FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN

TISSUE CONFOCAL IMAGES

Comparison between a Custom Unsupervised and a Supervised SVM Approach

Santa Di Cataldo, Elisa Ficarra and Enrico Macii

Dep. of Control and Computer Engineering, Politecnico di Torino, Cso Duca degli Abruzzi 24, 10129, Torino, Italy

Keywords: Tissue segmentation, tissue confocal images, immunohistochemistry, K-means clustering, Support Vector

Machine.

Abstract: In this paper we present a fully-automated method for the detection of tumor areas in immunohistochemical

confocal images. The image segmentation provided by the proposed technique allows quantitative protein

activity evaluation on the target tumoral tissue disregarding tissue areas that are not affected by the pathol-

ogy, such as connective tissue. The automated method, that is based on an innovative unsupervised cluster-

ing approach, enables more accurate tissue segmentation compared to traditional supervised methods that

can be found in literature, such as Support Vector Machine (SVM). Experimental results conducted on a

large set of heterogeneous immunohistochemical lung cancer tissue images demonstrate that the proposed

approach overcomes the performance of SVM by 8%, achieving on average an accuracy of 90%.

1 INTRODUCTION

Detecting tumor areas in cancer tissue images and

disregarding non pathological portions such as con-

nective tissue are critical tasks for the analysis of

disease state and dynamics. In fact, by monitoring

the activity of proteins involved in the genesis and

the development of multi-factorial genetic patholo-

gies we can obtain a useful diagnostic tool. It leads

to classify the pathology in a more accurate way

through its particular genetic alterations, and to cre-

ate new opportunities for early diagnosis and per-

sonalized predictive therapies (Taneja et al., 2004).

An approach for monitoring and quantifying the

protein activity in pathological tissues is to analyze,

for example, images of the tissue where the localiza-

tion of proteins is highlighted by fluorescent marked

antibodies that can detect and link the target pro-

teins. The antibodies are marked with particular

stains whose intensity is related to protein activity

intensity. This procedure is called immunohisto-

chemistry (IHC).

The increased use of immunohistochemistry

(IHC) in both clinical and basic research settings has

led to the development of techniques for acquiring

quantitative information from immunostains and

automated imaging methods have been developed in

an attempt to standardize IHC analysis.

Tissue segmentation for tumor areas detection is

the first fundamental step of automated IHC image

processing and protein activity evaluation. In fact

the quantification of a target protein activity should

be performed on tumor portions of the tissue without

taking into account the non pathological areas even-

tually present in the same IHC images. In Figure 1

are reported examples of IHC tissue images where

connective tissue (i.e. non tumoral tissue) is outlined

in black (for details about these images see Section

2).

Several methods have been proposed in the last

few years to perform automated segmentation of

tissue images (Demandolx et al., 1997; Nedzved et

al., 2000; Malpica et al., 1997; Dybowzki, 2000;

Nattkemper 2004). However the most accurate ap-

proaches are those that provide a well-suited frame-

work for incorporating primary expert knowledge

into the adaptation of algorithms, such as supervised

learning algorithm (e.g. Neural Networks, Machine

Learning, kernel-based) (Nattkemper 2004). The

most prominent algorithm among these is the sup-

port vector machine (SVM) proposed by V.Vapnik

(V.Vapnik, 1998) for binary classification. SVM is a

theoretically superior machine learning method

which has often been shown to achieve great classi-

116

Di Cataldo S., Ficarra E. and Macii E. (2008).

FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN TISSUE CONFOCAL IMAGES - Comparison between a Custom Unsupervised and a

Supervised SVM Approach.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 116-123

DOI: 10.5220/0001068501160123

 SciTePress

fication performance compared to other learning

algorithms across most application fields and tasks,

including image processing and tissue image proc-

essing in particular (Angelini et al., 2006; Osuna,

1997). Moreover, the SVM method is more able to

handle very high dimensional feature spaces than

traditional learning approaches (Muller et al., 2001;

Cai, 2003). This is in fact the case of the images

targeted by our work.

However, the IHC tissue images we considered

in our study present an intrinsic complexity, such as

very different characteristics of staining, intensity

distribution, considerable variation of tissue shape

and/or size and/or orientation and, finally, consider-

able variation of the signal intensity within the same

tissue areas due for example to superimposed stain-

ing.

Because of the heterogeneity of the representa-

tive features related to each tissue, it is very difficult

for the supervised methods to obtain a satisfying

fixed classifier able to distinguish between tumor

areas (i.e. epithelial tissue) and non cancerous tissue

portions (such as connective tissue).

For this reason we designed a fully-automated

unsupervised approach that is based only on the

characteristics of the input image rather than on a

fixed model of the ground truth.

In this paper we present our fully-automated un-

supervised method and we compare its performance

to that provided by a SVM approach applied on the

same IHC tissue image target. We demonstrate that

our method enables more accurate tissue segmenta-

tion compared with traditional SVM. Experimental

results conducted on a large set of heterogeneous

immunohistochemical lung cancer images are re-

ported and discussed in Section 4. In Section 2 we

detail our fully-automated unsupervised method and

we briefly introduce the SVM method. The imple-

mentation and the set-up are discussed in Section 3.

Finally, the Conclusions are reported in Section 5.

2 METHOD

The images we analyzed in this work were acquired

through high-resolution confocal microscopy and

show lung cancer tissue cells stained with marked

antibodies (see Figure 1). They are characterized by

a blue hematoxylin stain as a background colour and

a brown DAB stain in cellular regions where a re-

ceptor of the EGF-R/erb-B or TGF-alpha family is

detected (i.e. membranes or cytoplasm, respec-

tively). Cellular nuclei are blue-coloured and show a

staining intensity darker than background.

In all the images a remarkable portion of connec-

tive or other no cancer tissue components is present,

which appears as a blue-coloured mass (since brown

DAB-stained cells are only in cancerous tissue) with

quite well-defined borders. Connective tissue is usu-

ally characterized by shorter inter-cellular distances

and smaller nuclei than epithelial component; how-

ever, a generalization of this remark is impossible

because shape and dimensions distributions of can-

cer cells are often not predictable. As we outlined in

the Introduction, in order to perform accurate and

robust cell segmentation and protein activity quanti-

fication (Ficarra, 2006) these non cancerous tissue

portions have to be identified and isolated from the

representative epithelial tissue. Here we present two

different segmentation approaches to perform this

critical task: i) an unsupervised procedure based on a

K-means clustering of brown intensities followed by

some morphological and edge-based refinement

steps (see Figure 3); ii) a supervised classification of

RGB features through Support Vector Machine (see

Figure 5).

Experimental results obtained with each approach

on the same real-life datasets are presented and

compared in Section 4.

Figure 1: IHC tissue images with connective tissue manually outlined in black (from the left, x400 image with EGF-R posi-

tive reactions; x400 image with EGF-R positive reactions; x200 image with TGF-alpha positive reactions).

FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN TISSUE CONFOCAL IMAGES - Comparison

between a Custom Unsupervised and a Supervised SVM Approach

117

2.1 Unsupervised Procedure

Since non cancerous cells do not show positive reac-

tions at the EGF-R/TGF-alpha receptors, the mono-

chromatic pure-DAB component instead of the

original RGB image can be analyzed to perform

tissue segmentation: in fact in this simpler color

space connective components can be easily identi-

fied as wide bright regions with a quite homogene-

ous appearance (see Figure 2(b)).

An unsupervised learning algorithm (K-means,

in our work) can be efficaciously applied to isolate

bright regions; then areas which show morphologi-

cal and edge characteristics which are typical of

connective tissue can be selected to refine tissue

segmentation.

DAB-COMPONENT

SEPARATION

PREPROCESSING

K-MEANS

CLUSTERING

REFINEMENT BY

SIZE AND

CIRCULARITY

ANALYSIS

REFINEMENT BY

GRADIENT

MAGNITUDE

ANALYSIS

DAB-COMPONENT

SEPARATION

PREPROCESSING

K-MEANS

CLUSTERING

REFINEMENT BY

SIZE AND

CIRCULARITY

ANALYSIS

REFINEMENT BY

GRADIENT

MAGNITUDE

ANALYSIS

Figure 3: Unsupervised procedure based on K-means clus-

tering

Main steps of the proposed procedure are (see

Figure 3):

1) DAB-Component Separation. To separate pure-

DAB from pure-hematoxylin component a color

deconvolution algorithm based on stain-specific

RGB absorption is applied on the original RGB im-

age (Ruifrok 2001, 2004); differently from classical

color segmentation approaches based on transforma-

tion of RGB information to HSI or to another spe-

cific color representation (Brey, 2003), this method

has been demonstrated to perform a good color sepa-

ration even with colocalized stains. This critical

condition, due to chemical reactions of stains linking

the target proteins and to the tissue superposition

during the slicing of samples before image acquisi-

tion, is very common in the images targeted by our

method.

For this step, the free color deconvolution plugin

developed by G. Landini was integrated to our algo-

rithm.

2) Preprocessing. In pure-DAB images, connective

tissue can be differentiated from epithelial tissue

through its higher intensity (see Figure 2(b)); any-

way some preprocessing is needed in order to ho-

mogenize and separate the intensity distributions of

the two tissues, thus improving K-means’ perform-

ance.

First of all, a mean filter is performed: this opera-

tion replaces each pixel value with the average value

in its neighbourhood, thus smoothing intensity peaks

and decreasing the influence of single non-

representative pixels. Then a minimum filter is ap-

plied. The filter replaces pixels values with the

minimum intensity values in their neighbourhood:

this transformation reduces the intensity dynamic

and performs a further separation of connective and

epithelial intensity distributions, since the former

shows minimum values higher than the latter.

3) K-Means Clustering.

To isolate bright pixels

belonging to connective tissue a K-means clustering,

the well-known unsupervised learning algorithm

(Jain, 1988) which iteratively partitions a given

dataset into a fixed number of clusters, is applied;

this iterative partitioning minimizes the sum, over all

clusters, of the within-cluster sums of point-to-

cluster-centroid distances. Thus the procedure

minimizes the so-called objective function, J in

Equation 1, where k is the number of clusters, n is

the number of data points and the quadratic expres-

sion is the distance measure between a data point x

(j)

and the current cluster centroid c

(1)

The cluster with the highest centroid value is se-

lected as representative of the connective tissue (see

)(

∑∑

−=

cxJ

Figure 2: Unsupervised procedure: (a) original IHC image with connective regions manually outlined (in black); (b) pure-

DAB image (c) results after K-means clustering (pixels belonging to different clusters are mapped with grey intensity pro-

portional to the cluster centroid); (d) cluster with highest centroid value (in black); as outlined in section 2.1 point 4, some

small and round-shaped epithelial particles still have to be removed.

(a) (b) (c) (d)

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

118

Perimeter

Area

yCircularit

Figure 2(c)). The number of clusters k was empiri-

cally set to four (see Section 3.1 for details about the

parameter set-up).

4) Refinement by Size and Circularity Analysis.

Bright epithelial regions with low EGF-R/TGF-

alpha activity have to be removed from the connec-

tive cluster to refine tissue segmentation. As shown

in Figure 2(d), a large number of these regions are

approximately round-shaped and are considerably

smaller than connective mass: then a selective re-

moval of particles with a low area and a high circu-

larity compared to threshold values T

and T

performed (parameters set-up in Section 3.1).

Equation 2 shows the proposed index for circu-

larity evaluation (a value of 1 indicates a perfect

circle, a value approaching 0 an increasingly elon-

gated polygon).

(2)

5) Refinement by Gradient Magnitude Analysis.

Other bright epithelial regions can be removed from

the connective cluster through their edge characteris-

tics, since connective tissue usually shows a well-

defined boundary w.r.t. epithelial background in

terms of intensity gradient variation. On the base of

this remark, in this step areas which show along

their boundary a percentage of edge pixels (i.e. pix-

els with high gradient intensity variation w.r.t. back-

ground) lower than a threshold value T

are selec-

tively removed from connective cluster (parameter

set-up in Section 3.1). Edge detection is performed

through a Sobel detector followed by automated

intensity global thresholding.

2.2 Supervised Procedure

An alternate approach for tissue segmentation is

supervised learning; for this purpose a Support Vec-

tor Machine (SVM) classification is proposed.

The SVM (Vapnik, 1998) is a theoretically supe-

rior machine learning method which has often been

shown to achieve great classification performance

compared to other learning algorithms across most

application fields and tasks including image process-

ing (Statnikov, 2005).

Here we propose a procedure based on binary

SVM classification, in which the input elements (in

this work, small tissue regions) are associated to one

of two different classes, connective or epithelial, on

the base of a set of representative characteristics, the

features vector. To perform a reliable classification,

the SVM is previously trained with a set of elements

whose class is well-known, the so-called training

instances.

The classification is based on the implicit map-

ping of data to a higher dimensional space via a ker-

nel function and on the consequent solving of an

optimization problem to identify the maximum-

margin hyperplane that separates the given training

instances (see Figure 4).

Optimal margin

CLASS 1

CLASS 2

Optimal hyperplane

Support vector

Optimal margin

CLASS 1

CLASS 2

Optimal hyperplane

Support vector

Figure 4: Maximum-margin hyperplane in SVMs (linearly

separable case). The boundary training instances (support

vectors) are indicated by an extra circle.

This hyperplane is calculated on the base of

boundary training instances (i.e. elements with char-

acteristics which are border-line between the two

classes), the so-called support vectors; new instances

are then classified according to the side of the hy-

perplane they fall into.

In order to handle linearly nonseparable data, the

optimization cost function includes an error minimi-

zation term to penalize the wrongly classified train-

ing instances.

See the references provided in the text for a

technical description of SVMs.

TRAINING

FEATURES

EXTRACTION

TRAINING CLASSIFICATION

TRAINING

FEATURES

EXTRACTION

TRAINING CLASSIFICATION

Figure 5: Supervised procedure based on SVM.

Our proposed supervised procedure for tissue

segmentation consists in three main steps (see Figure

5):

1) Training Features Extraction. In order to ob-

tain a good generalization of the SVM, a skilled op-

erator was asked to select from a large number of

real-life tissue images small rectangular regions

wherein both connective and epithelial tissue were

present. The images showed various staining levels

and very different characteristics of tissue shape and

intensity distribution.

In each representative sample the operator manu-

ally traced the boundaries of connective and epithe-

lial tissue. Then a NxN square sliding window was

FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN TISSUE CONFOCAL IMAGES - Comparison

between a Custom Unsupervised and a Supervised SVM Approach

119

horizontally and vertically shifted over the samples

(shift value s), thus covering the entire surface of the

image; for each shifted window, a features vector

was generated with the RGB values of 256 equally-

spaced pixels (see Figure 6, parameters set-up in

Section 3.2).

In this way, a features vector of 3x256 variables

was created for each single shift.

A +1 label was assigned to windows with a

prevalence of epithelial tissue pixels, a -1 label to

windows with a prevalence of connective tissue pix-

els.

2) Training. The labelled features vectors were fed

into the SVM for the training; for details about the

parameters set-up see Section 3.2.

3) Classification. The optimized SVM obtained in

the training step is used to perform tissue classifica-

tion for new images.

For this purpose, the input images are processed

to generate features vectors as in step 1 which are

fed into the trained SVM. At the end of the classifi-

cation, the SVM automatically associates positive

labels to epithelial patterns and negative labels to

connective patterns. The output is then processed to

reconstruct a two-dimensional result as in Figure 8.

3 IMPLEMENTATION

The algorithm was implemented in Java as a plugin

for ImageJ, a public domain image analysis and

processing software which runs on all the standard

operating systems (Windows, Mac OS, Mac OS X

and Linux): therefore it is totally hardware-

independent, flexible and upgradeable. We inherited

the whole class hierarchy of the open-source ImageJ

1.37 API and the free plugins for color deconvolu-

tion (Landini) and K-means clustering (Sacha) and

we implemented our own functions and classes. A

user-friendly interface enables the user to set differ-

ent parameters values without modifying the source

code.

For the supervised procedure we used the cSVM

tool for binary classification (Anguita, 2005), since

it uses the state-of-art optimization method SMO,

i.e. Sequential Minimal Optimization (Platt, 1999).

This cSVM tool implements the algorithm described

in (Wang, 2004), which was successfully used to

solve different real world problems. Our ImageJ

plugins for features vectors generation and output

reconstruction were integrated to the SVM tool.

The parameters of the proposed algorithms were

empirically tuned by a skilled operator after running

several experiments on a large dataset of real tissue

images which showed very different characteristics

of staining intensity, resolution, EGF-R/TGF-alpha

activity level, tissue shape. In the following subsec-

tions, we report some details about the implementa-

tion of both the unsupervised and the supervised

classification procedures and we outline the experi-

mental set-up of the main parameters.

3.1 Unsupervised Procedure

The number of clusters k (see Section 2.1 point 3)

was set to 4 after running the algorithm with values

varying from 2 to 5 and evaluating each time K-

means performance in terms of sensibility (power to

detect connective components) and selectivity

(power to avoid misclassification of epithelial com-

ponents). For values lower than 4 we often experi-

enced a very good sensibility but a not sufficient

selectivity; for higher values the sensibility was fre-

quently poor. A

k value equal to 4 assured a good

performance of K-means in all the tested images.

The size threshold T

(see Section 2.1 point 4)

was varied from 1000 to 5000 pixels with a step of

1000 and was finally set to 3000. Increasing values

led to a progressive improvement of selectivity in

the connective tissue selection; with values higher

Figure 6: Generation of the features vectors for SVM training. A NxN square window is horizontally and vertically shifte

on the sample, thus covering the entire surface of the image. For each shift a features vector is generated with RGB values o

256 equally spaced pixels, as for Window A and Window B. Epithelial instances are labelled with a +1, connective instances

with a -1.

Features vector : [

…

256

]

Label : -1 (CONNECTIVE TISSUE)

Features vector : [

…

256

]

Label : +1 (EPITHELIAL TISSUE)

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

120

than 3000 the lack in sensibility was often not ac-

ceptable. Similarly, the circularity threshold T

(see Section 2.1 point 4) was decreased from 0,9 to

0,3. A value of 0,7 assured a good selectivity en-

hancement without altering sensibility in any of the

images.

The edge threshold T

(see Section 2.1 point 5)

was increased from 20% to 35% with a step of 5%,

evaluating each time the parameter performance in

terms of selectivity enhancement and sensibility

preservation. A value of 25% assured the best im-

provement in selectivity without altering sensibility

in any of the tested images.

3.2 Supervised Procedure

The window size N for features vectors generation

(see Section 2.2 point 1) should grant a visible

differentiation between connective and epithelial

tissue; since nuclei are blue-colored and quite

similar in both the tissues, the window has to be

large enough to contain a whole nucleus and some

surrounding tissue. On the other hand, lower-sized

windows allows a better selectivity.

After running several experiments with values

varying from 16 to 72 pixels, N was set to 32 for

x200 images and to 64 for x400 images.

Since the optimal window size depends on image

resolution, x200 and x400 images were respectively

classified with SVM trained with x200 and x400

samples.

The shift value s (see Section 2.2 point 1) was

set to N/4, which granted the best compromise

between selectivity of classification and

computational time.

After running experiments with linear, gaussian

and polynomial kernels, we finally chose the

normalized polynomial kernel shown in Equation 3,

where x

and x

are feature vectors, n=768 is the

input space dimension and p=2 is the kernel

hyperparameter; see (Wang, 2004) for technical

details).

(3)

4 EXPERIMENTAL RESULTS

We tested the performance of both the algorithms on

a large dataset extracted from real tissue images

which presented positive reactions at the EGF-R or

at the TGF-alpha receptor activation (see Figure 1

for examples); reactions are localized in cellular

membranes for EGF-R and in cytoplasm for TGF-

alpha. Images were acquired from different samples

with two different enlargements, x200 or x400.

A skilled operator was asked to manually draw

the boundaries of connective tissue in each of the

testing datasets. The manual segmentations per-

formed by the operator were pixel-by-pixel com-

pared to those obtained by both the unsupervised

and the supervised algorithms. Connective tissue

selection was evaluated in terms of sensibility (i.e.

power to detect connective tissue) and selectivity

(power to avoid misclassification of non-connective

tissue): for this purpose, the percentage of respec-

tively connective and non-connective pixels which

were equally classified by manual and automated

segmentation was calculated. The segmentation ac-

curacy was then calculated as weighted average of

sensibility and selectivity, as in Equation 4.

(4)

Different weights were used because sensibility

is more critical for automated measures of protein

()

nxxnxx

nxx

xxK

+⋅+⋅

+⋅

2211

ySelectivitySensibilitAccuracy ⋅+⋅=

Table 1: Experimental results of unsupervised and supervised classifications. As outlined in Section 3.2, in supervised

classification two different SVMs trained respectively with x200 and x400 samples were used (the number of training

instances extracted from each dataset and the total number of training instances are reported for both x200 and x400 classi-

fiers). Training instances were removed from the validation dataset, which was considerably larger.

UNSUPERVISED ALGORITHM SUPERVISED ALGORITHM

Dataset

Sensibility

(%)

Selectivity

(%)

Accuracy

(%)

Number of

training instances

Number of

validation instances

Sensibility

(%)

Selectivity

(%)

Accuracy

(%)

1 81,89 90,54 84,77 1692 28308 57,91 91,38 69,07

2 94,64 84,94 91,41 912 20263 94,05 79,20 89,10

3 95,21 97,99 96,14 220 20192 91,09 94,75 92,31

4 86,60 87,32 86,84 408 19142 84,41 91,18 86,66

x200

tot

3232 91137

5 91,77 86,20 89,91 558 6942 67,48 82,35 72,43

6 91,30 78,56 87,05 640 6860 66,48 90,02 74,32

7 99,67 93,33 97,56 252 7248 93,53 87,46 91,51

x400

8 89,21 86,28 88,23 300 5888 87,29 85,39 86,66

tot

1750 28688

FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN TISSUE CONFOCAL IMAGES - Comparison

between a Custom Unsupervised and a Supervised SVM Approach

121

activity, which is the principal application targeted

by our method: in fact, in order to obtain a reliable

measure, it is fundamental to eliminate as much as

possible non representative tissues from the range of

interest; on the contrary, erroneous removal of some

epithelial regions is more tolerable, since it has a

lower influence on the final measure.

Results obtained for both the automated algo-

rithms are reported in Table 1. The number of train-

ing instances extracted from each dataset and the

total number of training instances are reported too

for both x200 and x400 SVMs. The classification

performance was evaluated on a large validation

dataset which did not include the patterns used for

training.

Some examples of tissue segmentation are shown

in Figure 8.

Accuracy of tissue segmentation

0,00

10,00

20,00

30,00

40,00

50,00

60,00

70,00

80,00

90,00

100,00

12345678

Da ta s e t

UNSUPERVISED

SUPERVISED

Figure 7: Accuracy of tissue segmentation; comparison

between unsupervised and supervised procedure.

As shown in Table 1 and Figure 7, our unsuper-

vised procedure achieved the best results: this

method performed tissue segmentations highly com-

parable with those provided by the skilled operator

in all the testing datasets; mean accuracy was

90,24%, with values generally around 90% and al-

ways above approximately 85%. SVM performed

worse in all the tested datasets; mean accuracy was

about 7,5% lower than our unsupervised method.

As we previously outlined, SVM is a theoreti-

cally superior machine learning method which has

often been shown to achieve great classification

performance compared to other learning algorithms

across most application fields and tasks including

image processing (Angelini et al., 2006; Cai, 2003;

Muller et al., 2001; Osuna, 1997). However, in this

case its classification performance was poor because

of the intrinsic complexity of the images targeted by

our method: in fact, these images showed very dif-

ferent characteristics of staining, tissue shape and

intensity distribution. Because of the heterogeneity

of the representative features of each class, it was

impossible for the supervised method to obtain a

satisfying separability of connective and epithelial

tissue.

Images heterogeneity was less critical for the un-

supervised approach, since differently from SVMs it

is based only on the characteristics of the input im-

age and not on a fixed model of the ground truth.

On the other hand, our unsupervised method’s

selectivity is influenced by tissue composition: in

fact, since the number of clusters is a-priori fixed,

some epithelial regions with low brown staining are

often misclassified in images without any connective

tissue.

Despite this eventuality is unlikely, since pure-

epithelial tissue samples are very uncommon (and

we reasonably suppose that the operator would es-

cape the automated tissue segmentation in this case),

we are working on the solution of the problem: in

particular, the introduction of an adaptive number of

clusters is in development.

As regards the supervised approach, other learn-

ing methods such as neural networks and artificial

neural networks (ANN) will be tested in the future.

5 CONCLUSIONS

We presented a fully-automated unsupervised tissue

image segmentation method that allows to distin-

guish tumor areas in immunohistochemical images

and disregard non pathological areas such as connec-

tive tissue. This procedure is critical for automated

protein activity quantification in tumor tissues in

Figure 8: Examples of tissue segmentation performed by the unsupervised (a) and the supervised (b) algorithm (manual

segmentation in red, automated segmentation in black).

(1-a) (2-a) (2-b)(1-b)

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

122

order to analyze the pathology dynamics and devel-

opment.

We described the original processing steps we

designed. Finally, we carried out an extensive ex-

perimental evaluation on a large set of heterogene-

ous images that demonstrated the high accuracy

achievable by the proposed technique (90% on aver-

age) compared to a more traditional approach based

on Support Vector Machines (SVM).

As future work, we will compare the proposed ap-

proach to artificial neural networks (ANN), and we

will eventually study the possibility of their integra-

tion.

ACKNOWLEDGEMENTS

We acknowledge the Dep. of Pathology of the

S.Luigi Hospital of Orbassano in Turin, Italy, for

providing IHC images and for the helpful and stimu-

lating discussions.

REFERENCES

Angelini, E, Campanini, R., Iampieri, E., Lanconelli, N.

Masotti, M., Roffilli, M., 2006. Testing the perform-

ances of different image representation for mass clas-

sification in digital mammograms. Int. J. Mod. Phys.

17(1):113-131.

Anguita, D., Boni, A., Ridella, S., Rivieccio F., Sterpi, D.,

2005. Theoretical and Practical Model Selection

Methods for Support Vector Classifiers. Springer,

Studies in Fuzziness and Soft Computing, Support

Vector Machines: Theory and Application.

Brey, E.M., Lalani, Z., Hohnston, C., Wong, M., McIntire,

L.V., Duke, P.J., Patrick, C.W., 2003. Automated se-

lection of DAB-labeled tissue for immunohistochemi-

cal quantification. In J. Histochem. Cytochem., 51(5),

pp.575-584.

Cai, C.Z., W.L.Wang, Y.Z: Chen, 2003. Int.J.Mod.Phys.

14:575.

Demandolx D, Davoust J. , 1997. Multiparameter image

cytometry: from confocal micrographs to subcellular

fluorograms. Bioimaging. 4:159-169.

Dybowski R., 2000. Neural computation in medicine:

perspectives and prospects. Proc. ANNIMAB-1. pp.

27-36.

E. Osuna, R. Freund, F. Girrosi, 1997. Training Support

Vector Machines: an Application to Face Detection.

IEEE Computer Society Conference on Computer Vi-

sion and Pattern Recognition (CVPR'97). pp. 130.

Ficarra, E., Macii, E., De Micheli, G., 2006. Computer-

aided evaluation of protein expression in pathological

tissue images. In Proc. of IEEE CBMS‘06., pp.413-

418.

Jain, A.K., Dubes, R.C., 1988. Algorithms for clustering

data, Prentice Hall.

Landini, G., 2007. Software, http://www.dentistry.bham.ac

.uk/landinig/software/software.html

Malpica N, de Solorzano CO, Vaquero JJ, Santos A,

Vallcorba I, Garcia-Sagredo JM, del Pozo F, 1997.

Applying watershed algorithms to the segmentation of

clustered nuclei. Cytometry. 28(4): 289-297.

Muller K.R:, S. Mika,G. Ratsch, K. Tsuda, 2001. IEEE

Trans. Neural Networks, 12:181.

Nattkemper, T.W., 2004. Automatic segmentation of digi-

tal micrographs: A survey. Medinfo, 11(Pt 2):847-51.

Nedzved A, Ablameyko S, Pitas I., 2000. Morphological

segmentation of histology cell images. ICPR. 1:500-3.

Platt, J., 1999. Fast training of support vector machines

using sequential minimal optimization. In Scholkopf,

B., Advances in kernel methods-support vector learn-

ing. MIT Press, Cambridge, MA, USA.

Rasband, W.S., ImageJ, U. S. National Institutes of

Health, Bethesda, Maryland, USA, rsb.info.nih.gov/ij/.

Ruifrok, A.C., Johnston, D.A., 2001. Quantification of

histochemical staining by color deconvolution. In

Anal.Quant.Cytol.Histol., 23(4), pp.291-299.

Ruifrok, A.C., Katz, R., Johnston, D., 2004. Comparison

of quantification of histochemical staining by Hue-

Saturation-Intensity (HSI) transformation and color

deconvolution. In Appl. Immunohisto. M. M., 11(1),

pp.85-91.

Sacha, J., K-means clustering, http://ij-plugins.sourceforge

.net/plugins/clustering/index.html.

Statnikov, A., Aliferis, C.F., Tsamardinos, I., Hardin, D.,

Levy, S., 2005. A comprehensive evaluation of multi-

category classification methods for microarray gene

expression cancer diagnosis. In Bioinformatics, 21(5),

pp.631-643.

Taneja, T.K., SK.Sharma Markers of small cell lung can-

cer. World Journal of Surgical Oncology, Vol(2):10.

Vapnik, V., 1998. Statistical learning theory, Wiley-

Interscience, New York, NY, USA.

Wang, L., 2004. Support vector machines: theory and

applications, Springer.

FULLY-AUTOMATED SEGMENTATION OF TUMOR AREAS IN TISSUE CONFOCAL IMAGES - Comparison

between a Custom Unsupervised and a Supervised SVM Approach

123