cells. Nature,

453(7194):475–480.

Kolega, J. (2006). The Role of Myosin II Motor Activity

in Distributing Myosin Asymmetrically and Coupling

Protrusive Activity to Cell Translocation. Molecular

Biology of the Cell, 17(10):4435–4445.

Kuijper, A. and Heise, B. (2008). An automatic cell seg-

mentation method for differential interference contrast

microscopy. In ICPR 2008. 19th International Confer-

ence on Pattern Recognition, pages 1–4. IEEE.

Lane, J. D. and Stebbings, H. (2006). Transmitted light

imaging. In Stephens, D., editor, Cell Imaging, Meth-

ods Express. Scion, Bloxham, Oxfordshire.

Li, K. and Kanade, T. (2009). Nonnegative mixed-norm

preconditioning for microscopy image segmentation.

In Proceedings of the 21st Biennial International Con-

ference on Information Processing in Medical Imag-

ing (IPMI), pages 362–373.

Limouze, J., Straight, A. F., Mitchison, T., and Sellers,

J. R. (2004). Speciﬁcity of blebbistatin, an inhibitor of

myosin II. Journal of muscle research and cell motil-

ity, 25(4-5):337–341.

Middleton, C. and Sharp, J. (1984). Cell Locomotion in

Vitro: Techniques and Observations. Croom Helm

Ltd, Kent.

Murphy, D. B. (2001). Fundamentals of Light Microscopy

and Electronic Imaging, chapter Differential Inter-

ference Contrast (DIC) Microscopy and Modulation

Contrast Microscopy, pages 153 – 172. Wiley-Liss,

New York, New York.

Pincus, Z. and Theriot, J. A. (2007). Comparison of quan-

titative methods for cell-shape analysis. Journal of

Microscopy, 227(2):140–156.

Sacan, A., Ferhatosmanoglu, H., and Coskun, H. (2008).

Celltrack: an open-source software for cell track-

ing and motility analysis. Bioinformatics, (14):1647–

1649.

Salmon, E. D. and Tran, P. (2007). High-Resolution Video-

Enhanced Differential Interference Contrast Light Mi-

croscopy. In Sluder, G. and Wolf, D. E., editors, Dig-

ital Microscopy, volume 81 of Methods in Cell Biol-

ogy, pages 335 – 363. Academic Press, San Diego,

California, third edition.

Schwartz, S., Murphy, D., Spring, K., and Davidson, M.

(2003). de s

enarmont bias retardation in dic mi-

croscopy. http://www.microscopyu.com/pdfs/

DICMicroscopy.pdf. Nikon MicroscopyU.

Simon, I., Pound, C. R., Partin, A. W., Clemens, J. Q.,

and Christens-Barry, W. A. (1998). Automated im-

age analysis system for detecting boundaries of live

prostate cancer cells. Cytometry, 31(4):287–294.

Stephens, D. J. and Allan, V. J. (2003). Light Microscopy

Techniques for Live Cell Imaging. Science, 300:82 –

86.

Straight, A. F., Cheung, A., Limouze, J., Chen, I., West-

wood, N. J., Sellers, J. R., and Mitchison, T. J.

(2003). Dissecting Temporal and Spatial Control

of Cytokinesis with a Myosin II Inhibitor. Science,

299(5613):1743–1747.

Tsien, R. Y. (1998). The Green Fluorescent Protein. Annual

Review of Biochemistry, 67:509 – 44.

Wu, K., Gauthier, D., and Levine, M. D. (1995). Live cell

image segmentation. Biomedical Engineering, IEEE

Transactions on, 42(1):1–12.

Young, D. and Gray, A. J. (1996). Cell identiﬁcation in

Differential Interference Contrast microscope images

using edge detection. pages 66.1–66.10. 7th British

Machine Vision Conference, BMVA Press.

Zimmer, C., Zhang, B., Dufour, A., Thebaud, A.,

Berlemont, S., Meas-Yedid, V., and Marin, J.-C.

(2006). On the digital trail of mobile cells. Signal

Processing Magazine, IEEE, 23(3):54 –62.

AutomatedSegmentationofCellStructureinMicroscopyImages

105