Finding cells, finding molecules, finding patterns

Abstract

Many modern molecular labeling techniques result in bright point signals. Signals from molecules that are detected directly inside a cell can be captured by fluorescence microscopy. Signals representing different types of molecules may be randomly distributed in the cells or show systematic patterns indicating that the corresponding molecules have specific, non-random localizations and functions in the cell. Assessing this information requires high speed robust image segmentation followed by signal detection, and finally pattern analysis. We present and discuss this type of methods and show an example of how the distribution of different variants of mitochondrial DNA can be analyzed.
Accepted for publication in International Journal of Signal and Imaging Systems Engineering (IJSISE), 2006 (http://www.inderscience.com/browse/index.php?journalID=185)

Full text

Page 1

Int. J. Signal and Imaging Systems Engineering, Vol. 1, No. 1, 2008 11
Copyright © 2008 Inderscience Enterprises Ltd.
Finding cells, finding molecules, finding patterns
Carolina Wählby*
Department of Genetics and Pathology,
Centre for Image Analysis,
Uppsala University, Sweden
E-mail: Carolina.Wahlby@cb.uu.se
*Corresponding author
Patrick Karlsson
Centre for Image Analysis,
Uppsala University, Sweden
E-mail: Patrick.Karlsson@cb.uu.se
Sara Henriksson, Chatarina Larsson
and Mats Nilsson
Department of Genetics and Pathology,
Uppsala University, Sweden
E-mail: Sara.Henriksson@genpat.uu.se
E-mail: Chatarina.Larsson@genpat.uu.se
E-mail: Mats.Nilsson@genpat.uu.se
Ewert Bengtsson
Centre for Image Analysis,
Uppsala University, Sweden
E-mail: Ewert.Bengtsson@cb.uu.se
Abstract: Many modern molecular labelling techniques result in bright point signals.
Signals from molecules that are detected directly inside a cell can be captured by fluorescence
microscopy. Signals representing different types of molecules may be randomly distributed in
the cells or show systematic patterns, indicating that the corresponding molecules have specific,
non-random localisations and functions in the cell. Assessing this information requires high
speed robust image segmentation followed by signal detection, and finally, pattern analysis.
We present and discuss these types of methods and show an example of how the distribution
of different variants of mitochondrial DNA can be analysed.
Keywords: mass data analysis; image analysis; cytometry; single molecule detection; padlock
probes; pattern analysis.
Reference to this paper should be made as follows: Wählby, C., Karlsson, P., Henriksson, S.,
Larsson, C., Nilsson, M. and Bengtsson, E. (2008) ‘Finding cells, finding molecules,
finding patterns’, Int. J. Signal and Imaging Systems Engineering, Vol. 1, No. 1, pp.11–17.
Biographical notes: Carolina Wählby at Uppsala University received her MSc in Molecular
Biotechnology 1998 and a PhD in Image Analysis 2003. She is currently an Assistant Professor
at the Centre for Image Analysis and part time researcher at the Department of Genetics and
Pathology, both at Uppsala University. Her interests include development of algorithms
for applied digital image cytometry, and she has 13 international reviewed publications
in this field.
Patrick Karlsson is currently working towards his PhD Degree in Applied Digital Image
cytometry at the Centre for Image Analysis. He received his MSc Degree in Engineering Physics
from Uppsala University in 2002. He is a member of IAPR.
Sara Henriksson has a Masters Degree in Chemical Engineering and she is currently a PhD
student at the Department of Genetics and Pathology, Uppsala University. Her research interests
include design and optimisation of padlock probes for in situ applications.

Page 2

12 C. Wählby et al.
Chatarina Larsson is currently a PhD student at the Department of Genetics and Pathology,
Uppsala University. Her research interests include in situ genotyping of single DNA molecules
using padlock probes, and she has three international reviewed publications in this field.
Mats Nilsson received his PhD in Medical Genetics in 1998 and visited Leiden University as a
Post Doctoral Fellow in 1999–2000. He is an Associate Professor at the Department of Genetics
and Pathology, Uppsala University. His research aims at developing methods for single-cell and
single-molecule analysis. The research has resulted in more than 30 peer-reviewed original
papers, several book chapters, and four patents. In 2006 he received the Feulgen Prise from the
Society for Histochemistry. He is supervisor for seven PhD students and two post doctoral
fellows, and he is on the Board of Directors for Olink AB.
Ewert Bengtsson received his MSc and PhD in Engineering Physics from Uppsala University in
1974 and 1977, respectively. He continued the work on cell image analysis from his thesis, first
as a researcher at Uppsala University and then in spin-off companies, from 1983–1988.
He returned to Uppsala University as Adjunct Professor and established the centre for image
analysis where, since 1995, he is active as a full Professor. His research interests include all kinds
of biomedical image analyses. He has published over 120 papers in the field and supervised more
than 25 PhD students.

1 Introduction
Data mining can be defined as the science of extracting
useful information from large data sets. In this case, the
input data are digital images of cells captured using
fluorescence microscopy, and the information we aim
to retrieve is that of spatial distribution patterns of
different variants of fluorescence labelled molecular targets.
New probing and staining techniques allow a large variety
of molecular targets to be visualised in situ and imaged by
fluorescence microscopy. Biological processes can be
studied at the ultimate level of single molecules, and with
sufficient precision to distinguish even closely similar
variants of molecules. It is thus possible to study the
inter- or sub-cellular context of molecules that otherwise
may go undetected at the level of populations of molecules
and cells. At the same time, large numbers of cells have to
be analysed to retrieve statistically significant information.
Extracting information from the resulting image data will,
therefore, require efficient and robust cell segmentation as
well as signal detection and finally, pattern analysis.
Before signals can be assigned as coming from
a particular cell, each cell has to be delineated.
Segmentation is the process in which an image is divided
into its constituent objects, or parts, and background. Cells
can be visualised in many different ways, using different
kinds of probes or stains that bind to structures within a cell.
It is therefore difficult to define a single algorithm that will
always find the individual cells in an image, independent of
method for visualisation and cell morphology. Instead, cell
image segmentation can be seen as a modelling problem
where different approaches more or less explicitly are based
on models of the cells. For example, thresholding methods
can be seen as being based on a model stating that cells have
an intensity that is different from the surroundings. More
robust segmentation can be obtained if a combination of
features, such as intensity, edge gradients, and cellular
shape, is used.

In many applications in cell biology, where fluorescence
marked probes are applied, the resulting images are
composed of signals, seen as spots of different shapes and
intensities. The localisation of these regions can yield
important biological information. In multiple labelling
experiments in particular, measurements of relative
positions of regions labelled with different marker
molecules can provide insight into the functional
relationship between organelles and/or processes. Visual
inspection is, apart from being tedious, beset with various
sources of error. The positions of signals in an image should
be determined automatically to derive objective information
and allow further extraction of image information, such as
signal intensity distribution, relative positioning and pattern
analysis. The human mind is exceptional at finding patterns;
it will even find patterns in data that are completely random.
It is, therefore, valuable to have computerised methods that
can search for patterns in a more objective way.
We present an image based data mining example where
the distribution of different variants of the genetic
information contained in mitochondria (i.e., mtDNA) has
been examined. MtDNA is present in multiple copies in
the mitochondrion of the cell. It is inherited together with
the cytoplasm during cell replication. Genetic diseases are
often caused by mutations where one single nucleotide has
been substituted by another, a so-called point mutation.
To be able to study and diagnose such diseases with limited
material from patients, there is a need for methods to detect
point mutations in situ. Padlock probes and Rolling Circle
Amplification (RCA) combine highly specific target
sequence recognition with a high signal-to-noise ratio.
Padlock probes have been successfully used for detecting
point mutations in mitochondrial DNA, by Larsson
et al. (2004). We combine cell segmentation, padlock
probes, signal detection and pattern analysis to examine the
distribution of mtDNAs. These types of methods could also
be used in applications ranging from detection of infectious
organisms to studies of tumours.

Page 3

Finding cells, finding molecules, finding patterns 13
2 Methods
The methods section is divided into three parts, describing
methods for segmentation of cells, detection of signals, and
evaluation of patterns in the detected signal distribution.
A specific example is thereafter brought up in the
‘Experiments and results’ section.
2.1 Cell segmentation: finding cells
The difficulty of the segmentation problem is highly
dependent on the type of specimen that is to be analysed,
and the result of the segmentation usually determines the
eventual success of the final analysis. If we are dealing with
cytological specimens where the cells are lying singly on a
clean background with well stained nuclei, and if the
analysis task is limited to nuclear properties, then a
simple automatic thresholding method may be sufficient.
Thresholding is often based on histogram characteristics of
the pixel intensities of the image, see Sahoo et al. (1988).
In order to get a satisfactory segmentation result by
thresholding, a sufficiently uniform background is required.
The transition between object and background may be
diffuse, making an optimal threshold level difficult to find,
even after background correction. At the same time, a small
change in the threshold level may have a great impact on
the further analysis; feature measures such as area
and volume are directly dependent on the threshold.
Adaptive thresholding, i.e., local automatic thresholding,
can be used to circumvent the problem of varying
background or as a refinement to a coarse global threshold,
see Ortiz de Solorzano et al. (1999). The problems of
segmenting clustered objects and choosing a suitable
threshold level for objects with unsharp edges will,
however, remain.
If we model the objects as consisting of connected
regions of similar pixels we obtain region growing methods.
A popular region growing method which has proved to be
very useful in many areas of image segmentation is the
so-called watershed algorithm. The method was originally
suggested by Digabel and Lantuéjoul, and extended to
a more general framework by Beucher and Lantuéjoul
(1979). Watershed segmentation has then been refined and
used in very many situations, see Meyer and Beucher (1990)
and Vincent (1993) for an overview. If the intensity
of the image is interpreted as elevation in a landscape,
the watershed algorithm will split the image into regions
similar to the drainage regions of this landscape. To avoid
over-segmentation, i.e., splitting of the image into too many
regions, water can be allowed to rise only from places
marked as seeds (Beucher, 1992; Landini and Othman,
2003; Lockett et al., 1998; Meyer and Beucher, 1990;
Vincent, 1993). Seeds may be found manually or by
automated methods. Over-segmentation can also be reduced
by rule-based merging, e.g., shown by Najman and
Schmitt (1996).
Cell nuclei are usually convex and fairly round or
elliptic and the shape can, therefore, be used as part of the
object model. Touching nuclei that are not separated by an
intensity threshold can be separated by distance
transforming (Borgefors, 1986) the binary image and
applying watershed segmentation, see work done by
Malpica et al. (1997), Ortiz de Solorzano et al. (1999) and
Wählby et al. (2004).
None of the above described methods will alone
produce a satisfactory result on the more difficult types
of cell and tissue images. We may, for instance, have
problems if
•
the cells are clustered
•
the image background is varying
•
there are intensity variations within the cells.
By combining the methods, more powerful models can be
created, and more complex segmentation problems solved.
Our experience is that the seeded watershed approach is
a useful core component in such segmentation models.
Complex segmentation methods often require a large
number of input parameters that have to be optimised for
each type of input data. In case based reasoning, the
segmentation step is initiated by classifying each image as
belonging to one of a number of pre-defined cases,
and input parameters optimised for the particular case are
applied during segmentation, see Perner (1999).
2.2 Signal detection: finding molecules
The most common method for finding structures such as
proteins and organelles in situ is using antibodies
labelled with fluorescing molecules. Fluorescence labelled
secondary antibodies can be used to amplify the signal and
increase signal to noise ratios. The genetic information
contained in the DNA in a cell can be stained as a whole
using non-specific chemical dyes, or in a more specific way
using oligonucleotide probes that search for a particular
DNA sequence. Fluorescence In Situ Hybridisation (FISH)
is such a method, and can detect larger mutations such as
duplications, translocations and deletions, but it is not
sensitive enough to distinguish between single nucleotide
sequence variations. Primed In Situ Labelling (PRINS)
reaction uses a specific primer that will initiate synthesis of
DNA from fluorescent labelled nucleotides at the site of
sequence detection, see Koch et al. (1988). The method does
not, however, give signals from single-copy genes that are
distinguishable from noise caused by insertions of
fluorescing nucleotides in other places in the genome.
In the Oligonucleotide Ligation Assay (OLA), as shown by
Landegren et al. (1988), oligonucleotides are hybridised
juxtaposed with the junction at the point mutation. If there is
a perfect match, the two probes can be enzymatically
hybridised and detected. There is, however, a risk of wrong
probes being ligated, especially when trying to find many
different sequence variants in the same sample. This can be
avoided by, instead of using two separate probes, using a
single linear probe, a so called padlock probe. The padlock
probe has ends that are designed to hybridise juxtaposed
at the point mutation and if correctly base paired at the point

Page 4

14 C. Wählby et al.
mutation, the two ends can be enzymatically ligated,
forming a circular DNA molecule, see Nilsson et al. (1994).
The specifically reacted circular DNA can thereafter be
amplified using RCA generating molecules that are bound
by hundreds of fluorescing probes, see Banér et al. (1998).
These signals can be detected by fluorescence microscopy
as bright spots at, or below, the resolution of the
microscope, the image resolution being limited by the point
spread function of the microscope.
An image that contains multiple, and sometimes
clustered spots with different maximum intensities, can be
segmented in many different ways. Regions found by
procedures such as intensity thresholding often contain
more than one local maximum of intensity, indicating that
the region consists of more than one spot. Top-hat
transforms, see Haralick and Shapiro (1992), in combination
with threshold procedures fail to divide the image into
separate domains, each containing one local maximum of
intensity, as the top-hat transform is unable to distinguish
a local maximum from a saddle-point. If each spot contains
a single local maximum, watershed segmentation, as
described above, in combination with a background
threshold, may be used to delineate individual signals.
Another approach is the largest contour segmentation
by Manders et al. (1996), where the domain of each signal
is defined by a local maximum and an iterative
region-growing. If two or more signals are clustered into
a spatially large signal, where the individual signals do not
contain individual intensity maxima (due to tight clustering
or signal saturation), the shape of the signal can provide
clues as to how the signals should be detected. In the work
by Karlsson and Lindblad (2004), the curvature of the edge
of each signal cluster is examined, and signals are
positioned within the cluster starting from the position
where the greatest curvature is found.
2.3 Hypotheses testing: finding patterns
Patterns in image data can be evaluated by interpreting
the signal distribution as image texture, and by using
texture measurements. Some of the most commonly
used texture measures are derived from the Grey Level
Co-occurrence Matrix (GLCM). The GLCM is a tabulation
of how often different combinations of pixel brightness
values (grey levels) occur in a pixel pair in an image,
see Haralick and Shapiro
discretisation has also shown to give useful information for
cell pattern classification, see Perner et al. (2002). Different
kinds of distance measures can also be used to evaluate
spatial relationships between signals once they have been
detected. In the case where we want to know whether
red and green signals are randomly distributed in the
cytoplasm or not, we can simply count how often a red
signal has a green signal as its closest neighbour,
and how often a red signal has a red signal as its closest
neighbour (and the other way around). To evaluate the
outcome, we have to know what distributions can be
expected. By creating a virtual cell, where possible positions
and number of signals of different types are given as inputs,
(1992). Multi-interval
different hypotheses can be tested. We can then compare the
spatial relationships between signals in real cells with those
in a virtual cell with the same input parameters. Signals in
the virtual cell can be positioned based on a hypothesis,
i.e., either randomly, or according to a pattern. Thousands of
randomised virtual cells can thereafter be created, and the
probability of the real cell having the hypothesised signal
distribution pattern can be examined. Factors such as
staining efficiency and noise may also be added to the
virtual cell for comparison.
3 Experiments and results
To illustrate the concepts discussed we will, here, describe
a project where model based cell segmentation is combined
with padlock probing for molecule detection, model based
signal detection and pattern analysis, to examine the spatial
distribution of mtDNAs.
3.1 Finding cells
In the presented experiment, no general stain defining the
cytoplasm is available. We do, however, have a general
stain defining the nucleus of each cell. Combining this
information with the fact that the over all signal variance is
higher within the cytoplasm than in the image background,
a model defining cytoplasms is created. The three markers
(i.e., nuclear stain, padlock probe 1, and padlock probe 2)
are shown as three images, see Figure 1(a)–(c), each
captured with a different filter set in the fluorescence
microscope. Cells segmentation is initiated by intensity
thresholding of the image showing the nuclear stain.
A suitable threshold is found using Otsu’s method, which
searches for the threshold level that minimises the
intra-class variance of foreground as well as background,
see Otsu (1979). The resulting binary image is shown in
Figure 1(d). Intensity thresholding is not enough to separate
nuclei that are very close to one another. Thanks to their
round shape, touching nuclei can be separated by applying
watershed segmentation to a distance map of the binary
image. Over-segmentation due to multiple local maxima is
avoided by smoothing of the distance map. The distance
map is shown in Figure 1(e), and the result after watershed
segmentation is shown in Figure 1(f). The region
surrounding each nucleus, not belonging to the image
background, is the cytoplasm. The image background has
less variance than the parts of the image containing cells,
and can thus be found by variance filtering. The variance
map of the sum of B and C (Figure 1(g)) is thresholded
using Otsu’s method. Small, disconnected regions
are removed by morphological opening, and larger
disconnected regions are re-connected by dilation. In both
cases, a disc-shaped structure element of radius 10 pixels
was used. The choice of structure element depends on size
and density of signals.
Each pixel of the resulting binary image (Figure 1(h)) is
assigned to the closest nucleus by seeded watershed
segmentation, using the segmentation result from the
segmentation of the nuclei as seeds. Non-seeded regions

Page 5

Finding cells, finding molecules, finding patterns 15
are discarded as background. Figure 1(i) shows the final
segmentation result on top of a projection of the three input
images.
Figure 1 (a) Part of an image of DAPI stained cell nuclei;
(b) image of the same cells showing logarithm of
signals from padlock probe 1 (stained with Cy3);
(c) logarithm of signal from padlock probe 2 (stained
with FITC); (d) binary image after thresholding of a;
(e) shows the same image after distance
transformation; (f) is the result after watershed
segmentation on the distance map, i.e., the final
segmentation of the nuclei; (g) the cytoplasm is found
by combining the two images showing padlock probes
(b and c) and applying a variance filter; (h) potential
cytoplasm after thresholding of variance map and
morphological opening to remove noise and (i) final
segmentation result based on shape of nuclei and
variance of cytoplasm

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)
3.2 Finding molecules
In the presented experiment, a model system with padlock
probes was used. It consists of different detection sequences
that represent real point mutations. Four different padlock
probes were used for testing efficiency of staining and
evaluation of signal distribution patterns. Two of the
padlock probes hybridise to different sites on the same
mtDNA fragment, i.e., they are non-competing. One is
detected using Cy3 (red), and one using FITC (green).
The other two probes bind to the same site, and are,
therefore, competing.
Signal detection was initiated by first reducing the
background variation present in the images. As the cells are
cultured on a glass surface, they are comparably flat.
Despite this, it is necessary to image them in more than one
focal plane to make sure that all signals are detected.
In the presented study, the slides were studied in
a fluorescence microscope (Axioplan II Zeiss) using a
63x objective. Images were collected with Axiovision 4.3
software as a 16 layers z-stack with 0.5 um between
consecutive layers. The nuclear stain DAPI emission was
collected at 360 nm excitation wavelength for 200 ms, green
padlock detection fluorochrome FITC at 470 nm for 200 ms
and red padlock detection fluorochrome Cy3 at 546 nm
for 450 ms. The background was reduced in each z-image
separately by morphological tophat filtering, using a
disc of radius 10 pixels. Tophat filtering removes intensity
variations that have a spatial extent greater than that
of the disc. The 16 layers were thereafter combined using
maximum intensity projection. Projection of the 3D
information to a single image will result in loss of spatial
information in the z-direction. As the extension of cultured
cells in z-direction is only a fraction of their extension i
x-, and y- directions, the 3D information was considered less
important. This would, however, not be the case if cells in
the tissue were observed.
The result after pre-processing and maximum intensity
projection of a small fraction of an image is shown in
Figure 2(a). Simple intensity thresholding will separate the
signals from the image background, but signals that are
clustered will not be separated from each other. To separate
clustered signals, Watershed segmentation, starting from all
local maxima, is applied to the image, and the watershed
regions are allowed to extend until they reach a predefined
background threshold. The resulting signal centres after
watershed segmentation of Figure 2(a) are shown in
Figure 2(b).
Figure 2 (a) Enlarged image showing signals from padlock
probes after image pre-processing and (b) positions
of detected signals using watershed segmentation

(a)

(b)
3.3 Finding patterns
The patterns of red and green signal distributions were
examined by searching for aggregations of signals, i.e., the
existence of groups of signals with the same colour.
The affinity of red and green signals was measured as the
number of red signals with a green nearest neighbour,
the number of red signals with a nearest red neighbour, the
number of green signals with a nearest green neighbour, and
finally the number of green signals with a nearest red
neighbour. To normalise the observed result and evaluate
the probability of non-random pattern, virtual cells with
truly random patterns were created. Virtual cells with

Page 6

16 C. Wählby et al.
random signal distributions were created by keeping the
number of red and green signals the same as in the real cell.
Red and green signals were then randomised within an
area corresponding to the cytoplasm region with the 10%
greatest variance. Virtual cells were re-created 1000 times,
neighbourhood relations were examined, and the resulting
distributions were compared with observed distributions.
A number of restrictions have to be taken into
consideration when creating the virtual cells. First of all
there is a limit in closeness between signals in the real data
due to the point-spread function. Two signals that are of the
same colour will not be separated if they are closer than the
width of a single signal. This has to be compensated for in
the randomised data, or else it will affect the outcome of the
analysis of the neighbour relations. Randomised signals that
appeared closer to one another than the two closest signals
in the real data were simply removed, and a new pair of
random signals was created and tested for closeness with the
existing randomised signals. Figure 3(a), top, shows the true
signal distribution within the cytoplasm of a cell with
competing padlock probes, red and green signals as + and o
respectively. Figure 3(a), bottom, shows one of the 1000
virtual cells with randomised signals. Figure 3(b), top,
shows the true signal distribution within the cytoplasm of a
cell with non-competing padlock probes, red and green
signals as + and o respectively. Figure 3(b), bottom, shows
one of the 1000 virtual cells with randomised signals in the
non-competing case. As can be seen, it is not trivial to pick
out the cells showing random distributions compared to
those showing a non-random affinity between red and green
signals. Comparing the randomised data with the true
signal distributions shows that the pattern falls within the
randomised distribution in the case with competing probes,
while the non-competing probes show a red-green affinity
three standard deviations greater than that of the randomised
distribution. This agrees with what one would expect
as the non-competing probes can bind to the same mtDNA
fragment.
Figure 3 Cells treated with two competing: (a) or non-competing
and (b) padlock probes in a 50–50 concentration.
Top: true signal distribution, bottom: randomised
signal distribution. Red and green signals as + and o
respectively. Nucleus and cytoplasm outlined

(a)
Figure 3 Cells treated with two competing: (a) or non-competing
and (b) padlock probes in a 50–50 concentration.
Top: true signal distribution, bottom: randomised
signal distribution. Red and green signals as + and o
respectively. Nucleus and cytoplasm outlined
(continued)

(b)
4 Conclusions
Patterns and spatial relationships between molecules in cells
are of great interest in many types of analysis. One way of
examining patterns in cells is by visualising the molecules
of interest using highly specific detection probes, and
comparing observed signal distributions with randomised
distributions in a virtual cell. Before patterns can be
examined, signals from detection probes have to be found,
and clustered signals separated. More than one cell is often
observed simultaneously, and automated identification of
each individual cell in an image provides efficient analysis
of large data sets with little impact from observer bias.
In order to obtain a successful cell segmentation method
it is important to use as much a priori information as
possible about the appearance of the objects that are to be
segmented, without resorting to models that are too complex
or too difficult to train or apply.
References
Banér, J., Nilsson, M., Mendel-Hartvig, M. and Landegren, U.
(1998) ‘Signal amplification of padlock probes by rolling
circle replication’, Nucleic Acid Research, Vol. 26,
pp.5073–5078.
Beucher, S. (1992) ‘The watershed transformation applied
to image segmentation’, Scanning Microscopy, Vol. 6,
pp.299–314.
Beucher, S. and Lantuéjoul, C. (1979) ‘Use of watersheds
in contour detection’, International Workshop on Image
Processing: Real-time and Motion Detection/Estimation,
September, Rennes, France.
Borgefors, G. (1986) ‘Distance transformations in digital images’,
Computer Vision, Graphics and Image Processing, Vol. 34,
pp.344–371.

Page 7

Finding cells, finding molecules, finding patterns 17
Haralick, R.M. and Shapiro, L.G. (1992) Computer and Robot
Vision, Addison-Wesley, Reading, Massachusetts, USA,
Vols. 1–2, pp.1291–1294.
Karlsson, P. and Lindblad, J. (2004) ‘Segmentation and separation
of point like fluorescent markers in digital images’,
Proceedings of 2004 IEEE International Symposium on
Biomedical Imaging, Washington DC, USA, pp.453–470.
Koch, J., Kolvraa, S., Petersen, K., Gregersen, N. and
Bolund, L. (1988) ‘Oligonucleotide-priming methods for the
chromosome-specific labeling of alpha satellite DNA in situ’,
Chromosoma, Vol. 98, pp.259–265.
Landegren, U., Kaiser, R., Sanders, J. and Hood, L. (1988)
‘A ligase-mediated gene detection technique’, Science,
Vol. 241, pp.1077–1080.
Landini, G. and Othman, I.E. (2003) ‘Estimation of tissue layer
level by sequential morphological reconstruction’, Journal
of Microscopy, Vol. 209, No. 2, pp.118–125.
Larsson, C., Koch, J., Nygren, A., Janssen, G., Raap, A.K.,
Landegren, U. and Nilsson, M. (2004) ‘In situ genotyping
individual DNA molecules by target- primed rolling-circle
amplification of padlock probes’, Nature Methods, Vol. 1,
pp.227–232.
Lockett, S.J., Sudar, D., Thompson, C.T., Pinkel, D. and
Gray, J.W. (1998) ‘Efficient,
three-dimensional segmentation of cell nuclei in thick
tissue sections’, Cytometry, Vol. 31, pp.275–286.
Malpica, N., Ortiz de Solorzano, C., Vaquero, J.J., Santos, A.,
Vallcorba, I., Garcia Sagredo, J.M. and del Pozo, F. (1997)
‘Applying watershed algorithms to the segmentation of
clustered nuclei’, Cytometry, Vol. 28, No. 4, pp.289–297.
Manders, E.M.M., Hoebe, R., Strackee, J., Vossepoel, A.M. and
Aten, J.A. (1996) ‘Largest contour segmentation: a tool for
the localization of spots in confocal images’, Cytometry,
Vol. 23, pp.15–21.
Meyer, F. and Beucher, S. (1990) ‘Morphological segmentation’,
Journal of Visual Communication and Image Representation,
Vol. 1, No. 1, pp.21–46.





















interactive, and
Najman, L. and Schmitt, M. (1996) ‘Geodesic saliency of
watershed contours and hierarchial segmentation’, IEEE
Transactions on Pattern Analysis and Machine Intelligence,
Vol. 18, No. 12, pp.1163–1173.
Nilsson, M., Malmgren, H., Samiotaki, M., Kwiatkowski, M.,
Chowdhary, B.P. and Landegren, U. (1994) ‘Padlock probes:
circularizing oligonucleotides for localized DNA detection’,
Science, Vol. 265, pp.2085–2088.
Ortiz de Solorzano, C., Garcia Rodriguez, E., Jones, A., Pinkel, D.,
Gray, J.W., Sudar, D. and Lockett, S.J. (1999) ‘Segmentation
of confocal microscope images of cell nuclei in thick tissue
sections’, Journal of Microscopy, Vol. 193, pp.212–226.
Otsu, N. (1979) ‘A threshold selection method from gray-level
histograms’, IEEE Trans. on System Man and Cybernetics,
Vol. 9, No. 1, pp.62–69.
Perner, P. (1999) ‘An architecture for a CBR image segmentation
system’, Journal on Engineering Application in Artificial
Intelligence, Engineering
Intelligence, Vol. 12, No. 6, pp.749–759.
Perner, P., Perner, H. and Müller, B. (2002) ‘Mining knowledge
for HEp-2 cell image classification’, Journal Artificial
Intelligence in Medicine, Vol. 26, pp.161–173.
Sahoo, P.K., Soltani, S., Wong, A.K.C. and Chen, Y.C. (1988)
‘A survey of thresholding techniques’, Computer Vision,
Graphics and Image Processing, Vol. 41, pp.233–260.
Vincent, L. (1993) ‘Morphological grayscale reconstruction in
image analysis: applications and efficient algorithms’,
IEEE Transactions on Image Processing, Vol. 2, No. 2,
pp.176–201.
Wählby, C., Sintorn, I-M., Erlandsson, F., Borgefors, G. and
Bengtsson, E. (2004) ‘Combining intensity, edge, and shape
information for 2D and 3D segmentation of cell nuclei on
tissue sections’, Journal of Microscopy, Vol. 215, No. 1,
pp.67–76.
Applications of Artificial