© 2010 Kelsey et al, publisher and licensee Dove Medical Press Ltd. This is an Open Access article
which permits unrestricted noncommercial use, provided the original work is properly cited.
Pathology and Laboratory Medicine International 2010:2 99–105
Pathology and Laboratory Medicine International Dovepress
submit your manuscript | www.dovepress.com
Dovepress
99
O r I g I n A L r e s e A r c h
open access to scientific and medical research
Open Access Full Text Article
11116
Proliferating cell nuclear antigen (PcnA)
allows the automatic identification of follicles
in microscopic images of human ovarian tissue
Thomas W Kelsey1
Benedicta caserta2
Luis castillo2
W hamish B Wallace3
Francisco cóppola
gonzálvez4
1school of computer science,
University of st Andrews,
st Andrews, scotland, UK; 2Pereira
rossell hospital, Asse, Ministry of
Public health, Montevideo, Uruguay;
3Division of child Life and health,
Department of reproductive and
Developmental sciences, University
of edinburgh, edinburgh, scotland,
UK; 4Department of Obstetrics
and gynecology c, school of
Medicine, University of the republic,
Montevideo, Uruguay
correspondence: Tom Kelsey
school of computer science, University
of st Andrews, north haugh, st Andrews,
KY16 9sX United Kingdom
Tel +44 1334 463249
Fax +44 1334 463278
email tom@cs.st-andrews.ac.uk
Background: Human ovarian reserve is defined by the population of nongrowing follicles
(NGFs) in the ovary. Direct estimation of ovarian reserve involves the identification of NGFs
in prepared ovarian tissue. Previous studies involving human tissue have used hematoxylin and
eosin (HE) stain, with NGF populations estimated by human examination either of tissue under
a microscope, or of images taken of this tissue.
Methods: In this study we replaced HE with proliferating cell nuclear antigen (PCNA), and
automated the identification and enumeration of NGFs that appear in the resulting microscopic
images. We compared the automated estimates to those obtained by human experts, with the
“gold standard” taken to be the average of the conservative and liberal estimates by three human
experts.
Results: The automated estimates were within 10% of the “gold standard”, for images at both
100× and 200× magnifications. Automated analysis took longer than human analysis for several
hundred images, not allowing for breaks from analysis needed by humans.
Conclusion: Our results both replicate and improve on those of previous studies involv-
ing rodent ovaries, and demonstrate the viability of large-scale studies of human ovarian
reserve using a combination of immunohistochemistry and computational image analysis
techniques.
Keywords: histology, feature detection, ovarian reserve, immunohistochemistry, biological
clock
Introduction
The human ovary contains a fixed number of nongrowing follicles (NGF) established
before birth. This number declines with increasing age culminating in the menopause
at 50–51 years.1 Ovarian reserve is defined by the remaining population at a given age.
There is no technique known for direct in vivo estimation of ovarian reserve; indirect
indicators include antral follicle counts, ovarian volume and levels of hormones such
as follicle-stimulating hormone and anti-Mullerian hormone.2 A model describing the
age-related population of NGFs in the human ovary from conception to menopause,
in which the NGF populations of the 325 ovaries studied were all estimated using
variations on the standard methodology developed by Block in the early 1950s,4,5 has
recently been published.3 After oophorectomy (or post-mortem) the ovary is fixed, thin
slices (between 5 and 20 microns) are taken at regular intervals, and these are stained
with hematoxylin and eosin (HE). Sample regions are either inspected manually, or
photographed, with the NGFs appearing in the tissue counted by hand.

Pathology and Laboratory Medicine International 2010:2submit your manuscript | www.dovepress.com
Dovepress
Dovepress
100
Kelsey et al
Table 1 sample comparison of automated counts to human counts
Mag. Image Human 1 Human 2 Human 3 Automated Human
MeanCon Lib Mean Con Lib Mean Con Lib Mean Con Lib Mean
200 1 0 0 0 0 0 0 0 0 0 0 0 0 0.0
2 0 0 0 0 0 0 0 0 0 0 0 0 0.0
3 0 0 0 0 0 0 0 0 0 0 1 0.5 0.0
4 0 0 0 0 0 0 0 0 0 0 0 0 0.0
5 0 0 0 0 0 0 0 0 0 0 0 0 0.0
6 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0.5
7 0 2 1 0 2 1 0 2 1 0 2 1 1.0
8 0 2 1 0 2 1 0 3 1.5 0 2 1 1.2
9 0 3 1.5 0 3 1.5 0 3 1.5 0 3 1.5 1.5
10 6 8 7 5 6 5.5 6 7 6.5 5 9 7 6.3
11 0 0 0 0 0 0 0 0 0 0 0 0 0.0
12 0 2 1 0 0 0 0 0 0 0 2 1 0.3
13 0 0 0 0 0 0 0 0 0 0 0 0 0.0
14 0 0 0 0 0 0 0 0 0 0 1 0.5 0.0
15 5 7 6 0 6 3 3 6 4.5 0 6 3 4.5
16 2 5 3.5 1 3 2 2 3 2.5 1 5 3 2.7
17 1 1 1 0 1 0.5 0 1 0.5 0 1 0.5 0.7
18 0 0 0 0 0 0 0 0 0 0 0 0 0.0
19 0 0 0 0 0 0 0 0 0 0 2 1 0.0
20 0 0 0 0 0 0 0 0 0 0 0 0 0.0
21 0 0 0 0 0 0 0 0 0 0 0 0 0.0
22 1 8 4.5 1 6 3.5 2 6 4 1 8 4.5 4.0
23 4 5 4.5 3 5 4 3 5 4 3 5 4 4.2
24 2 4 3 2 2 2 2 3 2.5 2 4 3 2.5
25 2 6 4 2 3 2.5 2 5 3.5 2 6 4 3.3
26 1 3 2 0 1 0.5 1 3 2 0 3 1.5 1.5
27 0 3 1.5 0 0 0 0 2 1 0 3 1.5 0.8
28 1 3 2 1 2 1.5 1 3 2 1 3 2 1.8
29 1 1 1 1 1 1 1 1 1 1 1 1 1.0
30 0 8 4 0 3 1.5 0 6 3 0 11 5.5 2.8
31 5 7 6 4 6 5 5 6 5.5 4 7 5.5 5.5
32 0 0 0 0 0 0 0 0 0 0 0 0 0.0
33 0 4 2 0 2 1 0 3 1.5 0 4 2 1.5
34 1 1 1 0 1 0.5 1 1 1 0 1 0.5 0.8
35 0 3 1.5 0 2 1 0 3 1.5 0 3 1.5 1.3
36 0 0 0 0 1 0.5 0 0 0 0 0 0 0.2
37 0 2 1 0 1 0.5 0 2 1 0 2 1 0.8
38 4 4 4 2 5 3.5 4 5 4.5 2 4 3 4.0
39 1 2 1.5 0 1 0.5 1 1 1 0 2 1 1.0
40 1 1 1 0 1 0.5 0 1 0.5 0 1 0.5 0.7
41 9 13 11 7 11 9 7 13 10 6 11 8.5 10.0
42 0 0 0 0 0 0 0 0 0 0 0 0 0.0
43 2 5 3.5 0 2 1 1 3 2 0 5 2.5 2.2
Total 49 114 81.5 29 80 54.5 42 98 70 28 119 73.5 68.7
Notes: Three human experts performed both liberal and conservative ngF population estimates from 42 microscopy images of a single human ovary taken at 200×
magnification. The average of the averages of these counts are given in the final column; the average of the automated estimates are given in the penultimate column.
Abbreviations: Mag, magnification; Con, conservative estimate; Lib, liberal estimate.
Assuming an even distribution of NGFs throughout the
ovary, the full population is then estimated using solutions
of the corpuscle problem for 3-dimensional specimens. This
process is time consuming, and suffers from human mis-
classification, integration error due to small sample sizes,
and the inconsistent assumption of even distribution. In his
seminal paper from 1952,4 Block provided the motivation
for this study:
The distribution of these follicles in human ovaries is so
uneven that reliable values cannot be obtained until all the

Pathology and Laboratory Medicine International 2010:2 submit your manuscript | www.dovepress.com
Dovepress
Dovepress
101
PcnA improves image analysis of human ovaries
follicles are counted. This requires complete serial sectioning,
which for a woman of fertile age means one thousand five
hundred to two thousand five hundred 20 micron sections per
ovary. Under these circumstances any large-scale investiga-
tion is impracticable....
Our aim is to automate part of the estimation process, so
that the use of modern image preparation and analysis tools
and techniques can reduce the human workload involved in
more accurate ovarian reserve studies. We report a combined
process of tissue staining and automatic feature detection,
which gives results comparable to human counts. Our
process works at low magnifications (thereby reducing the
number of images needed per section), and can, in principle,
be used to obtain almost exact NGF populations from fully
sectioned ovaries.
Material and methods
We studied biopsies of tissue from three intact ovaries (post-
oophorectomy) serially sectioned, obtained after routine
surgery for cancer patients. None of the subjects had cancer
of the ovary. The ages of the patients for whom oophorectomy
was performed were 12, 18, and 20 years. Ovarian tissue
was received unfixed from theatre, and was on the same day
fixed in buffered formalin for between 24 and 48 hours and
embedded in paraffin. At a later date, the ovaries were sec-
tioned into 10–12 slices, from which 5 microns thick slide
tissue was obtained using a Microtomo knife (Leitz GmbH
and Co KG, Baden-Wurtemberg, Germany).
Immunohistochemistry
Proliferating Cell Nuclear Antigen (PCNA) was used as
the primary stain, in line with a successful study on rodent
ovaries.6 Our tissue preparation methods differ from this
study only in that we counterstained with hematoxylin for
60 seconds rather than three minutes and we used 1:100
dilution of PCNA (instead of 1:400) as recommended by the
stain supplier (BioCare Medical LLC, California, USA). The
preparation sequence was
1. Dis-paraffination and hydration
2. Heat induced Antigen retrieval for 60 minutes with
tris(hydroxymethyl)aminomethane-ethylenediaminet-
etraacetic acid (4 molar TRIS-EDTA) buffer solution
(pH 9)
3. Wash in distilled water 10 minutes followed by buffer
wash (0.1 molar phosphate buffered saline (PBS))
4. Incubation at room temperature for 60 minutes with pri-
mary antibody (mouse monoclonal PCNA concentrate,
dilution 1:100, clone PC 10 (BIOCARE)) followed by
buffer wash (PBS)
5. Incubation using Dual Link Heat-Stable Protein (HPr)
(DAKO Envision™ (DAKO Denmark, Glostrup, Den-
mark)) for 30 minutes at room temperature followed by
buffer wash (PBS)
6. Application of diaminobenzidine (DAB) chromogen
(DAKO Denmark A/S) for 10 minutes followed by wash
in distilled water
7. Counterstain nuclear-Mayer hematoxylin for 1 minute
followed by 10 minutes under running water
8. Dehydration in alcohol
9. Application of Xylene
10. Mounting on a standard coverslipped slide
External positive controls were performed on tissue from
patients with breast cancer and cancer of the colon.
Image preparation
Slides were viewed using a 1.3 megapixel Infinity 1 camera
(Lumanera Corp., Ottawa, Canada) attached to an Olympus
CX31 microscope (Olympus Imaging Corp., Tokyo, Japan).
Images were captured using the Infinity Analyze software
package supplied with the camera. The default exposure,
hue, saturation, brightness and contrast settings were used.
The white balance was adjusted from a default of Red 2.28,
Green 1.80, Blue 2.69 to Red 1.84, Green 1.84, Blue 3.30.
Light intensity was set to 3.6 for 100× images, and to 4.0 for
200× images. Images were saved as 32-bit-per-pixel RGB
1280 × 1024 pixel TIFF files.
Table 2 summarized comparison of automated counts to human counts for microscopy images
Mag. No.
Images
Human 1 Human 2 Human 3 Automated Human
MeanCon Lib Mean Con Lib Mean Con Lib Mean Con Lib Mean
100 220 191 399 295 182 377 279.5 180 370 275 189 416 302.5 283.2
200 97 70 211 140.5 69 201 135 83 206 144.5 73 230 151.5 140.0
Notes: Three human experts performed both liberal and conservative ngF population estimates from microscopy images of PcnA stained sections from three human ovaries at
200× and 100× magnifications. The average of the averages of these counts are given in the final column; the average of the automated estimates are given in the penultimate column.
Abbreviations: Mag, magnification; No, number of; Con, conservative estimate; Lib, liberal estimate.

Pathology and Laboratory Medicine International 2010:2submit your manuscript | www.dovepress.com
Dovepress
Dovepress
102
Kelsey et al
Image analysis
Our computational methodology varied with the magni-
fication used. We used the ImageJ (National Institutes of
Health, Washington DC, USA) suite of image analysis tools
throughout, making extensive use of the morphology soft-
ware extensions developed by Dr G. Landini of the University
of Birmingham, Birmingham, UK.
For 200× images we: (1) compute the maximum entropy
threshold, (2a) identify regions of restricted size having low
aspect ratio, high modulus ratio, high sphericity, and which
do not contain too much blue, (2b) identify regions that
have some circularity, for which blue average is not low,
and which are not background (ie, green and red kurtosis
are positive), (3) combine the two sets of isolated regions.
Processes 2a and 2b isolate NGF nuclei and zona pellucida
(ZP) respectively. Parts of the image that consist of nucleus
plus ZP are classified as NGFs, as are regions consisting only
of ZP: these are NGFs that have been sliced in an area not
containing the nucleus. Isolated nuclei thus represent false
positives and are discarded.
For 100× images we: (1) compute the triangle entropy, (2)
identify regions of restricted size as in 2a above, (3) filter out
any particles with low compactness and circularity and/or high
aspect ratio (values chosen are liberal, since we apply a color
filter to the survivors), and (4) filter by color: median RGB
must be lower than 70, 60, and 55, respectively (giving a very
dark brown). For both magnifications, we run the code twice
– with liberal and conservative settings – and take the average
as our estimate of the number of NGFs in the image.
ngF counts by hand
Laboratory staff performed two counts for each image. One
conservative (including only those regions of the image that
certainly represented NGFs), and the other liberal (including
both definite NGFs and regions that could equally be either
an NGF or a sectioned blood/lymphatic vessel). These counts
were added together, and the average taken.
Results
We obtained excellent results for both 200× (Figure 1) and
100× images (Figures 2 and 3). For this small sample, the
automatic identification code with conservative settings
consistently agreed to within 5% with the average conserva-
tive human count. With liberal settings the code consistently
agreed to within 10%. Taking the average of these counts
(both human and automatic) to be a good estimate of the
true number of NGFs present, the automated image analysis
count was indistinguishable from averages of expert human
counts, being neither more conservative nor more liberal
than the average human counter. There was wider variance in
population estimates at 100× for both human and automated
image analysis counters.
The automated analysis was, on average, a factor of two
times slower than the time taken by human experts. However,
these timings do not take into account the breaks needed for
a human when analyzing tens of thousands of images. If
we assume that a human can work accurately for less than
12 hours per day, then the automated analysis becomes the
faster method.
Figure 1 Automatic NGF identification in PCNA stained human ovarian tissue (original image taken at 200× magnification) with liberal settings. Panel (a) is the original image.
Panels (b) and (c) show the identification of NGF nuclei by color, size and shape. Panels (d) and (e) show the identification of light areas (either ZP or sectioned blood/lymphatic
vessels), also by color, size and shape. Panel (f) shows the identified NGFs with liberal settings applied: a light area of the correct size and shape is classified as an NGF that has
not been sectioned through the nucleus. human expert estimates for the number of ngFs in this image range from 5 to 8.

Pathology and Laboratory Medicine International 2010:2 submit your manuscript | www.dovepress.com
Dovepress
Dovepress
103
PcnA improves image analysis of human ovaries
Discussion
Ovarian tissue consists of stroma cells, NGFs – consisting
of an oocyte surrounded by zona pellucida (ZP) – and
growing follicles, supported by an extracellular matrix.7
Blood and lymphatic vessels are also present. The standard
stain, HE, hinders computational image analysis even at
high magnification, since sub-regions of NGFs can have
the same color (and size and morphology) as stroma cells
held in the extracellular matrix. Moreover, an obvious can-
didate as a computational technique – color deconvolution
into shades of pink and blue – cannot be fully automated
since HE is a nonstoichiometric stain, and hence a priori
empirical derivation of stain vectors is needed for (at least)
each batch of images.
PCNA, however, stains the nuclei of the stroma cells and
NGFs in shades of brown since these cells are in the G
1
,
S or G
2
interphase stages of cell development. The nuclei of
NGFs are typically stained a darker brown than the nuclei
of the stroma cells, allowing us to differentiate by color
as well as morphology and size. The slight hematoxylin
counterstain that we used gives a blue color to the extra-
cellular matrix, leaving the ZP an almost unstained light
color. This distinction of regions of images by color allows
us to add color differentiation to the size and morphology
Figure 2 Automatic NGF identification in PCNA stained human ovarian tissue (original image taken at 100× magnification) with liberal settings. Panel (a) is the original image.
Panel (b) shows the result of triangle thresholding. Panels (c) through (e) show filtering by size, shape and color respectively. Panel (f) shows 17 identified NGFs with liberal
settings applied. human expert estimates for the number of ngFs in this image range from 14 to 17.
Figure 3 Automatic NGF identification in PCNA stained human ovarian tissue (original image taken at 100× magnification) with conservative settings. Panel (a) is the original
image. Panel (b) shows the result of triangle thresholding. Panels (c) through (e) show filtering by size, shape, and color, respectively. Panel (f) shows 14 identified NGFs with
conservative settings applied. human expert estimates for the number of ngFs in this image range from 14 to 17.

Pathology and Laboratory Medicine International 2010:2submit your manuscript | www.dovepress.com
Dovepress
Dovepress
104
Kelsey et al
attributes that are the only tools available for HE-stained
tissue.
Human identification of NGFs is far from un ambiguous.
A study involving five monkey ovaries8 reported average
populations of 15,735 NGFs, with a standard deviation of
6,214. A study involving 10 rat ovaries6 reported average
NGF populations of 871 (SD, 279) with HE stain, and 1132
(SD, 290) using PCNA. Both studies reported a normal
distribution of estimates, indicating no human tendency to
consistently over- or underestimate the true population. The
problem is therefore the precision of the estimates rather
than their accuracy, and hence averaging multiple counts
is almost certain to be more accurate than a single count.
However, the inherent uncertainty in individual estimates
hinders the reporting of exact results when comparing
human counts with those obtained by a computer program.
Our approach in this study, therefore, was to mimic two
human observers (one conservative, the other liberal) and
estimate the true population as the average of these two
counts. It should be noted, however, that users of the code
who prefer to count, for example, only textbook examples
of NGFs may simply use the conservative settings and take
the results of these counts as their population estimate.
Previous studies have investigated the use of computa-
tional techniques to estimate NGF populations in images of
rodent ovarian tissue. The study on rats6 published in 2008
provides no details on the image analysis performed, but does
refer to it as semiautomatic rather than automatic. A more
recent study8 involving mouse ovaries stained with mouse
vasa homolog (MVH) generates comparable data with con-
ventional methods of NGF counting, and the authors provide
a full description of the imaging techniques used. This study
reports a semiautomatic rather than automatic image analysis
method, noting that light micrographs will differ from dark
micrographs, so that computer settings have to be altered
for each batch of images depending on the intensity of the
staining of the nuclei in that batch. The results obtained
for our study were completely automatic: no human input
was required to adjust settings before the automated counts
(Tables 1 and 2). This could, however, be due to the small
number of ovaries that we examined, and it may well be the
case that, in general, a level of human involvement is needed
to pre-process a batch of images before accurate automatic
counting can proceed.
The main strength of our study is that we have used
human ovarian tissue. The main limitation of this study is
the small number of slides examined, from a small number
of ovaries. Clearly, rodent ovaries are more easily obtainable
than those of human subjects, but our aim is to address the
more important research question of how best to estimate
human ovarian reserve.
Conclusions
To our knowledge, we present the first combination of
PCNA staining combined with fully automated image
analysis to estimate human NGF populations from histo-
logical images. Neither of our methods (for images taken
at 200× and 100×) requires pre-processing before use:
the thresholding automatically gives good results for our
PCNA stained tissue. By running our code on a cluster of
computational nodes, it is entirely feasible to automatically
estimate NGF populations from all the images obtained
from every section of a human ovary. It may be possible
that differences in stain levels across many ovaries, and/or
across multiple laboratories will mean that some human
input is needed to regularize each batch of images, as found
by the recent study involving mice.9 Further research is
needed into this possibility.
Acknowledgments/disclosures
The authors report no conflicts of interest in this work.
TWK is supported by EPSRC grants EP/CS23229/1 and
EP/H004092/1. The funders had no role in study design,
data collection and analysis, decision to publish, or prepa-
ration of the manuscript. We would like to acknowledge
the critical discussions with, and the technical support
from, A Sica, M Cedeira, D Mazal, S de la Peña, S
Bianco, V Carbonati, C Carrera, M Correa, R Gabrielli,
R Nesti and M Cuello. We thank Mr TM Shariah Sazzad
and Ms P Wright for help with the human counting of
NGFs.
References
1. Trelour A. Menstrual cyclicity and the pre-menopause. Maturitas.
1981;3:249–264.
2. Broekemans FJ, et al. A systematic review of tests predicting ovar-
ian reserve and IVF outcome. Hum Reprod Update. 2006;12:
685–718.
3. Wallace WHB, Kelsey TW. Human Ovarian Reserve from Conception
to the Menopause. PLoS ONE. 2010:5(1):e8772.
4. Block E. Quantitative morphological investigations of the follicular
system in women: variations at different ages. Acta Anat (Basel).
1952;14:108–123.
5. Block E. A quantitative morphological investigations of the follicular
system in newborn female infants. Acta Anat (Basel). 1953;17:
201–206.
6. Picut CA, Swanson CL, Scully KL, Roseman VC, Parker RF,
Remick AK. Ovarian follicle counts using proliferating cell nuclear
antigen (PCNA) and semi-automated image analysis in rats. Toxicol
Pathol. 2008;5:674–679.

Pathology and Laboratory Medicine International
Publish your work in this journal
Submit your manuscript here: http://www.dovepress.com/pathology-and-laboratory-medicine-international-journal
Pathology and Laboratory Medicine International is a peer-reviewed,
open access journal focusing on innovative basic research and
translational research related to pathology or human disease. The
journal includes original research, updates, case reports, reviews
and commentaries on current controversies. The Academic Sponsor
of this journal is the Chinese American Pathology Association
(CAPA). The manuscript management system is completely
online and includes a very quick and fair peer-review system.
Visit http://www.dovepress.com/testimonials.php to read real quotes
from published authors.
Pathology and Laboratory Medicine International 2010:2 submit your manuscript | www.dovepress.com
Dovepress
Dovepress
Dovepress
105
PcnA improves image analysis of human ovaries
7. Robboy SJ, Mutter GL, Prat J, Bentley RC, Russell P, Anderson MC,
editors. Robboy’s Pathology of the Female Reproductive Tract, 2nd ed.
Amsterdam: Churchill Livingstone; 2009.
8. Miller PB, Charleston JS, Battaglia DE, Klein NA, Soules MR. An
Accurate, Simple Method for Unbiased Determination of Primordial Fol-
licle Number in the Primate Ovary. Biol Reprod. 1997;56:909–915.
9. Skodras A, Giannarou S, Fenwick M, Franks S, Stark J, Hardy K. Object
recognition in the ovary: quantification of oocytes from microscopic
images. Proceedings of the 16th International Conference on Digital
Signal Processing; 2009 Jul 5–7; Santorini, Greece, p. 345–350.