Engineering, 2013, 5, 400-403
http://dx.doi.org/10.4236/eng.2013.510B081 Published Online October 2013 (http://www.scirp.org/journal/eng)
Algorithms for Ch romosome Classification*
Wenzhong Yan, Lei Bai
Department of Computer, North China Institute of Science and Technology, Beijing, China
Received 2013
ABSTRACT
Automated chromosome classification has been an important pattern recognition problem for decades. In order to im-
prove the performance of automated chromosome classification, artificial intelligence and machine learning methods
have been widely used in the computer-assisted chromosome detection and classification systems. This paper is focused
on these algorithms, especially on artificial neur al network (ANN) and wavelet transform algorithms. The princ iple and
the realization of these algorithms are analyzed. Results of these algorithms are compared and discussed.
Keywords: Chromosome; Classification; ANN; Wavelet; M-FISH
1. Introduction
Chromosomes are genetic information carriers and chro-
mosome analysis constitutes an important procedure in
clinical and cancer cytogenetics studies. Chromosome ka-
ryotyping refers to the classification and subsequently a
formatted display of the chromosomes found in a cell
spread. A karyotype is required to assign each chromo-
some to one of 24 classes (22 autosomes and two sex
chromosomes). Figure 1 shows a sample result of the
karyotype. Since karyotyping is a time consuming pro-
cedure, computer-based classifiers have been proposed.
Most of these classifiers make use of an intuitive trans-
formation of the chromosome image density distributions
into a set of features to be used by some sort of statistical
discriminator. These types of classifiers have not shown
high perf o rmance re sults [1,2].
In order to improve the performance of automated
chromosome classification, artificial intelligence and ma-
chine learning methods have been widely used in area.
Among them, artificial neural networks (ANN) and wave-
let transform algorithms are the most popular tools. This
paper is focused on these algorithms. The principle and
the realization of these algorithms are analyzed. Results
of these algorithms are compared and discussed.
2. Artificial Neural Network Based
Algorithms
1) Basic Th eor y
Artificial neural networks (ANN) have been developed
as generalizations of mathematical models of biological
nervous systems. The basic processing elements of neur-
al networks are called artificial neurons, or simply neu-
rons or nodes. In a simplified mathematical model of the
neuron, the effects of the synapses are represented by
connection weights that modulate the effect of the asso-
ciated input signals, and the nonlinear characteristic ex-
hibited by neurons is represented by a transfer function.
The neuron impulse is then computed as the weighted
sum of the input signals, transformed by the transfer
function. The learning capability of an artificial neuron is
achieved by adjusting the weights in accordance to the
chosen learning algorithm [3].
2) Classification Algorithms Based on ANN
Backpropagation training method is commonly used to
train ANNs. In multi-layer feed-forward ANNs, the num-
ber of output neurons is often fixed (from 1 to 24), but
the number of input neurons, hidden neurons, steepness
of the activation function, learning rate, momentu m term,
number of learning iter ation s and upp er boun d of train ing
error are all programmable. Determining these training or
optimization parameters is important for the performance
(a) (b)
Figure 1. (a) A metaphase cell spread; (b) A karyotype of
the chromosomes in (a).
*This research is
supported by the Fundamental Research Funds for the
Central Universities (2011A010).

W. Z. YAN, L. BAI
401
and robustness of an ANN used in chromosome classifi-
cation [4].
In order to improve the performance of traditional
multilayer ANNs, a number of other more sophisticated
neural networks have been proposed and tested in this
area.
A hierarchical multi-layer neural network with an error
back-propagation training algorithm has been adopted for
the automatic classification of Giemsa-stained human
chromosomes. Firstly, chro mosomes data is classified into
7 major groups based on their morphological features
such as relative length, relative area, centromeric index,
and 80 density profiles. Then each 7 major groups are
classified into 24 subgroups using each group classifier.
Figure 2 shows the two steps of chromosome classifica-
tion. The classification error decreased by using two
steps of classification and the classification error was
5.9% [5].
A fuzzy Hopfield neural network is a combination
model of neuro and fuzzy computing. Its main difference
from the traditiona l ANN is that it holds fuzzy clustering
capability and learning mechanism of acquiring know-
ledge about the targets (human chromosomes) from the
noisy training samples. It develops a Classifier with the
Fuzzy Hopfield Network (CFHN) to identify each ob-
served human chromosome and assign it to one of the 24
human chromosome classes. In a test involving 100 hu-
man chromosomes, the fuzzy Hopfield neural network
produces a very low u nidenti f ication rate of 3. 33 % [6].
3. Wavelet Transforms Based Algorithms
Some researchers have set out to explore the use of
wavelet-based band pattern descriptors for chromosome
classification. Compared with other methods, wavelet
transforms use different basis functions that lead to the
desirable property of characterizing and localizing signal
Figure 2. Architecture of the hierarchical multi-layer neural
network.
features simultaneously in both the space and transform
domains. Furthermore, they offer a means of signal re-
presentation that facilitates multi-resolution analysis [7].
An expanded basis function system that allows high
resolution decomposition of a signal is called the wavelet
packet transform. These are computed by iterating not
only down the lowpass scaling function branch of Mal-
lat’s Discrete Wavelet Transform (DWT ) algorithm tree,
but also down the highpass wavelet branch [7]. Thus, the
wavelet packet transform offers more basis functions for
signal analysis than the wavelet transform. Compared
with the wavelet transform, the wavelet packet transform
uses only full-width basis functions. They are all ortho-
gonal, and all but the first have zero area. These are in-
tuitively more satisfying weighting functions than the
ever-narrowing wavele t basis f unc t ions.
In one study, researchers describe their recent study to
employ wavelet packets as basis function sets to compute
chromosome band pattern features. During the study, they
evaluated a total of 28 wavelet packet basis function sets,
including the well known Haar and Daubechies’4 and
Daubechies’6 wavelet packets. They conducted experi-
ments on two benchmark chromosome datasets and com-
pare the experimental results with the results of the cur-
rently best performing Weighted Density Distribution
(WDD) method. Table 1 summarizes the experimental
results. For the sake of clarity and page limit, this table
includes only those of the well-known Haar, Daubechies’4
(D4), Daubechies’6 (D6), and the best-performing wave-
let packet (BPWP), in comparison to the WDD results
[8].
Another research proposes a method for chromosome
classification based on the chromosome shape analysis.
This approach is based on wavelet packet transform (WPT)
and best basis algorithm (BBA). First, the chromosome
image is preprocessed and binarized. Second, the contour
of the chromosome is detected and the signature of the
contour is produced. Figure 3 shows a binarized chro-
mosome, its contour and its signature of the contour.
Third, the signature is decomposed by the wavelet packet
transform and its best basis is found. Finally, the coeffi-
cients of the best tree of wavelet packet transform cor-
respondent to the signature of chromosomes are com-
pared in order to classify the chromosomes. The results
obtained show that the proposed method provides an
effective chromosome classification based on WPT and
BBA of the shape signature of the chromosomes [9].
Table 1. Summary of the chromosome classification expe-
rimental results based on the two data sets
WDD BPWP Haar D4 D6
Copenhagen set 96.8% 96.2% 95.7% 95.2% 94.0%
Genzyme set 85.3% 84.6% 83.8% 79.9% 80.0%

W. Z. YAN, L. BAI
402
(a)
(b)
Figure 3. (a) A binarized chromosome and its contour; (b)
The signature of the contour in (a).
4. Other Algorithms
Besides the ANN and wavelet transforms based algo-
rithms, there are also other algorithms used for chromo-
some classification.
One research is interested in classification of chromo-
somes from either complete or incomplete cells. Research-
ers investigate globally op timal algorithms for automated
classification and pairing of human chromosomes. Even
in cases where the cell data are incomplete as often en-
countered in practice, they can still formulate the prob-
lem as a transportation problem, and hence find the glo-
bally optimal solution in polynomial time. In addition, a
technique of homologue pairing via maximum-weight
graph matching is proposed. It obtains the globally op-
timal solution by forming all homologue pairs simulta-
neously under a maximum likelihood criterion, rather than
finding one pair at a time as in existing heuristic algo-
rithms. After the optimal homologue pairing, chromosome
classification can also be done by maximu m-weight graph
matching. This new graph theoretical approach to chro-
mosome pairing and classification is more robust than
the transportation algorithm [10].
Traditional chromosome imaging has been limited to
grayscale images. In the mid-1990s, a new technique for
staining chromosomes was introduced. It produced an
image in which each chromosome type appeared as a
distinct color [11]. This multispectral staining technique
is called multiplex fluorescence in-situ hybridization, or
MFISH, which made analysis of chromosome images
easier, not only for visual inspection of the images by
humans, but also for computer analysis of the images.
M-FISH uses five color dyes that attach to various chro-
mosomes differently to produce a multispectral image,
and a sixth dye that attaches to all chromosomes to pro-
duce a grayscale image. Thus, it is possible to envision
new and improved methods for the location, segmenta-
tion and classification of chromosome images by ex-
ploiting the color information in M-FISH images.
One study addresses the topics of segmentation and
classification of MFISH chromosome images. It intro-
duces a probabilistic model of M-FISH chromosomes
that allows for simultaneous segmentation and clas sifica-
tion. The additional information provided by multiple
spectra in chromosome images makes it feasible to dis-
tinguish chromosomes that overlap and touch within clus-
ters. Fig ure 4 shows the comparison of two types of
cluster information. Thus, researchers develop a joint
segmentation-classification algorith m that optimizes proba-
bilistic information obtained from the multispectral chro-
mosome pixels, and enables the decomposition of over-
lapping and touching chromosomes, and moreover, pro-
vides estimates of confidence in the chromosome seg-
mentation-classification [12].
Another study presents a new segmentation method
between chromosomes and background and a novel un-
supervised classification method based on a fuzzy logic
classifier specifically designed for M-FISH images. Uti-
lizing the chromosome boundaries, the initial classifica-
tion results improved significantly after the prior adjusted
reclassification while keeping the translocations intact.
Figure 5 shows the fuzzy logic classification and prior
adjusted reclassification. This study also presents a new
segmentation method that combines both spectral and
edge information. Ten M-FISH images from a publicly
available database were used to test our methods. The
segmentation accuracy was more than 98% on average
[13].
5. Discussion and Conclusion
The problem of automated chromosome classification
has been investigated in many studies. A large number of

W. Z. YAN, L. BAI
403
(a) (b)
Figure 4. Comparison of two types of cluster information. (a)
Boundary of cluster; (b) Multispectral information in cluster.
Figure 5. Fuzzy logic classification and prior adjusted rec-
lassification.
novel techniques have been investigated by a number of
research groups around the world. In this paper we re-
viewed some typical algorithms, such as ANN and wave-
let transform algorithms etc. We analyzed the principle
and the realization of these algorithms and also discussed
the results of these algorithms.
REFERENCES
[1] M. Zardoshti-Kermani and A. Afshordi, “Classification of
Chromosomes Using Higher-Orde Neural Networks”.
[2] O. Sjahputera and J. M. Keller, “Evolution of a Fuzzy
Rule-Based System for Automatic Chromosome Recog-
nition,” IEEE International Fuzzy System Conference
Proceedings, 1999, pp. 129-134.
[3] P. H. Sydenham and R. thorn, “Handbook of Measuring
System Design,” John Wiley & Sons, Ltd., 2005.
http://dx.doi.org/10.1002/0471497398
[4] J. Cho, “Chromosome Classification Using Backpropaga-
tion Neural Networks,” IEEE Engineering in Medicine
and Biology Magazine, Vol. 19, 2000, pp. 28-33.
http://dx.doi.org/10.1109/51.816241
[5] J. Cho, S. Y. Ryu and S. H. Woo, “A Study for the Hie-
rarchical Artificial Neural Network Model for Giemsa-
Stained Human Chromosome Classification,” Proceeding
of the 26th Annual International Conference of the IEEE
EMBS, 2004, pp. 4588-4591.
[6] X. Ruan, “A Classifier with the Fuzzy Hopfield Network
for Human Chromosomes, Intelligent Control and Auto-
mation,” Proceedings of the 3rd World Congress on In-
telligent Control and Automation, Vol. 2, 2000, pp. 1159-
1164.
[7] C. S. Burrus, R. A. Gopinath and H. Guo, “Introduction
to Wavelets and Wavelet Transforms,” Prentice-Hall, En-
glewood Cliffs, NJ, 1997.
[8] Q. Wu and K. R. Castleman, “Automated Chromosome
Classification Using Wavelet-Based Band Pattern De-
scriptors,” 13th IEEE Symposium on Computer-Based
Medical Systems, 2000, pp. 189-194.
[9] L. V. Guimaraes, J. A. Schuck and A. Elbern, “Chromo-
some Classification for Karyotype Composing Applying
Shape Representation on Wavelet Packet Transform,”
Proceedings of the 25th Annual International Conference
of the IEEE EMBS, 2003, pp. 941-943.
[10] X. L. Wu, P. Biy a ni and S. Dumitrescu, “Globally Op-
timal Classification and Pairing of Human Chromo-
somes,” Proceedings of the 26th Annual International
Conference of the IEEE EMBS, 2004, pp. 2789-2792.
[11] M. R. Speicher, S. G. Ballard and D. C. Ward, “Karyo-
typing Human Chromosomes by Combinatorial Multi-
fluor FISH,” Nature Genetics, Vol. 12, 1996, pp. 368-375.
http://dx.doi.org/10.1038/ng0496-368
[12] C. S. Wade, C. B. Alan and L. E. Brian, “Maximum-
Likelihood Techniques for Joint Segmentation-Classifi-
cation of Multispectral Chromosome Images,” IEEE
Transaction on Medical Imaging, Vol. 24, No. 12, 2005,
pp. 1593-1610.
http://dx.doi.org/10.1109/TMI.2005.859207
[13] H. Choi, K. R. Castleman and A. C. Bovik, “Segmenta-
tion and Fuzzy-Logic Classification of M-FISH Chromo-
some Images,” IEEE International Conference on Image
Processing, 2006, pp . 69-72.