Automated Cell Segmentation of Fission Yeast Phase Images

Segmenting Cells from Light Microscopy Images

Jennifer O’Brien

1,2

, Sanual Hoque

, Daniel Mulvihill

and Konstantinos Sirlantzis

School of Engineering and Digital Arts, University of Kent, Canterbury, U.K.

School of Biosciences, University of Kent, Canterbury, U.K.

Keywords:

Automated Segmentation, Light Microscopy, Fission Yeast.

Abstract:

Robust image analysis is an important aspect of all cell biology studies. The geometrics of cells are critical for

developing an understanding of biological processes. Time constraints placed on researchers lead to a narrower

focus on what data are collected and recorded from an experiment, resulting in a loss of data. Currently,

preprocessing of microscope images is followed by the utilisation and parameterisation of inbuilt functions of

various softwares to obtain information. Using the ﬁssion yeast, Schizosaccharomyes pombe, we propose a

novel, fully automated, segmentation software for cells with a signiﬁcantly lower rate of segmentation errors

than PombeX with the same dataset.

1 INTRODUCTION

Fission yeast are rod-shaped, unicellular organisms

that share the same molecular and cellular pathways

and proteins as more complex eukaryotes (Hayles

et al., 2013). However, they lack differentiated cells

that can complicate experimental results making them

ideal model organisms for studying fundamental cel-

lular processes.

The cell cycle of all eukaryotes has the same four

stages: gap 1 (G1), DNA synthesis (S), gap 2 (G2)

and mitosis (M) (Figure 1). In G1 cells grow to twice

their original size, duplicating all of the internal struc-

tures except the nucleus. Chromosomal duplication

occurs during S phase and this process is checked du-

ring G2. During mitosis the nucleus divides and the

cell separates into two identical daughter cells.

Fission yeast cells grow exclusively from their

ends. Therefore, knowing the length, width and area

of a cell can inform researchers as to how the cell

is growing, and in what stage of the cell cycle it is.

By monitoring these parameters in cultures of mutant

cells, researchers can rapidly determine whether cell

growth or cell cycle regulation are perturbed by the

mutations within the cells genome.

Labelling proteins with ﬂuorescent proteins has

had a signiﬁcant impact on the types of experiments

bioscientists can design. Previously, 4,6-diaminidino-

2-pheylindole (DAPI) staining was used to visualise

the nuclei of cells. As with all DNA binding agents,

Figure 1: Progression of the cell cycle through stages.

DAPI perturbs DNA organisation and so only single-

time point images of one generation can be analy-

sed (Zink et al., 2003). Using ﬂuorescent proteins

such as green ﬂuorescent protein (GFP) to tag intrin-

sic SNA binding proteins, such as histones, allows

cells to continue growing and dividing and therefore

time-lapse images can be taken to analyse the effect

of mutants on the DNA cycle.

Automating the analysis of microscopy images in-

creases the volume of data that can be scrutinised.

These larger datasets increase the probability of rare,

atypical occurrences being detected and reduce the

risk of skewed data. Another beneﬁt of such programs

is the reduced workload to the researcher, which al-

lows for more experiments to be planned and execu-

ted, and money to be saved.

O’Brien J., Hoque S., Mulvihill D. and Sirlantzis K.

Automated Cell Segmentation of Fission Yeast Phase Images - Segmenting Cells from Light Microscopy Images.

DOI: 10.5220/0006149100920099

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 92-99

ISBN: 978-989-758-215-8

1.1 Previous Research

There have been a number of software previously de-

veloped that were designed to segment cells in mi-

croscopy images. Some are designed for use with

rod-shaped cells including, or even speciﬁcally, ﬁs-

sion yeast; whilst others are more generic. Many of

these programs, however, require ﬂuorescent images

too for segmentation. Those which are phase only en-

counter problems because ﬁssion yeast are unicellu-

lar organisms so images should have a sparser density

than tissue samples which most software is designed

for.

The use of a high content analysis system by

PombeX involves accompanying ﬂuorescently label-

led nuclei images for the phase images. An upda-

ted version now means that the ﬂuorescent images

are not necessary for segmentation. Both methods

distinguish foreground and background pixels based

on intensities and then use an active contour algo-

rithm (Peng et al., 2013). To ensure segmentation

measurements are as reliable as possible, machine le-

arning classiﬁers remove incorrectly segmented cells

as far as possible.

LineageTracker is another software which has

been designed speciﬁcally for ﬁssion yeast, but has an

exclusive remit. It was intended for analysing ﬂuores-

cently labelled spindle poles (Schmitter et al., 2013).

Due to the speciﬁcity of the program, cell segmenta-

tion requires exactly two ﬂuorescent markers per cell,

which have to be manually selected.

CellSegm is a segmentation software which is de-

signed to segment cells from 2-dimensional or 3-

dimensional ﬂuorescent images. It was intended for

surface stained cells, however, it has been modi-

ﬁed to work without staining by utilising stained nu-

clei (Hodneland et al., 2013). It also has modules

which output statistical data.

RodCellJ is an ImageJ based segmentation pro-

gram. It segments ﬂuorescent nuclei and tracks them,

however, it does not segment the cells from a phase

image (Schmitter et al., 2013). This means that re-

searchers are unable to identify if the cells growth is

perturbed unless it disturbs the DNA cycle.

CellProﬁler is a user-friendly interface for cell

segmentation that allows users to create pipelines to

select modules for segmentation. These pipelines are

published and freely available, however, there is not

one currently available which is suited for ﬁssion ye-

ast segmentation (Carpenter et al., 2006).

1.2 Problem Formation

We aim to develop a fully automated system that seg-

ments ﬁssion yeast cells from phase images with no

additional inputs. These segmented cells should be

accurate to a degree that geometric parameters, such

as cell length, width and area, can be measured. A low

segmentation error rate is crucial as they skew results,

and result in manual segmentation being preferable.

2 METHODOLOGY

2.1 Terminology and Notations Used

EMMG Edinburgh minimal media with glutamate as

a nitrogen source.

°C degrees Celsius.

2.2 Growth and Mounting of Cells

Schizosaccharomyces pombe strain hht-cgfp (gene

encoding for histones is fused to cDNA for GFP at

the endogenous locus resulting in ﬂuorescent nuclei),

provided by the Mulvihill lab were cultured to log

phase in EMMG(Moreno et al., 1991) at 25°C. Cells

were mounted as stated in Fission Yeast: a laboratory

manual(Mulvihill, 2016).

2.3 Image Viewing and Capturing

Program development images were captured on a

photometrics quantEM CCD camera attached to an

Olympus IX71 inverted microscope with a 100x oil

lens. Program testing images were captured on a zyla

4.2 sCMOS camera attached to an Olympus IX71 in-

verted microscope with 60x air and 100x oil lenses.

MetaMorph (Molecular Devices, USA) software was

used to view and capture the images. Different ca-

meras were used to thoroughly test the usability of

the proposed Outline software regardless of the set

up found in different labs. Images were stored as

16-bit or 32-bit (developmental and testing respecti-

vely) TIFF ﬁles. Each image consists of a 512 x 512

or 1024 x 1024 array (developmental and testing re-

spectively) with each pixel representing 16 µm by 16

µm or 6.5 µm by 6.5 µm (developmental and testing

respectively).

Altering the focus of the images may result in dif-

ferent outcomes, however, for this study, the images

were captured under the conditions required for hu-

man segmentation. This ensured fair assessment of

segmentation by visual inspection.

Automated Cell Segmentation of Fission Yeast Phase Images - Segmenting Cells from Light Microscopy Images

Figure 2: Block diagram of Outline software implementing

the proposed method.

2.4 Proposed Automated Segmentation

Analysis Software

2.4.1 General Description of the Algorithm

The steps of the automated segmentation method pro-

posed in this paper are: Phase Image - the raw image

taken from the saved TIFF ﬁle from MetaMorph; Ini-

tial Thresholding was selected based on the version of

the Outline software run; Edge Detection using a Lap-

lacian of Gaussian ﬁlter; Structural Elements - hori-

zontal and vertical lines with widths of three pixels

are used to dilate, holes are ﬁlled as are the outer four

pixels on each side; Solidifying Outlines opens and

erodes the image with disks of 10 and three pixels re-

spectively; Segmented Outline Image then extracts the

perimeter and overlays the segmentation outlines over

the original image; Regionprops returns the length,

width, area and perimeter etc of the cells; Useable

Data is the output table of Regionprops.

2.4.2 Differences between Outline Versions

The main differences between the different versions

are in the initial thresholding. OutlineV1 has no thres-

holding, OutlineV2 requires a threshold value to be

provided by the user.The threshold values for Outli-

neV2 are selected visually by the user using the Image

Segmenter app in Matlab. The ﬁrst ﬁve images from

each new dataset were assessed and the median seg-

mentation selected for the whole dataset. OutlineV3

Figure 3: Block diagram of FinalOutline software.

automatically thresholds the images.

FinalOutline has the same thresholding as Out-

lineV3, but has additional changes in later stages.

There is a module to segment end-to-end touching

cells which is important for the analysis of just di-

vided cells. Another module which has been added

converts the segmented outlines to red which is easier

to evaluate the accuracy of the segmentation.

2.4.3 Schematic of FinalOutline

The differences and extra modules for the FinalOut-

line software are: Initial Thresholding uses histogram

equalisation; Separate Touching Cells compares the

number of objects before and after the structural ele-

ments, if the number is the same, it continues, if it

is different, it returns to the original image and de-

termines where the ends of the separate cells are in

Separate Touching Cells and creating an artiﬁcial di-

vide which is then incorporated in the Solidify Out-

lines module; Remove Segmentation Errors removes

and cell segmentations which have more than four cell

ends detected in Separate Touching Cells. All of the

modules which have not been described here remai-

ned the same between all versions of the Outline soft-

ware.

BIOIMAGING 2017 - 4th International Conference on Bioimaging

Figure 4: Segmentations to demonstrate the assessment cri-

teria; (a) perfect, (b) acceptable, (c) unacceptable segmen-

tation, (d) segmentation of background as cell.

3 RESULTS AND DISCUSSION

3.1 Assessing Segmentation Efﬁciency

Cells which were deemed perfect were segmented to

the standard considered reasonable by manual seg-

mentation. For this paper, we consider this to be

continuously touching the inner or outer line of the

cell wall (Figure 4a). Acceptable segmentations are

when the segmented line deviates slightly from per-

fect segmentation as determined by the researcher (Fi-

gure 4b). Segmentation errors include segmentati-

ons where it is obvious to the human eye that part

of the cell has been excluded or where segmentation

has occurred in a region where there are no cells (Fi-

gure 4c-d). These segmentations skew the data and

are therefore detrimental to program efﬁcacy. These

criteria are still subjective, however, they can be grou-

ped to valid (perfect and acceptable) and invalid (seg-

mentation errors) to demonstrate the usability of the

software to the end user.

Two development datasets were collected on dif-

ferent days to recreate the realistic variability in da-

tasets. A total of 119 images were taken, containing

408 cells - 140 single cells and 268 touching cells.

The results of cell segmentation by each version of

Outline for the combined dataset can be seen in Table

Another dataset was created for testing different

lenses. One dataset per lens was taken for the testing

datasets. The 60x air lens had a total of 300 images

while the 100x oil lens had 200. The segmentation

efﬁciency of both FinalOutline and PombeX on both

datasets can be seen in Table 2.

The percentages seen in both tables are based on

the total number of cells determined by visual in-

spection.

3.2 Differences between Outline

Software Versions

Automating the initial step of the program for Out-

lineV3 resulted in over a four-fold decrease in per-

fect segmentations and around a ten-fold increase in

erroneous segmentations (Table 1). When compared

to OutlineV3, FinalOutline produces better segmenta-

tion results. The number of perfect segmentations in-

creases from 32 to 177 with nine touching cells being

segmented separately, the slightly off segmentations

also improve from 91 to 137 and the segmentation

errors fall considerably from 50.98% to 13.48% (Ta-

ble 1). These numbers are still high when compared

to the ﬁrst two versions as there are still signiﬁcantly

more erroneous segmentations and missed cells.

Some of the missed cells in FinalOutline have

been removed from the segmented output images by

the program in response to Remove Segmentation Er-

rors. This is veriﬁable as OutlineV1 has a total of

247 cells segmented as touching, OutlineV2 has 249

while only 187 are segmented by FinalOutline (Table

1). Even if all of the erroneous segmentations of Fi-

nalOutline, and none of those from OutlineV1 or Out-

lineV2, were of touching cells, there are less touching

cells included in the output data of FinalOutline.

3.3 Comparison of FinalOutline with

PombeX

Images collected on 100x oil lens are of a higher qua-

lity but do not allow for high throughput. We tested

our program, FinalOutline, and PombeX on both ty-

pes of images. Neither program requires any input

parameters and needs only the phase images for ana-

lysis.

PombeX was able to segment signiﬁcantly more

cells from the 60x air lens dataset compared to the

100x oil lens dataset (over 10% from less than 1%).

FinalOutline was more successful on the 100x oil da-

taset (27.91% perfect for 100x oil, 6.67% for 60x air;

Table 2). For PombeX, the number of missed cells

was over three-fold more in the 60x air data then the

100x oil data. However, a total of 6.49% of cells mis-

sed in the 60x air data is much lower than the 93.91%

of FinalOutline for the same data.

The most concerning statistic for bioscientists are

the percentage of segmentation errors. These errors

are what skew results and make manual segmentation

preferable over current programs. FinalOutline had

0.58% and 2.25% error rates (100x oil and 60x air

data respectively) which are vast improvements on the

325.00% and 330.74% of segmentation errors (100x

oil and 60 air respectively) of PombeX (Table 2).

Automated Cell Segmentation of Fission Yeast Phase Images - Segmenting Cells from Light Microscopy Images

Table 1: Performance of the Outline software.

Table 2: Performance of FinalOutline compared to PombeX on images taken with either a 100x oil lens or a 60x air lens.

Two features FinalOutline has, which PombeX

does not, is an easy input function and saved output

ﬁles. The input function allows for a whole folder of

images to be automatically analysed. To set it up for

a new set of data takes less than one minute. Once

it has been conﬁgured, the computer can be left until

the researcher is ready.

The output ﬁles of FinalOutline include the seg-

mented images and two tables. The ﬁrst table contains

the image number and the number of cells while the

second table contains the length, width, perimeter and

area of every cell in the dataset. Having these ﬁles sa-

ved allows researchers to quickly assess the quality of

the segmentation outlines and remove erroneous data

from the tables based on the images to reduce skew-

ness. PombeX displays the segmented image but has

no statistical output data.

4 CONCLUSION

The proposed automated segmentation software

shows many characteristics beneﬁcial for bioscien-

tists. The software is fully automated and requires no

seeds for segmentation as many other programs do.

The number of segmentation errors are minimal in Fi-

nalOutline but reducing them further will increase the

usability of the program.

BIOIMAGING 2017 - 4th International Conference on Bioimaging

Figure 5: Example images from the developmental dataset; (L-R) original image, output from OutlineV1, output from Outli-

neV2, output from OutlineV3, output from FinalOutline.

ACKNOWLEDGEMENTS

This work was funded by 50th anniversary GTA fun-

ding from the University of Kent, secured by the

Industrial Biotechnology Centre. Support from the

Schools of Biosciences and Engineering and Digital

Arts.

Automated Cell Segmentation of Fission Yeast Phase Images - Segmenting Cells from Light Microscopy Images

Figure 6: (A) Example images from 60x air lens testing dataset; (T-B) original image, output from FinalOutline, output from

PombeX. (B) Example images from 100x oil lens testing dataset; (T-B) original image, output from FinalOutline, output from

PombeX. Images from FinalOutline output shown with contrast adjusted for clarity.

BIOIMAGING 2017 - 4th International Conference on Bioimaging

REFERENCES

Carpenter, A. E., Jones, T. R., Lamprecht, M. R., Clarke, C.,

Kang, I. H., Friman, O., Guertin, D. A., Chang, J. H.,

Lindquist, R. A., Moffat, J., Golland, P., and Sabatini,

D. M. (2006). CellProﬁler: image analysis software

for identifying and quantifying cell phenotypes. Ge-

nome Biology, 7:R100.

Hayles, J., Wood, V., Jeffery, L., Hoe, K.-L., Kim, D.-

U., Park, H.-O., Salas-Pino, S., Heichinger, C., and

Nurse, P. (2013). A genome-wide resource of cell cy-

cle and cell shape genes of ﬁssion yeast. Open Bio-

logy, 3(5):130053.

Hodneland, E., K

ogel, T., Frei, D. M., Gerdes, H.-H., and

Lundervold, A. (2013). CellSegm - a MATLAB tool-

box for high-throughput 3D cell segmentation. Source

Code for Biology and Medicine, 8(16).

Moreno, S., Klar, A., and Nurse, P. (1991). Molecular ge-

netic analysis of ﬁssion yeast Schizosaccharomyces

pombe. Methods Enzymol., 194:795–823.

Mulvihill, D. P. (2016). Fission Yeast: a laboratory manual.

In Hagan, I. M., Carr, A. M., Grallert, A., and Nurse,

P., editors, Fission Yeast, chapter 10. Live C, pages

220–222. CSH Press, Cold Spring Harbor, 1st edition.

Peng, J.-Y., Chen, Y.-J., Green, M. D., Sabatinos, S. A.,

Forsburg, S. L., and Hsu, C.-N. (2013). PombeX: Ro-

bust Cell Segmentation for Fission Yeast Transillumi-

nation Images. PLOS ONE, 8(12):e81434.

Schmitter, D., Wachowicz, P., Sage, D., Chasapi, A., Xe-

narios, I., Simanis, and Unser, M. (2013). A 2D/3D

image analysis system to track ﬂuorescently labeled

structures in rod-shaped cells: application to measure

spindle pole asymmetry during mitosis. Cell Div., 8:6.

Zink, D., Sadoni, N., and Stelzer, E. (2003). Visualizing

chromatin and chromosomes in living cells. Methods,

29(1):42–50.

Automated Cell Segmentation of Fission Yeast Phase Images - Segmenting Cells from Light Microscopy Images