2018 4 th International C onference o n Frontiers o f Signal Processing

Classification of Metaphase Chromosomes Using Deep Learning Neural Network

Kiruthika P J a y a n th i K B
D ept. o f E C E D ept. o f ECE,
K.S. R angasam y College o f T echnology K.S. R angasam y College o f Technology
T am ilnadu, India T am ilnadu, India
e-m ail: kiruthi1012@ gm ail.com e-m ail: jayanthikb@ gm ail.com

Nirmala Madian
D ept. o f E C E
Sri Sakthi Institute o f E ngineering and Technology
Tam ilnadu, India
e-m ail: nirm alaooty@ gm ail.com

Abstract—Karyotyping of Banded Metaphase Chromosomes is im ages are tak en in this study as they are seen easily and
one of the preliminary steps used in cytogenetics to analyze the have unique light and dark bands.
chromosomes for diagnostic purposes. Deep learning is a G enerally there are 23 pairs o f chrom osom es in every
subfield of machine learning concerned with structure and hum an cell [1]. The first 22 pairs are autosom es and the 23rd
function of brain. It exploits a way to automate predictive p air is the sex chrom osom e. C hrom osom al abnorm alities are
analysis. The key aspect of deep learning is that the layers of related to the structure and num ber o f these 23 pairs o f
features are not designed by human engineers. They are chrom osom es. K aryotyping [2, 3] is a standard profile o f
learned from data using a general purpose learning procedure. chrom osom es as show n in Fig. 1.
This paper proposes a convolution based deep learning to
classify the chromosomes for automated karyotyping. The
developed architecture allows us to train and test images that
helps in predicting the chromosome abnormality. The
performance analysis is based on loss and accuracy curves and
the graphical representation clearly exhibits better
classification results for this architecture.

Keywords-deep learning; chromosome; karyotyping;

convolutional neural networks

I. In t r o d u c t io n

K aryotyping o f chrom osom es com bines the study o f

chrom osom e m orphology and genetic diseases. T hough the
process o f karyotyping requires m ore m anual effort, tim e Figure 1. Karyotyped Image
consum ption, hum an visual perception and dom ain expert, it
rem ains as a very im portant task fo r cytogeneticists to A ll 23 pairs o f chrom osom es are differentiated using
perform this process efficently. A n approach for addressing various staining procedures. They are G (G iem sa) banding,
such problem is to create an autom atized system to classify Q (Quinacrine) banding, R banding and C banding as in Fig.
the chrom osom es using a classifier. R ecently, D eep 2. O u t o f all, G b anding is preferred because they give a
learning tool is m ainly em ployed to perform the various distinct p attern o f light and dark bands.
tasks by the process o f autom ation. K aryotyping is
perform ed b ased o n the various features extracted from
chrom osom es. T he im portant features to recognize the
chrom osom e are centrom ere position, length o f the
chrom osom es, centrom ere index, banding patterns.
Chrom osom es are visualized as a continuous sequence o f
light and dark bands and they becom e evident by staining
techniques. K aryotyping allow s us to determ ine w hether
there are any abnorm alities o r structural problem s in them.
N orm ally, K aryotyping process is carried out during the
Figure 2. Stained Images (a) G band image (b) Q band Image (c) R band
m etaphase stage o f cell division. M etaphase chrom osom e image (d) C band Image

978-1-5386-7853-4/18/$31.00 ©2018 IEEE 110

 D an Cire§an et.al presented the concept o f m ulticlass
deep neural netw ork fo r im proving the im age classification
benchm arks recognition fo r traffic signs [1]. Swati et.al
proposed a different straightening techniques th at w as
applied to chrom osom e im ages prior to classification o f
chrom osom es using siam ese m ethods [2]. B ut it is used only
fo r lim ited data. A ltan et.al provided a com prehensive review
o f past and recent research in the area o f neural netw orks
based autom atic hum an chrom osom e classification system
and feature extraction [3]. W enzhong Y an and L ei B ai
proposed a classification algorithm based on deep b elief
netw orks by extracting the features based o n H ilbert H uang
transform [4]. Perform ance o f classification w as exam ined to
diagnose the subjects w ith o r w ithout the coronary artery
disease. X ingw ei W ang et.al applied a higher order neural
netw ork fo r the classification o f hum an chrom osom es [5]. A
new autom ated chrom osom e karyotyping schem e w ith A N N
b ased two decision layer classifier [6] w as dem onstrated.
T his w as applied only to norm al chrom osom es and not tested
w ith abnorm al or cancerous m etaphase chrom osom es. The
p aper [7] presented a sum m arization o f m ajor advanced
classification m ethods and techniques used fo r im proving
classification accuracy. A lso it discussed im portant issues
affecting the success o f im age classifications.
In [8], neural netw orks have been applied to perform all
m ajor stages o f hum an chrom osom e analysis nam ely feature
extraction, im age segm entation and classification. A
com parative result analysis o f SVM , D T and K N N classifier
fo r im age classification w as perform ed [9]. The application
o f probabilistic neural netw ork to the classification o f norm al
hum an chrom osom es [10] w as described. V arious methods
w ere reported fo r classification o f chrom osom es b ased on
banding p attern [11, 12]. T he recognition rates achieved in
this study are superior to those reported using either the
m axim um likelihood o r back propagation neural netw ork
techniques.
D eep learning technology has been applied to m edical
diagnosis b ased o n a large am ount o f accum ulated X -rays,
CT scans, lab data and M RIs. The proposed w ork exploits
deep learning convolutional neural netw ork fo r classification
o f sex chrom osom es from the 23 pairs o f chrom osom es.
II. Ma t e r ia l s a n d Met hods
A. Deep Learning Neural Network
D eep learning is a type o f m achine learning in w hich a
m odel learns to perform classification tasks directly from
im ages, text, o r sound. D eep learning is usually im plem ented
using neural netw ork architecture. The term “deep” refers to
the num ber o f layers in the netw ork— the m ore layers, the
deeper the network. T raditional neural netw orks contain only
2 o r 3 layers, w hile deep networks can have hundreds. It is
necessary to develop m ore pow erful discrim inative
optim ization techniques to fin d b etter feature extraction
m odels at each layer. A deep neural netw ork com bines
m ultiple nonlinear processing layers w ith sim ple elem ents
operating in parallel and inspired by biological nervous
systems. It consists o f an input layer, several hidden layers,
and an output layer. T he layers are interconnected v ia nodes,
111
o r neurons, w ith each hidden layer having an output and
previous layer as its input. D eep-learning netw orks are
distinguished from the m ore com m onplace single-hidden-
layer neural netw orks by their d epth (i.e,) the num ber o f
node layers through w hich data passes in a m ulti-step
process o f pattern recognition.
B. Convolutional Neural Networks
The convolutional neural netw ork (CN N , o r ConvN et) is
one o f the m ost p opular algorithm s fo r deep learning w ith
im ages and video. A convolutional neural netw ork (CN N ) is
a type o f artificial neural netw ork used in im age recognition
and processing that is specifically designed to process pixel
data. The layers o f a c N n consist o f an input layer, an output
layer and a hidden layer that includes m ultiple convolutional
layers, pooling layers, flattened layer, fully connected layers
and norm alization layers.
B asically, the im age classification process involves two
steps, nam ely training and testing. The num ber o f output
layers in the CN N depends o n the num ber o f classes th at are
to be labeled and classified. In CNN, convolution operation
perform s a vital role. Convolution puts the input im ages
through a set o f convolutional filters, each o f w hich activates
certain features from the im ages. The convolution layer and
pooling layer help in the extraction o f features from patches
o f im age. The convolutional layer com putes the output
feature m ap by the follow ing equation
Z k = f ( l l 1 W k * x k) (1)
w here x denotes input im age , Z k is the K th output feature
m ap , W k is the w eight o f Kth feature map, * is a two
dim ensional convolutional operator and f(.) represents
nonlinear activation function.
P ooling sim plifies the output by perform ing nonlinear
dow nsam pling on the input data. N onlinear dow nsam pling is
done to im prove the extraction o f features. It progressively
reduces the spatial size o f the representation, thereby
reducing the feature m ap dim ensionality and com putational
com plexity o f the network, w hich in turn can im prove the
perform ance.
M ax pooling is perform ed w hich generalizes the results
from the convolutional filter, m aking the detection offeature
invariant to scale o r orientation changes. Thus, the
dim ensions o f the feature m ap reduces from (m,n) to
(m/k,n/k). k needs to be chosen in consistence w ith the
dim ensions o f the input feature map. In this, the input image
is dow n sam pled by a factor o f 2 along each direction. This
value is chosen to have m inim um pixel loss and get a precise
region w here the features are located, thereby reducing the
com plexity o f the m odel w ithout reducing its perform ance.
R ectified linear unit (ReLU ) allow s fo r faster and more
effective training by m apping negative values to zero and
m aintaining positive values. These operations are repeated
over tens or hundreds o f layers, w ith each layer learning to
detect different features. The convolution layer is
param eterized by the num ber o f filters, size o f each filter and
the activation function used. In the proposed architecture, 3
convolutional layer, 3 subsam pling layer and 2 fully
connected layer is used. The num ber o f filters used is 32, the
size o f each filter being 3*3, and the activation function is
the rectifier function, w hich is used at fully connected layer
Figure 3. Block diagram of proposed work
The last stage o f a convolutional neural netw ork is a
classifier, called as dense layer. It needs individual features
i.e., it needs a feature vector to perform classification. In
order to convert the output o f convolutional part o f the CN N
into a 1D feature vector, flattening operation is done. It gets
the output o f pooled im age pixels (2D array), flattens all its
structure to create a one dim ensional single feature vector to
b e used by the dense layer fo r the final classification. A fully
connected layer connects the set o f nodes got after the
flattening step. Fully connected layer perform s non-linear
transform ations o f the extracted features and classifies the
inputs. This process com pletes the building up o f the
convolution neural netw ork model. Im age classification w ith
CN N w orks quite w ell w hen enough training data is
p rovided.
Tw o m ajor categories o f im age classification techniques
include unsupervised (calculated by softw are) and
supervised (hum an-guided) classification. T raining im ages
are labeled in a supervised w ay by an analyst, b u t the feature
learning and classification are autom atically done by
softw are in an unsupervised way. W hile hum an visual im age
interpretation techniques rely o n shape, size, pattern, tone,
texture, shadows, and association, digital im age
interpretation relies m ainly o n color, i.e. o n com parisons o f
digital num bers found in different bands in different parts o f
an im age. In deep-learning networks, each layer o f nodes
trains o n a distinct set o f features b ased o n the previous
lay er’s output. The further advancem ent into the neural net,
the m ore com plex the features the nodes can recognize, since
112
to determ ine probabilistic confidence value o f output.
they aggregate and recom bine features from the previous
layer.
III. Re s u l t s a nd D is c u s s io n s
D iagnosis by deep learning are typically m ore objective
and accurate. To test the architecture, chrom osom e im ages
are collected and total num ber o f chrom osom e im ages are
random ly divided into groups fo r training, validation and test
sets respectively. The training set consists o f 175 im ages and
the testing set consists o f 83 im ages. E ach chrom osom e
im age is assigned a label from the 24 categories. F o r all the
experim ents, the resolution o f the im age is set to 64*64.
D ata augm entation is done to increase the num ber o f data
in the dataset. The netw ork is tested on various datasets
w hich in turn is tested fo r different num ber o f epochs
(iterations). H igher the learning rate, lesser num ber o f epochs
is required. L earning rate determ ines how quickly o r slowly
the update o f w eights have to takes place. F or sm aller values
o f learning rate, too m any iterations are needed to converge
to the b est values. Typically, learning rates are assigned at
random b ased on the user. A ll m odels w ere trained using
convolutional neural networks. To regularize the netw ork
training, a sufficiently large num ber o f epochs are provided
w hile training each model. M odels are trained fo r different
values o f epochs. B y observing the validation results (i.e)
loss and accuracy at each epoch, epoch can be varied to
obtain the highest validation accuracy. T his is im plem ented
using keras w ith tensorflow as backend.
 (b), it can be seen that the validation accuracy becam e
slightly stable after 4-5 epochs and rarely increases at certain
epochs. In the beginning, validation accuracy linearly
increases w ith loss, then it does not increase much. The
training accuracy is high and training loss is quite low. B ut
the validation loss and accuracy are not better com pared to
training loss and accuracy, im plying that the m odel is over
fitting.

Figure 4. Workflow of the proposed work

A fter the m odel is created, com pile it using an

optim ization algorithm . A dditionally loss type is specified
depending on the num ber o f classes to be classified. In this Figure 6(a). Plot between epoch and loss
proposed work, num ber o f classes is 2, (x and y chrom osom e)
and loss type is considered as binary. A ccuracy is the
m etric used fo r analyzing w hile training the model.
The m odel is trained using fit () function by storing the
results o f the function. It can be used to p lot the accuracy and
loss function plots betw een training and validation to analyze
the perform ance o f the model. T he m odel is evaluated and
graph is plotted fo r loss and accuracy betw een training data
and validation data. Probability value is obtained w hich
determ ines the type o f chrom osom e. In this case, if the
probability score is less th an 0.5, the predicted result is X
else Y w ill b e displayed.
Epoch 24/25
100/100 [==============================] - 5s 46ns/step - lo ss: 6.0752 - acc: 6.
9725 - va l_loss: 0.0101 - val_acc: 1.0000
Epoch 25/25 Figure 6(b). Plot between epoch and accuracy
100/106 • 5s 47ns/step - lo ss: 0.0526 - acc: 0.
9825 - val lo ss: 0.0061 ■ v a la c c : 1.0000
/hone/research/anaconda3/lib/python3.6/slte-packages/sklearn/cross_validation.py To overcom e this problem , dropout layer, a regularizing
¡41: Deprecationwarning: This nodule was deprecated in version 0.18 in favor o f param eter is added to m ake the m odel perform b etter w hile
the nodel_selection nodule in to which a l l the refactored classes and functions a
re noved. Also note th a t the in te rfa ce o f the new CV ite ra to rs are d iffe re n t fro
keeping all the other layers unchanged. B y adding the
n th a t o f th is nodule. This nodule w ill be removed in 0.20. dropout layer in this architecture, the loss and accuracy are
"This nodule w ill be removed in 0 .2 0 .", Deprecationwarning)
fairly consistent w hich is show n in Fig. 7(a) and 7(b).
predicted: [[0.00236736]] x

Figure 5. (a) output for X chromosome

Epoch 25/25
100/190 - 5s 47ns/step • loss: 0.0714 • acc:
9725 • val.loss: 8.8124 • val.acc: l.M M
/hwe/researck/anaconda3/Ub/p)rtton3.6/slte -packapes/sktear r/cr »ss_»aUdat Ion.
:41: DeprecattoeMarnlng: This module »as deprecated in wrston 0.181* fa w o
the nodel.selectUn nodule into thick a ll the refactored classes and functions
re noved. also note that the interface of the ne» (V iterators are different f
n that of this nodule. This nodule udii be renewed tn 0.20.
'this nodule t i l l be renoved 1» 0.20.’ . Oeprecettoodarntnj)
predicted: fii.9954995U v___________________________________________
Figure 5. (b) output for Y chromosome

Fig. 5 gives the classification output for the test

chrom osom e b ased o n the probability score. F rom the Fig. 6

113

Figure 7(b). Plot between epoch and accuracy
M ostly, m any classification algorithm s w ork w ell w hen
sm all am ount o f data are available. A ccuracy o f classifiers
goes dow n w hen large dataset is taken fo r classification. B ut
deep learning perform s b etter even it com es fo r image
classification w ith large am ount o f data. In the proposed
work, the chrom osom e prediction gives favorable results for
sex chrom osom es. F ew preprocessing steps like im age
straightening is adopted to im prove the perform ance. In
future work, deep learning convolutional neural netw ork can
b e em ployed to classify all 22 pairs o f chrom osom es.
IV. Co n c l u s io n
In this proposed w ork, classification o f chrom osom es is
done using convolutional neural networks. The input
considered fo r CN N is the individual m etaphase
chrom osom es. R ectified linear u n it (ReLU ) is the activation
function used fo r extracting the features in CNN. The
extracted feature helps in the classification o f chrom osom es.
The proposed w ork gives an accuracy o f 100% fo r sex
chrom osm es b u t not as favourable fo r autosom es as the
chrom osom es do n ot have a proper size and structure. In the
future w ork, a better deep learning algorithm can be
em ployed to extract better features and the netw ork w ill be
trained and tested w ith large num ber o f datasets for
im proved perform ance and accuracy.
Ac k n o w l ed g men t
T his project is supported by U niversity
Com m ission, India.
Re f e r ences
[1] Dan Cire§an, Ueli Meier, Jürgen Schmidhuber, Multi-column Deep
Neural Networks for Image Classification, CVPR '12 Proceedings of
the 2012 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 3642-3649, 2012
[2] Swati, Gaurav Gupta, Mohit Yadav, Monika Sharma, Lovekesh Vig,
Siamese Networks For Chromosome Classification, 2017 IEEE
conference on Computer Vision Workshops (ICCVW), 2017
[3] Altan, Novruz Allahverdi, Yakup Kutlu Diagnosis of Coronary
Artery Disease Using Deep Belief Networks, European journal of
engineering and natural sciences, Volume 2, Issue 1, pp. 29-36, 2017
[4] Wenzhong Yan, Lei Bai, Algorithms for Chromosome
Classification Engineering, 5, 400-403, 2013
[5] Xingwei Wang, Bin Zheng, Shibo Li, John J. Mulvihill, Marc C.
Wood, and Hong Liu, Automated Classification of Metaphase
Chromosomes: Optimization of an Adaptive Computerized Scheme,
Biomed Inform. 42(1): 22-31, 2009.
[6] D. Lu & Q. Weng ,A survey of image classification methods and
techniques for improving classification performance, International
Journal of Remote Sensing, 28, 5, 823-870, 2007
[7] Boaz Lerner, Toward A Completely Automatic Neural Network
Based Human Chromosome Analysis, IEEE Trans Syst 28(4):544-
552, 1998
[8] M. Zardoshti-Kermani and A. Afshordi, Classification of
Chromosomes Using Higher-Orde Neural Networks, Proceedings of
ICNN’95- International Conference on Neural Networks, IEEE, 1995.
[9] Sandeep Kumar, Zeeshan Khan, Anurag Jain, A Review of Content
Based Image Classification using Machine Learning Approach,
International Journal of Advanced Computer Research, Volume-2
Number-3 Issue-5, 2012
[10] Walter P. Sweeney Jr., Mohamad T. Musavi,' and John N. Guidi,
Classification of Chromosomes Using a Probabilistic Neural Network,
Journal of the international society for advancement of cytometry,
Volume 16, issue1,Pages 17-24,1994
[11] Maximo E. Drets, Margery W. Shaw, “Specific banding patterns of
human chromosomes (heterochromatin/ Giemsa stain/ chromosome
bands)”, 1971, Proc. Nat. Acad. Sci. USA, Vol. 68, No. 9, pp. 2073­
2077
[12] J.H. Tjio, A. Levan, The chromosome number in man, Hereditas 42
(1956) 1-6.
[13] William James Kenneth Cummino, Noi MaCumming Nevin, A
system for automated chromosome analysis, Humangenetik 7 (1969)
349-350.
[14] J. Graham, J. Piper, Automatic karyotype analysis, Humana, Totowa,
NJ, 1994.
[15] S. B. Kotsiantis, Supervised Machine Learning: A Review of
Classification Techniques, Informatica 31 (2007) 249-268
[16] Chrysa Daiou, Alexandras Lambropoulos, Christoforos Markou,
Christos Maramis, Anastasios Delopoulos, Automatic Chromosome
Classification using Support Vector Machines
[17] T Arora, R Dhir, A review of metaphase chromosome image selection
techniques for automatic karyotype generation Medical & biological
engineering & computing, Springer, 2016
[18] F Abid, L Hamami, A survey of neural networks based automated
systems for human chromosome classification, Artificial Intelligence
G rants Review, Springer, 2018.
114