Histopathology Grading Of Breast Cancer

Using Artificial Neural Network

Agoes Santika H Anny Setijo Rahaju Riries Rulaningtyas
Magister of Biomedical Department Anatomy Suryani Dyah Astuti
Engineering Phatology Physics Department
Faculty of Postgraduate Faculty of Medical Faculty of Science and
Surabaya, Indonesia Surabaya, Indonesia Technology
ag.santika.com [email protected] Surabaya, Indonesia
air.ac.id [email protected]
[email protected]

Abstract—Breast cancer is the leading uncontrolled development of breast tissue

type of malignant tumor observed in (tissue) Chintya, 2009).
women and effective treatment depends on In Indonesia, the most common type of
its early diagnosis. The gold standard of breast cancer histopalogy is Invasif Ductal
breast cancer is histopathologic Carsinoma Mamae. This type is a group of
examination of cancer cells. The malignant epithelial tumors that have the
determination of the grading in breast character of invasion (infiltration) to the
cancer is determined by three factors: surrounding tissue and is characterized by the
pleomorphic, tubular formation and cell tendency of experiencing extensive
mitosis. In this paper refer to pleumorfic metastases elsewhere. In general, this tumor is
and tubular formation by Breast cell
adenocarcinoma and is believed to originate
histopatologi images. The proposed system
from the epithelium of the breast
consist of four major steps : preprocessing,
segmentation, feature extraction and parenchyma, especially cells derived from the
classification. In preprocessing grayscale terminalis duktus lobular unite (TDLU).
and segmentation are applicated. Feature The prognosis of breast cancer is
extraction using Gray level Cooccurence associated with a wide range of clinical
Matrix (GLCM) and the 4 features are conditions and pathological factors. Grade
contrast, energy, entropy and homogeneity. histology is one of the important factors in
The final step is identification using determining the prognosis of the patient.
Backpropagation. Some of important There are two grading systems widely used in
parameters will be variated in this process recent years, the Bloom-Richardson system
such as learning rate and the number of and the Black system. The three histological
node in hidden layer. The research result factors studied hace been presented
suggest that extraction feature in 4 feature elsewhere(Bloom, 1950) and breifly as
is the best accuracy for based on classes follwors:
81,1% and especially in accuracy is 80%. a. Degree of structural differentation as
shown by the presence of tubular
Index key—Histopatologic breast cancer, arragement of the cells.
kmeans, GLCM, Backpropagation
b.Variantion in size, shape and staining of
BINTRODUCTION the nuclei
c. Frequency of hyperchromatic and mitotic
figures
reast cancer (breast cancer) is one of the
highest causes of human deaths besides
cervical cancer (servics). This case can occur
in men and women. Breast cancer is
somewhat the most common disease affecting
women, although men also have the
possibility of experiencing this disease with a
ratio of 1 in 1000. Breast cancer is a disease
where there is excessive growth or
 We got the data form Pathology Anatomy
of RSUD Soetomo. There are 150 patient
data with different grade classification. The
size image 804 x 565. The dataset comprised
150 tisuee images with 50 low, 50
intermediate, 50 hight grade breast cancers.
There are two data;primary data and training
data. Homedata is training data where this
data is intelligent system arficial neural
netwok. While for training data used for
Fig. 1. Image sel (J. Biomed. Eng. 1988, network training.
Vol. 10, October)

Hystologi is the study of the microscopic Nuclei Segmentation

anatomy of cell and tissues of organisme.
Hystological analysisis performed by Automatically segment nuclei in breast cancer
examining a thin slice(section) of tisuue under histopatological image is a hybrid active
a light or electron microscope. contour which combined boundary and region
Hystopathology, the microscopic study of information.First step, normalization methodis
biopsies to locate and classify disease, has employed to reduce the color variations in
roots in both clinical medicine and in basic tissue appearance due to variability in tissue
science. In histology image analysis for preparation. In this study segementasi image
cancer diagnosis, histopathologis visually using K means method. K-Means is an
examine the regularities of cell shapes and algorithm Clustering Repeatedly. The K-
tisuue distribution, decide whether tissue Means algorithm begins with a random
regiobs are cancerous and determine the selection of K, K here is the number of
malignancy level. clusters Which you want to establish. Then set
While histology image interpretation the values of K at random, for a while the
continues to be the standard for cancer value becomes the center of Cluster Or
diagnosis, current computer technology commonly called the centroid, mean or
towards this the task falls behind clinical "means". Calculate the distance of each
existing data on each centroid using the
need. Manual analysis of histology tissues is
Euclidian formula until it finds the closest
still the primary diagnosis method, and
distance from each data with the centroid.
depends heavily on the expertise and
Classify each data based on its proximity to
experience of histopathologist. Such manual the centroid. K means clustering groups
intervention has the disvantages of being very imange points into K cluster by minimizing
time consuming and being difficult to grade in the objective function.
reproducible manner;empirically, it is known
that there are substantial intra-and-inter
observation variations among experts. The
large and impratical demand on experts time
to interpret the images are making computer
assisted diagnosis(CAD) systems increasingly Where Ii is the intensity of image point xi in
important. Compared to manual the class sk
analysis,computer based system may provide
rapid and consistent cancer detection and Feature Extraction
grading result. In this research, into two After image segmentation, image
sequential subsystem tissue preparation and features are extracted from the regions of
image production following by image interest to detect and grade potential deseases.
processing and analysis. In this case, we using Gray Level Co-
Occurance Matrix. The GLCM method is
included in statistical methods where in
MATERIAL
statistical calculations use the gray degree
distribution (histogram) by measuring the
contrast, granularity, and roughness of a backwarded in the hidden layer then from the
region of the neighboring relationships hidden layer to the input layer.
between pixels within the image This After the cancer classification or grading, we
statistical paradigm is of unlimited use, can evaluate the performance of CAD system
making it suitable for unstructured natural by comparing their result with ground truth.
textures of sub patterns and set rules The grading result on the testing subset can be
(microstructures). classified into ine of four categories: true
Gray-Level Co-occurrence matrix is positive(TP), False positif(FP), True
the most widely used method for texture negatif(FN), False Negatif (FN).
analysis. This method was introduced by
Haralick in 1973 which is part of a NASA- EXPERIMENTAL DESIGN
supported project that is precisely NASA
Goddard Space Flight Center. Haralick uses CAD system are increasingly
images from NASA ERTS to classify with an employed in histopatology with the objective
accuracy of over 80%. of instant and consistent disesase
The GLCM matrix is calculated from identification analisis. In the experiment, the
the pixel value paired and has a certain dataset comprised 150 histopatologi
intensity value. Let d be the distance between preparation image consists of 56 grade 1, 70
two pixels (x1, y1) and (x2, y2) and Ѳ tetha is grade 2, 40 grade 3. The rest of 120 cases use
defined as the angle between the two, the training set and 30 testing data set. The
GLCM matrix is the spatial distribution of Pd proposed segmentation method is employed
Ѳ (i, j). on the captured BCH imaged by MATLAB.
1. Homogeneiti
The proposed method is compared with
Indicates the homogeneity of the intensity
comparable methods from the published
variation in the image. Homogeneiti
literature.
Equation:

Histologi Segmentation Fiture extraction

image K-means GLCM
................................. (1) Homogieniti,
2. Energy energy, entrophy,
contrass
Energy is a GLCM feature used to measure
concentration of intensity pairs on GLCM
matrix, and is defined as follows: . Identification
grading results

............................(2)
Fig 2. Conseptual framework
3. Contrast
Contrast to improve image contrast :
The segmentation result with K means
...............................(3) clustering are shown in Fig. 2
4. Entropy
Shows the size of the shape irregularity:
�9 = −Ʃ� Σ� �(�, �)log (�(�, �)).................(4)

Backpropagation is one of the artificial

neural network methods used to determine the
grading of breast cancer. Each
backpropagation network unit located in the Fig 3. K means clustering segmentation
input layer is connected to each unit in the
hidden layer connected to each unit in the K means clustering groups image point
output layer. When the network is given an into K cluster by minimizing the objective
input pattern as a training pattern, then the function. The image points in a d-dimensional
pattern goes to the hidden layer units for by certain probability density function where
subsequent output will give response as the dense regionscorrespond to the local maxima
output of artificial neural network. When the (modes) of the underlying distribution. Image
output is not as expected, the output will be
point associated to the same mode (by a
 greadient ascent procedure) are grouped into training stage was done by the variation of the
one cluster. backpropagation parameters of the maximum
After image segmentation, image epoh, hidden layer, and learning rate. In
features are extracted from gray level co- addition to the variation of parameters in the
occurance matrixs. Traditional features with training also conducted input variation to find
object size and shape,topological or graph out the texture features that really easy to
based features intensity and color features, recognize by backpropagation. The highest
texture feature(homogeneiti, energy, entrophy input accuracy was obtained from the training
and contrass). This is shown in the table 1, using the entropy feature of 81.5789%,
followed by homoginity of 78.9473% and
then energy of 77.6315%. The training using
variation of two highest accuracy inputs was
obtained from a variation of energy + IDM
feature of 84.2105%, followed by entropy +
IDM of 82.8947%, then energy + entropy of
81.5789%

Table 1 GLCM result

GLCM
CONCLUSIONS
Kode
No. Homogi
image
nitas Energi Entrophy Kontras This paper describes a computer-aided
diagnosis system for quantitative analizyng
0.51192 the Breast Cancer Histopatologi imanges. K-
1 A1 2 0.57663 1,72944E+05 4869.57 Means segmentation methode is developed for
0.52789 0.55312 detection and segmentation of cell nuclei. The
2 A11 4 9 1,51681E+05 922.96
segmentation result of proposed method have
0.55176 0.58621
3 A12 3 5 1,49035E+05 2285.49 higher segmentation accuracy. Shape, size and
0.61001 texture feature selection based form
4 A13 0.53804 8 1,51737E+05 2389.1 GLCM(Gray Level Co-Occurance Matrixs).
The Identification use Artificial Neural
In this study more emphasis on the calculation Network can show significantly accuracy
of pleumorfix and tubular cells. Input ANN
Comes from the output of fiture extraction Acknowledgment
that the output is a number. In this paper, the
This research was support by member of
image are categorized with three main type,
Anatomi Pathologi Laboratory, RSUD
namely Grade 1,
Soetomo Surabaya.

References
[1] Anonim, 2008, Deteksi Kanker Leher
Rahim Dan Kanker Payudara, http://
www.depkes.go.id/index.php
option=news&task=viewarticle&sid=2965,1
5 September 2009

[2] Adenoid cystic carcinoma, 2010 [cited on

2010, July 27]. Available from:
http://www.wikipedia.com/.

[3] Automated grading of breast cancer

Fig. 4 Grade I grafik
histopathology using cascaded ensemble
with combination of multi-level image
Backpropagation training in this study aims to
features, Tao Wana, Jiajia Caob, Jianhui
get the weight that is able to produce the
Chenc, Zengchang Qinb,2016
output calculation as close as possible to the
target so as to produce high accuracy. The
[4] Chandrasoma P, Taylor CR. The Breast,
initial weights used were randomly selected
Concise Pathology, McGraw-Hill
with a range of values from -1 to 1. The
International Edition; 2001(3): 815-29.
 Irwanto, Yudhi Purwananto dan Rully
[5] Ellis IO, et all. Invasive breast carcinoma. Soelaiman,2012
In: Tavasolli FA, Devilee P. Pathology and
Genetic of Tumours of the Breast and [15] Paulus, E., Nataliani, N. Cepat mahir
Female Genital Organs, WHO Classification GUI Matlab. Penerbit Andi,
of Tumours, IARC Press; 2003: 18-19, 23- Yogyakarta.2007.
43.
[16] Patten BM. Human Embryology, Mc
[6] Grading in histopathology, Simon S Cross, Graw-Hill; Philadelphia; 2004(2): 240-41.
Ksenija Benes, Timothy J Stephenson
Robert F Harrison,2011 [17] Pettinato, Guido, Carlos J. Manivel,
Invasif Micropapillary Carcinoma of the
[7] Grading and Prognosis of Invasive Ductal Breast, Am J Clin Pathol ; 2004, 121 : 6 :
Mammary Carcinoma by Nuclear Image 854 - 66.
Analysis in Tissue Sections, K. D. Kunze,
G. Haroske, V. Dimmer, W. Meyer and F. [18] Putra, D. Pengolahan Citra Digital.
Theissig, Institute ofPathological Anatomy, Penerbit Andi, Yogyakarta. 2010
Medical Academy "Carl Gustav Carus';
Dresden, GDR,1989 [19] Rosai J. The Breast, Rosai and
Ackermans’s Surgical Pathology, Volume
[8] Histology image analysis for carcinoma two, Mosby; 2004(9): 9-10, 1797, 1802-18.
detection and grading, Lei He, L. Rodney
Long, Sameer Antani, George R. [20] Rosen PP. Invasive Mammary
Thoma,2011 Carcinoma, Breast Pathology, Volume I,
Lippincott; Philadelphia; 2001(2): 236 - 56.
[9] Histological Grading And Prognosis In
Breast Cancer A Study Of 1409 Cases Of [21] Schnitt SJ, Mills RR, Hanby AM,
Which 359 Have Been Followed For 15 Oberman HA. The Breast. In: Mills SE, et
Years,H. J. G. Bloom And W. W. all, 2004, Stenberg’s Diagnostic Surgical
Richardson, From The Meyerstein Institute Pathology, Volume IA, Lippincott Williams
Of Radiotherapy And The Bland-Sutton & Wilkins; 2004(4): 330 -67.
Institute Of Pathology Of The Middlesex
Hospital, London, W,1957 [22] Sloane JP. The Breast, Biopsy Pathology
of The Breast, Biopsy Pathology series 24,
[10] Lester SC. The Breast. In: Kumar V, Arnold, 2001(24): 62 - 9.
Abbas AK, Fausto N. Robbins and Cotran
Pathologic Basis of Disease. Elsevier [23] Slide Apocrine Carcinoma of Breast,
Saunders; Philadelphia; 2005(7): 1120, 2010 [cited on 2010, July 27]. Available
1142-49. from: http://www.webpathology.com/.

[11] Kissane J M. The Breast, Anderson’s [24] Thor AD, Osunkoya AO. The Breast. In:
Pathology, Volume II, Mosby, 1990(9): 1726 Rubin E, Strayer DS. Farber. Editors.
– 48 rubin’s Pathology: Clinicopathologyc
Fondation of Medicine. JB Lippincott
[12] Montag A, Kumar V. The Female Genital Williams & Wilkins; Philadelphia; 2008(5):
System and Breast. In: Kumar V, Abbas 842-53.
AK, Fausto N, Mitchell RN. Robbins Basic
Pathology. Saunders Elsevier; Philadelphia;
2007(8): 743-49.

[13] Nasser, S., Alkhadi, R. A Modifed Fuzzy

K-means Clustering using Expectation
maximization. University of Nevada Reno,
Reno RV 89557, USA, hal 471,2008.

[14] Optimasi Kinerja Algoritma Klasterisasi

K-Means untuk Kuantisasi Warna Citra