You are on page 1of 5



Rajendra Talware
Dept. of E&Tc, Engg, VIIT, Pune
Ph.D. student COE, Pune

Abstract Cellular image analysis is being attempted by

several Image Processing (IP), statistical approaches and
machine learning (ML), Neural Networks (NN) techniques. IP
based algorithms are getting trapped into subjectivity due to
reasons of cell image propoerties and methods of extractions
and classification techniques. ML based methods are also well
explored, but it suffer from large training sets and feature
selection to reduce dimensionality of neural network.
This paper proposes design and integration of Evolutionary
methodologies with temporal networks and on-the-fly learning
for getting desired confidence interval with minimal training.
IP based method is experimented with morphological feature
extraction. Results revels that, utility limited to spatial and
noise-free data. This emphasize the need of evolutionary
methodologies with temporal learning and for optimizing
weights, structure and learning of NN. Paper presents results
extracted from IP algorithm with its shortcoming and proposes
framework based on Evolutionary ANN to overcome the
subjectivity issue reported in text and experimentation.
Keywords-cellular image analysis, evolutionary methodologies,
wavelet based framework, cancer)

Multiple biomedical imaging forms an essential part
of cancer clinical protocols and is able to furnish
morphological, structural, metabolic and functional
information. Early detection of cancer through screening
based on imaging is probably the major contributor to a
reduction in mortality for certain cancers. [1] Diagnosis
usually requires the histological examination of biopsy
samples; a pathologist typically assesses the deviation in the
cell structures and/or the change in the distribution of the
cells. [2] Integration with diagnostic tools assists in clinical
decision-making. For such applications it is desired to have
intelligent systems, which can handle temporal data by doing
dynamic adjustment in the learnt pattern using underlying
criticalities. Evolutionary Artificial Neural Networks (EANN)
has ability to be able to adapt to the situation, where evolution
is another fundamental form of adaptation in addition to
learning [3].
The paper is organized as, section II presents brief review
of work in cancer detection and diagnosis.Section III
This Project is partially funded by University of Pune, Under Grant No.
BCUD/183 Year 2009-2011

Dr. Aditya Abhyankar

Dean R&D, VIIT, Pune
Adjunct Professor, COE, Pune
pressents the proposed methodology and mathematical model
for it. Section IV reports the experimentation carried out using
parametric methods of image processings, and template based
learning for pattern classification, Subsequently Section V
discusses results from experimentation and Finally Section VI
presents conclusion and future direction of work.
Digital image processing and pattern recognition
techniques are widely used in pathological screening for
cancer. A special issue on the same compiles the recent trends
in [4]. Study for cancer diagnosis and Gleason grading of the
histological images of prostate reported in [5], dealt with
color, texture, and morpho-metric features at both the image
(global) and histological object levels. Paper compared the
performance of Gaussian, nearest neighbor, and Support
Vector Machine (SVM) classifiers together with the sequential
forward feature selection algorithm.
Yu-Len Huangt et al[6], employed the image retrieval
technique to classify breast tumors as benign or malignant
lesions. Oncology/ Pathologist located regions-of-interest
(ROI) of ultrasound images. The textual features from
manually annotated regions-of-interest (ROI) sub-image are
utilized to classify breast tumors. The principal component
analysis (PCA) is used to reduce the dimension of textural
In attempt of reducing the diagnosis time and classifying
mass in breast to either benign, or malignant with high
accuracy. Afzan Adam et al [7], used Back Propagation
Neural Network (BPNN) model with Genetic Algorithm (GA)
namely GAwNN for faster classifier, without downgrading the
classification performance.
C. Scott, [8], introduces a hierarchical wavelet-based
framework for modeling patterns in digital images. Wavelet is
used for efficient image representations. Unknown model
parameters are inferred from labeled training data using
TEMPLAR, a template learning algorithm with linear
complexity. It employs minimum description length (MDL)
[3]. After sufficient training with different patterns, it provides
a low-dimensional subspace classifier, invariant to unknown
pattern transformations and background clutter. The
TEMPLAR essentially considers a pattern confined at centre,
else requires a preprocessing, as it aiming at making template

invariant to global transformation. Local transformations need

to be taken care by wavelet models. [16]


Tf f 0

t xt xt

Estimation Error

III A. Mathematical model:


C Known a priori
1,t State parameter uncertainties
2,t Measurement disturbances
Vector responsible for matching f 0 (, ) & f (, )
Evolutionary dynamic update strategy for network learning is
given by

xt f ( xt , ut ) 1,t
yt C xt 2,t

f f 0

Nominal dynamics of wavelet neural network

f 0 x, u | : A0 x w10 1 ( x) w20 1 ( x)u

( x)

A0 x w10 w20 1

1 ( x)u

1 x 2 x l1 x x

Wavelet Differential evolutions with orthogonal basis are


( x) ( x)u (t ) l2 x x

( ij , kl ) ij kl i, j , k , l Z
h0 ( z )




( x)


h0 ( x, )

( x)dx


i j i j ( x)


xt converges asymptotically and w1 should

converge and w2 should be adaptive.

i M1 j N1

NN approximation error condition

Parameter adjustment wi,t provides function approximation.

( M 1 , M 2 , N1 , N 2 ) h0 ( x, )

Wi ,t K i ,t ( P1 PW
2 1,t i ,t f 2 xt ) i
21 K 1 K W

Upper bound on error modeling

f0 f1 x t

f0 f1 1f

( f2 2f f1 1f

xt Axt w1,t ( x ) w2,t 2 ( x ) ut
y cxt
k1[ yt cxt ] k2 t
yt cxt



Observer structure from NN estimator

h ( x)


f 2 x t

i ,t


i ,t

i ,t

Where P1P2 are positive solution of Riccati equation

(max i )C T et etT c
(min 2i 1) i Ti
i ,t :

x t

f0 f t f 2 x t
f0 , f1 R f0 , f1

Average Estimation Error ()

: lim g.

s 0

xs xs Q0 ds

Training of Network

t Yt X tT t t t X t X tT t

( f * g )( x)

f ( y ) g ( x y )dy

f ( x y ) g ( y ) dy

Use un-sharpening
filter for sharpening
and contrast

Training Quality

w (0)

}T dt
tr { t t1[t1 ]T
T t 0

Stability analysis of learning lawas by Lyapunav method

(Bounded weight behavior)

V1 : tr {w1,Tt w1,t } [ K1,t K1, min ]2

V1 : t r {w 1,t ( K1,t 1 (t , x ) 2 K i ,t K 1,t w 1,t )}

2 ck1,t [ k1,t k1,min ] k1,t tr {w 1,t 1 (t , xt )


2 K1,t (


t r {w 1,t w 1,t } c[ k1,t k1,min ] )


Fig.1 Shows (Top left) the original image of cancerous

cell,(Top right) grey-scale of it, (Bottom Left) de-noised
image after convolution, (Bottom Right) Un-sharped image
after filtering
IV C. Morphological Feature Extraction Subjectivity

IV A. Data collection and Domain Expert support:
Huge volume of data is required.. Data collected from
Histopathology Dept, Cancer Research Centre, Ruby Hall
Clinic, Pune, in the form of glass slides. Images are processed
using microscope-CCD setup. Manual annotation plays a key
role in establishing a ground truth. Validate the results against
the state-of-the-art, which may lead to give best predicted
pathological solutions by avoiding the subjectivity in
processing and underlying algorithms. Due to evolutionary
nature of algorithm, manual annotation is required for training
and validation for every new class of pattern.

Morphological feature extraction [11] is foremost common

and important step in cancer detection. Parametric approach
was selected to extract the features. Image filtered using
anisotropic diffusion,[12] thresholding. Staining methods
routinely used in pathology. This causes problems in image
segmentation for the quantitative analysis and detection of
cancer. Morphological features are extracted using dilation
and erosion with different parameter values, from KP=1000,
100 and 08 respectively.
In the next operation, the
morphological constructor is changed and the change in shape
of the extracted shape can be seen. The Morphological
constructors used are disk, arbitrary and diamond.

IV B. Denoising: Due to a considerable amount of noise

that arises from the staining process, it is usually necessary to
reduce the noise prior to the focal area identification [12]. The
noise can be removed using either by thresholding, filtering,
Morphology operations. Filter is implemented the Gaussian
Probability density function (PDF) [14]for one dimension and
repeated for the other with

f ( x; , 2 )

e ( x ) /(2 )

Where > 0 is the standard deviation, the real

parameter is the expected value, and (x) = (2) 1/2ex2/2 is
the density of the standard normal distribution. Convolution
filtering is used to remove the noise from the image. Resulting
image is free from noise, but contains blurring and un-sharp
objects in the image [13].

Fig.2 shows (top row) change in dilation parameter from 08,

100, 1000 from left to right resp. (bottom row) shows
changein shape of object based on different constructors
IV D. TEMPLAR Template based learning for
classification and synthesis

During process of capture, an image can appear at any

locations, orientations, scales. These uncertainties in
pattern observations are modeled with a hierarchical
framework, based on the notion of deformable templates
from pattern theory [4]. A template is a noise-free
observation of a pattern that can be transformed into an
arbitrary observation of the same pattern by applying a
deformation to the template in the form of rotation,
scaling, translation and shear, as well adding observation
noise. For translation, the image was treated as a torus.
An image was translated to the right, those pixels on the
right edge of the original image wrap around and appear
on the left edge of the translated image. For rotation,
when an image was rotated through an angle which was
not a multiple of 90 degrees, the corners of the original
image that are cropped off by rotation are mapped in a
one-to-one manner back to the empty corners of the
rotated image. Number of training images used are 10,
with depth of the wavelet as 5. The figure 3 shows
translation applied on the input image, resulting in nonparametric way of template learning. When observed
image and training images are passed through NISTSTS, [15] it is observed that, the training images formed
are passing certain non-parametric tests, which could not
be through for original captures image.

cells. Furthermore, no fixed shape, size and color fir particular

type of cell or level of deviation. Most of the parameters are
non-modular. Segmentation becomes a tedious task for this
case and none of the algorithms claimed confidence for
general framework of pathological data in general, rather most
of the algorithms reported success for particular disease.
Staining is very much required for pathology, but adds burden
of filtering the image. Morphological feature extraction, which
is based on mainly thresholding and segmentation, detection
of continuities gets trapped into subjectivity. Results shows,
high subjectivity in results as the methods are strictly modular.
Morphological operations are not representing the physical
meaning of the processes. It suffers from dependence on the
non-topographical content in an image.
In light of these results from experimentation as well reported
in [16]. Non-parametric approaches are followed. Template
based learning using Hierarchical wavelet-based model used
for pattern modeling as represented in[8]. It is observed that,
the model is not enough to take care of local transformation in
images as well, it requires pre-processing in order to confine
the pattern to centre. The training images generated from
TEMPLAR framework are observed invariant to translation
and rotation. Tested with NIST-STS suit, the training images
pass certain non-parametric test, whereas, the original image,
from which patterns are made, fails for the same. This, in a
way, confirms non-parametric approach and deals with nonmodular data. Another important issue with network training
is the different datasets and opinions of experts. Manual
annotation plays an important role in establishing the ground
truth. The proposed dynamic update model which will be
evolving on wavelet framework, is based on the nominal
model and Riccati matrix, on stone-wiestrass theorem of
uniform approximation by polynomials for continuous real
valued functions on a compact interval.

Fig. 3: Illustration of translation operation for generating

training images, Translation is from left to right in the images
starting from top left, top right to Bottom right
successively(original image : Figure1 top right)
During the process of image acquisition it was observed that,
the cell images can appear at anywhere in glass slide
whenever attempted with manual or motorized microscopy.
HCS data is also suffer from these acquisition related
problems of clutter and occlusion, touching and overlapping


In this paper, various challenges in detection and diagnosis

of cancer cells from histo-pathological data are discussed for
denoising and feature extraction. In particular, subjectivity in
segmentation and morphological feature extraction with
modular methods are discussed. Classification methodologies
based on image analysis algorithm and along-with limitations.
In light of the results of parametric methods, a TEMPLAR
framework, a hierarchical wavelet based approach is adopted
for taking advantage of sparse representation and to achieve
invariant to global transformation. Worked with wavelet based
model, we come up with a mathematical model for future
work on lines of Riccatti matrix and Stone-Wiestrass theorem
of approximation. Experimentation proved shortfalls of
traditional process of image analysis. Approach for designing
a class of EANN with objectives for classifying cell image
based on more objective features is planned for proposed work
using non-parametric methods.
We thank Ruby Hall Clinic, Pune for providing biopsy
samples data and manual annotation. This project is supported

by BCUD, University of Pune under grant for the year 20092011.


C. Scott, "A Hierarchical Wavelet-Based Framework for Pattern

Analysis and Synthesis," Masters Thesis, Rice University 2000.



U. Grenander and M. J. Miller, Representations of knowledge in

complex systems, J. Roy. Stat. Soc., vol. 56, no. 3, pp. 133, 1994.



Leonard Fass, Imaging and cancer: A review Molecular Oncology

(2008) 115-152
Cigdem Demir and Bulent Yener, Automated cancer diagnosis based
on histo-pathological images: a systematic survey, technical report,
Rensselaer Polytechnic Institute, dept. of computer science, TR-05-09.
Richard Duda, Hart, Stork Pattern Classification Wiley-Inter-science,
Jinshan Tang (editorial), Digital image processing and pattern
recognition techniques for the detection of cancer, Pattern recognition
Vol 42 (2009) 1015-

[5] Ali Tabesh, Mikhail Teverovskiy, Ho-Yuen Pang, Vinay P. Kumar,

David Verbel, Angeliki Kotsianti, and Olivier Saidi, Multifeature
Prostate Cancer Diagnosis and Gleason Grading of Histological
Images. IEEE Trans Med Imaging . 2007 Oct; 26(10):1366-78


Yu-Len Huangt et al Breast Cancer Diagnosis Using Image Retrieval

For Different Ultrasonic Systems 2004 International Conference on
Image Processing (ICIP) p.2957-p2961.
Afzan Adam et al, Computerized Breast Cancer Diagnosis with
Genetic Algorithms and Neural Network. Scientific common, 2008

[10] LIoyd K. Williams, Linear and Riccati matrix equation International

Journal of Mathematics and Mathematical Sciences, Volume 12 (1989),
Issue 1, Pages 131-136, doi:10.1155/S0161171289000189
[11] Jean-Philippe, Morphological Feature extraction for the classification
of digital images of cancerous tissue, IEEE transactions on Bio-medical
Engineering, Vol. 43, NO.10, Oct. 1996 pp 1011-1021
[12] Pietro Perona and Jitendra Malik, Scale-Space and Edge Detection
Using Anisotropic Diffusion , IEEE Transactions on Pattern Analysis
and Machine Intelligence, Vol. 12. No. 7. July 1990, pp 629-639.
[13] Rafael C. Gonzalez, Richard E. Woods, Steven L. Eddins, Digital
Image Processing Using MATLAB.

Bernardo, J. M.; A.F.M. Smith (2000). Bayesian Theory. Wiley.ISBN


[15] NIST-STS suite:

[16] Rajendra Talware, Aditya Abyankar, towards designing a spatiotemporal neural network for non-modular high content pathological
screening IET national conference, NCSIPA 2009,