Artificial neural network based cancer cell classification

Computer Engineering and Intelligent Systems www.iiste.org
ISSN 2222-1719 (Paper) ISSN 2222-2863 (Online)
Vol 3, No.2, 2012

Artificial Neural Network based Cancer Cell Classification
(ANN – C3)
Guruprasad Bhat1* Vidyadevi G Biradar2 H Sarojadevi2 Nalini N2
1. Cisco Systems, SEZ Unit, Cessna business park, Marathahalli-Sarjapur outer ring road, Bangalore,
Karnataka, India – 560 103
2. Nitte Meenakshi Insitute of Technology, Bangalore , Karnataka, India – 560 064
*guruprasadbharatibhat@gmail.com,
vgb2011@gmail.com,hsarojadevi@gmail.com,nalinaniranjan@hotmail.com

Abstract
This paper addresses the system which achieves auto-segmentation and cell characterization for prediction
of percentage of carcinoma (cancerous) cells in the given image with high accuracy. The system has been
designed and developed for analysis of medical pathological images based on hybridization of syntactic
and statistical approaches, using Artificial Neural Network as a classifier tool (ANN) [2]. This system
performs segmentation and classification as is done in human vision system [1] [9] [10] [12], which
recognize objects; perceives depth; identifies different textures, curved surfaces, or a surface inclination by
texture information and brightness.
In this paper, an attempt has been made to present an approach for soft tissue characterization utilizing
texture-primitive features and segmentation with Artificial Neural Network (ANN) classifier tool. The
present approach directly combines second, third, and fourth steps into one algorithm. This is a semi-
supervised approach in which supervision is involved only at the level of defining structure of Artificial
Neural Network; afterwards, algorithm itself scans the whole image and performs the segmentation and
classification in unsupervised mode. Finally, algorithm was applied to selected pathological images for
segmentation and classification. Results were in agreement with those with manual segmentation and were
clinically correlated [18] [21].
Keywords: Grey scale images, Histogram equalization, Gausian filtering, Haris corner detector, Threshold,
Seed point, Region growing segmentation, Tamura texture feature extraction, Artificial Neural
Network(ANN), Artificial Neuron, Synapses, Weights, Activation function, Learning function,
Classification matrix.

1. Introduction
In the modern age of computerized fully automated trend of living, the field of automated diagnostic
systems plays an important and vital role. Automated diagnostic system designs in Medical Image
processing are one such field where numerous systems are proposed and still many more under conceptual
design due explosive growth of the technology today. From the past decades, we have witnessed an
explosive growth of Digital image processing for analysis of the data that can be captured by digital images
and artificial neural networks are used to aggregate the analyzed data from these images to produce a
diagnosis prediction with high accuracy instantaneously where digital images serve as tool for input data
[20] [21]. Hence in the process of surgery these automated systems help the surgeon to identify the infected
parts or tumors in case of cancerous growth of cells to be removed with high accuracy hence by increasing
the probability of survival of a patient. In this proposal one of such an automated system for cancer cell
classification which helps as a tool assisting surgeon to differentiate cancerous cells from those normal
cells i.e. percentage of carcinoma cells, instantaneously during the surgery. Here the pathological images
serve as input data. The analysis of these pathological images is directly based on four steps: 1) image
filtering or enhancement, 2) segmentation, 3) feature extraction, and 4) analysis of extracted features by
pattern recognition system or classifier [21]. Since neural network ensembles are used as decision makers

7

Vol 3, No.2, 2012
even though network takes more time to adapt behavior, once it is trained it classifies almost
instantaneously due to electrical signal communication of nodes in the network.

2. System architecture
The ANN – C3 architecture is shown in figure 1. It comprises of five distinct components, as show below.
Each component is described briefly in subsequent sections.

Figure 1: ANN - C3 system architecture
2.1 Images used
This system is designed and verified to take grey scale pathological images as input. Grey scale
pathological images help to identify affected cells makes these images for analysis of cancerous growth of
cells.

2.2 Pre-processing
Grey scale pathological imaging process may be dirtied by various noises. Perform an image pre processing
task to remove noise in a pathological image first. To remove the noise the Histogram equalization or
Gaussian filter based median filtering is done [5] [6] [8] [19].

2.3 Segmentation
Segmentation includes two phases. First phase deals with threshold detection and the later one with similar
region identification. For threshold detection various methods like GUI selection, graphical method or
corner detectors can be used. GUI selection reduces automation and graphical method fails when multiple
objects are present in an input data. Since this design mainly deals with multiple objects (cells) in an input
image, Haris corner detectors are used to find threshold. In second phase, threshold points detected by
corners serve as seed point for segmentation. Four neighborhood based region growing segmentation is
used increase the speed compare to eight neighborhood and increase the accuracy compared to region split
and merge i.e. trade off between accuracy and speed. A brief discussion of Haris corner detector and 4-
neighborhood region growing Segmentation is done in section III [11] [13].

2.4 Feature Extraction
Neural network classifiers are those differ from traditional classifiers like Bayesian and k – nearest
neighborhood classifiers in various aspects from type of input data to output representation. Since the
neural networks are used as classifiers in this design which takes only numerical data as input rather than
any kind of data as input by Bayesian and k – nearest neighbor classifiers, the input image data has to be
converted to numerical form. This conversion is done by extracting tamura texture features. A brief
discussion of tamura texture feature is done in section IV.
Tamura texture features:
The human vision system (HVS) permits scene interpretation ‘at a glance’ i.e. the human eye ‘sees’ not
scenes but sets of objects in various relations to each other, in spite of the fact that the ambient illumination

8

Vol 3, No.2, 2012
is likely to vary from one object to another—and over the various surfaces of each object—and in spite of
the fact that there will be secondary illumination from one object to another. These variations in the
captured images are referred as tamura texture features, even the same texture features are observed by
surgeon to differentiate carcinoma cells and non-carcinoma cells.

2.5 Neural Network
Supervised feed-forward back-propagation neural network ensemble used as a classifier tool. As discussed
previously, neural network differs in various ways from traditional classifiers like Bayesian and k – nearest
neighbor classifiers. One of the main differences is linearity of data. Traditional classifiers like Bayesian
and k – nearest neighbor requires linear data to work correctly. But neural network works as well for non-
linear data because it is simulated on the observation of biological neurons and network of neurons. Wide
range of input data for training makes neural network to work with higher accuracy, in other words a small
set of data or large set of similar data makes system to be biased [22]. Thus neural network classifier
requires a large set of data for training and also long time to train to reach the stable state. But once the
network is trained it works as fast as biological neural networks by propagating signals as fast as electrical
signals.

3. Haris corner detector and 4-neighborhood region growing segmentation
3.1 Haris corner detector
A corner can be defined as the intersection of two edges. A corner can also be defined as points for which
there are two dominant and different edge directions in a local neighborhood of the point. An interest point
is a point in an image which has a well-defined position and can be robustly detected. This means that an
interest point can be a corner but it can also be, for example, an isolated point of local intensity maximum
or minimum, line endings, or a point on a curve where the curvature is locally maximal.
In practice, most so-called corner detection methods detect interest points in general, rather than corners in
particular. As a consequence, if only corners are to be detected it is necessary to do a local analysis of
detected interest points to determine which of these real corners are.
A simple approach to corner detection in images is using correlation, but this gets very computationally
expensive and suboptimal. Haris corner detector is one such corner detector, which uses differential of the
corner score with respect to direction directly, instead of using shifted patches. This corner score is often
referred to as autocorrelation.
The algorithm of haris corner detector as follows:
Without loss of generality, we will assume a grayscale 2-dimensional image is used. Let this image be
given by I. Consider taking an image patch over the area (u,v) and shifting it by (x,y). The weighted sum of
squared differences (SSD) between these two patches, denoted S, is given by:
(1)

I(u + x,v + y) can be approximated by a Taylor expansion . Let Ix and Iy be the partial derivatives of I, such
that
(2)
This produces the approximation
(3)

This can be written in matrix form:
(4)

Where A is the structure tensor,

9

Vol 3, No.2, 2012
(5)

This matrix (5) is a Harris matrix, and angle brackets denote averaging (i.e. summation over (u,v)). If a
circular window is used, then the response will be isotropic [16].
A corner (or in general an interest point) is characterized by a large variation of S in all directions of the
vector (x,y). By analyzing the eigenvalues of A, this characterization can be expressed in the following
way: A should have two "large" eigenvalues for an interest point. Based on the magnitudes of the
eigenvalues, the following inferences can be made based on this argument:
If λ1≈0 and λ2≈0 then this pixel (x , y) has no features of interest.
If λ1≈0 and λ2 has some large positive value, then an edge is found.
If λ1 and λ2 have large positive values, then a corner is found.
Haris and Stephens noted that exact computation of the eigenvalues is computationally expensive, since it
requires the computation of a Square root, and instead suggest the following function Mc, where κ is a
tunable sensitivity parameter:
Mc= λ1λ2 – k(λ1+λ2)2=det(A) – k trace2(A) (6)
Therefore, the algorithm does not have to actually compute the Eigen value decomposition of the matrix A
and instead it is sufficient to evaluate the determinant and trace of A to find corners, or rather interest points
in general.
The value of κ has to be determined empirically, and in the literature values in the range 0.04 - 0.15 have
been reported as feasible.
The covariance matrix for the corner position is A − 1, i.e.
(7)

Compute x and y derivatives of image
(8)
(9)
Compute product of derivatives of each image
(10)
(11)
3. Compute the sums of products of derivatives at each pixel
(12)
(13)
(14)
Define at each pixel (x, y) the matrix

(15)
Compute the response of the detector at each pixel
R=Det(H) – k(Trace(H))2 (16)
Threshold on value R. Compute nonmax suppression.
3.2 4-neigbourhood region growing Segmentation
Segmentation is the process of identifying the region of interest from the input image. Considering an input
image I being read and converted to the greyscale image .let’s assume the seed point to be (x , y). If the

10

Vol 3, No.2, 2012
seed point is provided by the GUI then a function getpts() will make sure the x and y axes values have been
fetched. To create a mask we’ll convert all the pixels in the image I’ to 0 and call the image J.
In order to discover the neighbors we will use four pixel connectivity [14]. Starting with the seed point the
algorithm looks for the 4 pixels surrounding the pixel in consideration. Every time a surrounding pixel is
considered, the region mean is calculated and checked with that of the pixel in consideration and added to
the region. Similarly as the pixel is added to the region corresponding pixel in the image J is highlighted to
1 which would result in the highest intensity hence illuminating the pixel. As the segmentation continues
the region into consideration is intensified in the image J resulting in the segmentation of the affected area,
which later can be combined with the original image and displayed to the user.

5. Tamura Textutre feature extraction
Tamura texture feature concepts proposed by Tamura et al in 1978. These tamura texture features
corresponding to human perception and these features examined by 6 different constituent features. Six
features are: [15]
Coarseness – Coarseness is the numerical value describing whether texture is coarse or fine.
Contrast – Contrast defines whether texture contrast is high or low.
Directionality – Directionality defines whether texture pallets are oriented in single direction or not i.e.
directional or non-directional.
Line-likeness – Line-likeness correspond to pattern elements i.e. whether texture formed by lines i.e. line-
like or blob-like.
Regularity – Regularity defines the interval in which patterns repeated. If patterns are repeated in regular
interval then the texture is regular else it is said to be Irregular.
Roughness – Roughness defines the whether the surface is rough or smooth.
In these six features, Coarseness, Contrast and Directionality correspond to strong human perception and
these features are calculated pixel-wise by creating 3-D histogram of these three features. Estimation of
these three features are described in subsequent sections.
Coarseness relates to distances of notable spatial variations of grey levels, that is, implicitly, to the size of
the primitive elements (texels) forming the texture. The proposed computational procedure accounts for
differences between the average signals for the non-overlapping windows of different size:
At each pixel (x,y), compute six averages for the windows of size 2k × 2k, k=0,1,...,5, around the pixel.
At each pixel, compute absolute differencesEk(x,y) between the pairs of nonoverlapping averages in the
horizontal and vertical directions.
At each pixel, find the value of k that maximises the difference Ek(x,y) in either direction and set the best
size Sbest(x,y)=2k.
Compute the coarseness feature Fcrs by averaging Sbest(x,y) over the entire image. Instead of the average of
Sbest(x,y, an improved coarseness feature to deal with textures having multiple coarseness properties is a
histogram characterising the whole distribution of the best sizes over the image.
Contrast measures how grey levels q; q = 0, 1, ..., qmax, vary in the image g and to what extent their
distribution is biased to black or white. The second-order and normalised fourth-order central moments of
the grey level histogram (empirical probability distribution), that is, the variance, σ2, and kurtosis, α4, are
used to define the contrast:

(17)
Where,
(18)

11

Vol 3, No.2, 2012
(19)

(20)
and m is the mean grey level, i.e. the first order moment of the grey level probability distribution. The value
n=0.25 is recommended as the best for discriminating the textures.
Degree of directionality is measured using the frequency distribution of oriented local edges against their
directional angles. The edge strength e(x,y) and the directional angle a(x,y) are computed using the Sobel
edge detector approximating the pixel-wise x- and y-derivatives of the image:
e(x,y)=0.5(|∆x(x,y)|+ |∆y(x,y)|) (21)
-1
a(x,y)=tan (∆x(x,y)/ ∆y(x,y)) (22)
where ∆x(x,y) and ∆y(x,y) are the horizontal and vertical grey level differences between the neighbouring
pixels, respectively. The differences are measured using the following 3 × 3 moving window operators:
−1 0 1 1 1 1
−1 0 1 0 0 0
−1 0 1 −1 −1 −1

A histogram Hdir(a) of quantised direction values a is constructed by counting numbers of the edge pixels
with the corresponding directional angles and the edge strength greater than a predefined threshold. The
histogram is relatively uniform for images without strong orientation and exhibits peaks for highly
directional images. The degree of directionality relates to the sharpness of the peaks:
(23)

where np is the number of peaks, ap is the position of the pth peak, wp is the range of the angles attributed to
the pth peak (that is, the range between valleys around the peak), r denotes a normalising factor related to
quantising levels of the angles a, and a is the quantised directional angle (cyclically in modulo 180o). Three
other features are highly correlated with the above three features and do not add much to the effectiveness
of the texture description.
The linelikeness feature Flin is defined as an average coincidence of the edge directions (more precisely,
coded directional angles) that co-occurred in the pairs of pixels separated by a distance d along the edge
direction in every pixel. The edge strength is expected to be greater than a given threshold eliminating
trivial "weak" edges. The coincidence is measured by the cosine of difference between the angles, so that
the co-occurrences in the same direction are measured by +1 and those in the perpendicular directions by -
1. The regularity feature is defined as Freg=1-r(scrs+scon+sdir + slin) where r is a normalising factor and each
s... means the standard deviation of the corresponding feature F... in each subimage the texture is partitioned
into. The roughness feature is given by simply summing the coarseness and contrast measures:
Frgh=Fcrs+Fcon . These features capture the high-level perceptual attributes of a texture well and are useful
for image browsing. However, they are not very effective for finer texture discrimination.
6. Artificial Neural Network
A neural network is a massively parallel distributed processor that has a natural propensity for storing
experiential knowledge and making it available for use. It resembles the brain in two respects [3] [4] [7]:
1. Knowledge is acquired by the network through a learning process.
2. Interneuron connection strengths known as synaptic weights are used to store the knowledge.
Benefits of neural network
Nonlinearity.
Input-output mapping.

12

Vol 3, No.2, 2012
Adaptivity.
Contextual information.
Fault tolerance.
VLSI implementability.
Uniformity of analysis and design.
Neurobiological analogy.
Model of a neuron
A neuron is an information-processing unit that is fundamental to the operation of a neural network. We
may identify three basic elements of the neuron model: [17] [18]

Figure 2:Non-linear model of a neuron.
A set of synapses, each of which is characterized by a weight or strength of its own. Specifically, a signal xj
at the input of synapse j connected to neuron k is multiplied by the synaptic weight wkj. It is important to
make a note of the manner in which the subscripts of the synaptic weight wkj are written. The first subscript
refers to the neuron in question and the second subscript refers to the input end of the synapse to which the
weight refers. The weight wkj is positive if the associated synapse is excitatory; it is negative if the synapse
is inhibitory.
An adder for summing the input signals, weighted by the respective synapses of the neuron.
An activation function for limiting the amplitude of the output of a neuron. The activation function is also
referred to in the literature as a squashing function in that it squashes (limits) the permissible amplitude
range of the output signal to some finite value.
Typically, the normalized amplitude range of the output of a neuron is written as the closed unit interval [0,
1] or alternatively [-1, 1].
The model of a neuron also includes an externally applied bias (threshold) wk0 = bk that has the effect of
lowering or increasing the net input of the activation function.
Since after feature tamura feature extraction data is in the form of numerical values, Artificial Neural
Network classifier suits well for classification. Also non – linearity of the data makes other traditional
classifiers like Bayesian and kth – nearest neighbor classifier inefficient compared to ANN classifier. Thus
in this system ANN classifier is used as classification tool.

13

Vol 3, No.2, 2012
7. Experimental results

Figure 3: Intermediate result
Figure 3 shows the intermediate result after corner detection. First image in figure 3 is the input image, 2nd
image displayed is histogram equalized image. From this histogram equalized image threshold points
detected and marked with red + marks as shown in third image of figure3.
From each seed point region is extracted and from extracted region tamura features are calculated. Each
feature vector consists of 4 features and n number of such feature vectors can be obtained from single
image which helps to prevent the system to be biased.
Extracted feature vectors are sent to neural network. The performance measurement with variable number
of hidden layer neurons with single layered feed forward back- propagation network is tabulated in table1:
Index Number of Percentage of correct
neurons classification
1. 20 96.4286%
2. 21 85.7143%
3. 22 92.8571%
4. 23 92.8571%
5. 24 96.4286%
Table 1- variable number of hidden layer neurons
The performance measurement with variable number of hidden layers with fixed number of neurons of 20
neurons in each layer, feed forward back- propagation network is tabulated in following table:
hidden layers classification
1. 1 96.4286%
2. 2 89.2857%
3. 3 85.7143%
Table 2- variable number of hidden layers

8. Conclusion
Even though there is no successful generalized neural network configuration, for a particular application
a neural network with acceptable level of accuracy can be designed by selecting suitable number of hidden
layers, number of neurons per hidden layer and transfer and learning functions. The performance also
depends on the training function parameters like whether it is a batch training or one input at a time. Also
we have witnessed the advantages of neural network classifiers over other traditional classifiers like
Bayesian and k – nearest neighbor classifiers.
This design can be extended to estimate the number of carcinoma cells per unit area. This estimation

14

Vol 3, No.2, 2012
helps in automated diagnosis systems like blood purifier in case of blood cancer. Also this can extended to
take color image as input with more feature added to feature vector to increase the accuracy of the output.

References
[ 1]. “An Approach for Discretization and Feature Selection Of Continuous-Valued Attributes in Medical
Images for Classification Learning”.
[ 2]. Basavaraj .S. Anami1 and Vishwanath.C.Burkpalli 2 1. Principal, K.L.E.Institute of Technology, Hubli-
580030, India 2. Research Scholar, Basaveshwar Engineering College, Bagalkot – 587102, India.
[ 3]. “Texture based Identification and Classification of Bulk Sugary Food Objects”, ICGST-GVIP Journal,
ISSN: 1687-398X, Volume 9, Issue 4, August 2009.
[ 4]. Bing Gong, School of Computer Science and Technology Heilongjiang University Harbin, China, “A
Novel Learning Algorithm of Back-propagation Neural Network” ,2009 IITA International Conference
on Control, Automation and Systems Engineering.
[ 5]. Weilin Li, Pan Fu and Weiqing Cao, “Tool Wear States Recognition Based on Genetic Algorithm and
Back Propagation Neural Network Model”, 2010 International Conference on Computer Application and
System Modeling (1CCASM 2010)
[ 6]. Acharya and Ray, “Image Processing: Principles and Applications”, Wiley-Interscience 2005 ISBN 0-
471-71998-6
[ 7]. Russ, “The Image Processing Handbook”, Fourth Edition, CRC 2002 ISBN 0-8493-2532-3
[ 8]. SIMON HAYKIN, Book on “Neural Networks”, 2nd edition, A comprehensive edition.
[ 9]. “Digital Image Processing” by Gonzalez & Woods 2nd edition
[ 10]. H. Tamura, S. Mori, and T. Yamawaki, "Texture features corresponding to visual perception," IEEE
Trans. On Systems, Man, and Cybernetics, vol. Smc-8, No. 6, June 1978.
[ 11]. J. Smith and S.-F. Chang, “Transform features for texture classification and discrimination in large image
database”.IEEE Intl. Conf. on Image Proc., 1994.
[ 12]. Castleman K R. “Digital image processing”. NJ: Prentice Hall, 1996.
[ 13]. Manjunath, B., Ma, W.: “Texture features for browsing and retrieval of image data”. IEEE Trans on
Pattern Analysis and Machine Intelligence 18 (1996) 837842
[ 14]. Alexander Suhre, A. Enis Cetin, Tulin Ersahin, Rengul Cetin-Atalay, “Classification of cell images using
a generalized harris Corner Detector”
[ 15]. R. Adams and L. Bischof, “Seeded Region Growing”, IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 16, pp. 641-647, 1994.
[ 16]. R. Haralick, “Statistical and structural approaches to texture”, Proceedings of the IEEE, vol. 67, pp. 786–
804, 1979.
[ 17]. C. Harris and M.J. Stephens. “A combined corner and edge detector”. In Alvey Vision Conference, pages
147–152, 1988. D.
[ 18]. Ballard and C. Brown, “Computer Vision”, Prentice-Hall, Inc., 1982, Chap. 6.
[ 19]. Davies, “Machine Vision: Theory”, Algorithms and Practicalities, Academic Press, 1990, Chap. 18.
[ 20]. A K Jain, “Fundamentals of Digital Image Processing”, Prentice-Hall, 1986, Chap. 9.
[ 21]. D. Vernon Zhi-Hua Zhou, Yuan Jiang, Yu-Bin Yang, Shi-Fu Chen , “Lung Cancer Cell Identification
Based on Artificial Neural Network Ensembles”, National Laboratory for Novel Software Technology,
Nanjing University, Nanjing 210093, P.R.China. Artificial Ingelligence in Medicine, 2002, vol.24, no.1,
pp.25-36. @Elsevier.
[ 22]. Leonard Fass , “Imaging and cancer: A review”, GE Healthcare, 352 Buckingham Avenue, Slough, SL1
4ER, UK Imperial College Department of Bioengineering, London, UK.
[ 23]. Jinggangshan, P. R. China , “Application of Neural Networks in Medical Image Processing” ISBN 978-
952-5726-09-1, Proceedings of the Second International Symposium on Networking and Network
Security (ISNNS ’10), 2-4, April. 2010, pp. 02.

15

Vol 3, No.2, 2012

Figure 1: ANN - C3 system architecture

neurons classification
1. 20 96.4286%
2. 21 85.7143%
3. 22 92.8571%
4. 23 92.8571%
5. 24 96.4286%
Table 1: Variable number of hidden layer neurons

hidden layers classification
1. 1 96.4286%
2. 2 89.2857%
3. 3 85.7143%
Table 2: Variable number of hidden layers

16

Artificial neural network based cancer cell classification

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Artificial neural network based cancer cell classification (20)

More from Alexander Decker (20)

Recently uploaded (20)

Artificial neural network based cancer cell classification