Content uploaded by Anunchai Assawamakin
Author content
All content in this area was uploaded by Anunchai Assawamakin on Nov 05, 2014
Content may be subject to copyright.
Estimation of Band Level Resolutions
of Human Chromosome Images
W. Sethakulvichaia, S. Manitpornsutb, M. Wiboonratc, W. Lilakiatsakund, A. Assawamakine, S. Tongsimaf
IT Graduate School, Mahanakorn University of Technology, Bangkok, Thailanda, d
Department of Computer and Multimedia Engineering, University of the Thai Chamber of Commerce, Bangkok, Thailandb
College of Graduate Study in Management, Khon Kaen University, Thailandc
National Center for Genetic Engineering and Biotechnology (BIOTEC), Klong Luang, Pathumthani, Thailande, f
wisan@sriracha.coma, suparerk_man@utcc.ac.thb, montwi@kku.ac.thc, woraphon@mut.ac.thd,
anunchaiice@gmail.come, sissades@biotec.or.thf
Abstract—Current estimation methods for band level resolutions
of human chromosome images in cytogenetic laboratories are
time consuming and required experienced specialists to manually
perform. To alleviate this problem, in this paper, a computerized
approach to estimate band level resolution is proposed. The
intensity gradient profile and sign profile of chromosome images
are utilized to count the number of bands. Then band level
resolutions of chromosome images are classified into three
categories: 400-, 500-, and 550-band levels by using k-nearest
neighbor algorithm. The experimental results show the accuracy
of the proposed algorithm. We also provide a discussion on how
to improve the overall accuracy.
Keywords; image processing, band level resolution, feature
extraction, medial axis determination, k-nearest neighbor.
I. INTRODUCTION
Automated systems are now widely integrated to almost
everywhere surrounding our everyday life. Most of the systems
are aiming to reduce a time consumption of repeated and
tedious tasks. Chromosome analysis is one of those laborious
and time-intensive activities, routinely conducted in
cytogenetic laboratories [1] – [5]. The evolution of computer
graphic in the 21st century has changed the way of human
chromosome analysis. Integrating computerized system such as
image processing into this area has been an open research topic
since 1960s [2]. However, there are plenty of aspects for
research in this area required for improvement, e.g. information
from the chromosome images, processing speed, accuracy rate,
reliability, easy to adaption, and competitive cost.
The band level of resolution of each human chromosome
image is one of the additional information that laboratories
need to specify in karyotype report. Actually this information
represents the total number of bands on chromosomes
approximately half of the cell [6]. Most laboratories determine
this number by identifying the specific landmarks on the
specific chromosomes. From many literature reviews,
automatic estimation system for band level resolution of human
chromosome image has not been mentioned much in this
research field. Furthermore, it is expected that knowing the
band level resolution before the classification process will help
managing the proper dataset for the chromosome classification
system.
The objective of this study is to demonstrate the method for
estimating the band level resolution of human chromosome
image automatically. This paper is organized as following.
Section II provides a background of the cell cycle, the
karyotype analysis and the important of the levels of resolution
of the chromosome images. The source of dataset and the
proposed algorithm are explained in Section III.
Experimentation and results are revealed in Section IV. Finally,
the explanation of results and discussion are summarized in
Section V.
II. BACKGROUND
Cytogenetic laboratories are one of those specialized
laboratories where many special tests for study micro-
biological behavior of all living things are usually conducted
[7]. Karyotype analysis is a special activity, consisting of
several steps, i.e. collecting cell samples, culturing cell
samples, staining chromosomes, acquiring image of
chromosomes, segmenting image into individual chromosomes,
classifying chromosomes into their classes, rearranging and
representing into karyogram and finally analyzing the
karyogram [8].
The results of the karyotype analysis give valuable
information in many aspects such as, for parental screening, sex
determination, identification of abnormality in chromosomal
diseases such as Down’s syndrome, Klinefelter syndromes,
malignant tumor, cancer, leukemia, ect [7].
G-banding is the technique that cytogenetic laboratories use
to stain the specific content of chromosomes so that they
express their distinctiveness pattern on their chromosome
images [1]. This information is the main information that
cytogenetic staff use to classify the chromosome into their
correct classes. Although, there are many more others staining
techniques available for karyotype analysis in this present, and
some of them give more information of individual
chromosomes, g-banding technique is the most widely used for
the karyotype analysis due to its low cost and short preparation
duration. In addition, information achieved after the analysis is
adequate for medial practitioner to make decisions. Therefore,
this work will be focused on the images of the g-banding
chromosome only.
276
978-1-4673-1921-8/12/$31.00 ©2012 IEEE
2012 Ninth International Joint Conference on Computer Science and Software Engineering (JCSSE)
During the cell culturing process, cells go through their
cells cycle processes as shown in Figure 1 [8]. Chromosomes
in each cells, during mitosis process, has different stages of
condensation [9]. Most of currently automatic karyotype
analysis research is focusing on the images of metaphase
chromosomes because the chromosome at this stage is in the
most condensed form. However, in reality, cytogenetic need to
conduct further examination on the pro-metaphase or even
prophase chromosomes in many cases [10] – [11]. At the
earlier stages of mitosis, the structures of chromosomes are less
condensed and the overall length of all chromosomes in one
cell is longer than in metaphase cell. Furthermore, the total
numbers of bands along the length of all chromosomes are also
different. The more number of bands the earlier the stage of
mitosis.
The level of resolution of human chromosome image is an
important number to indicate the stages of chromosomes being
analyzed. It is normally determined by counting the number of
bands seen in a haploid set (22 autosomes + X and Y) [6].
However, this is really difficult to perform in reality. Most
laboratories currently evaluate the level of resolution by
identifying the land marks on the chromosome images and
scoring the karyotype image by using the scoring table as
shown in Table I. At least three of the land marks criteria to be
obtained to indicate the band levels resolution [10]. There is
also a requirement to specify the number of band resolution in
the karyotype report to confirm that the analysis is appropriate
for the referral reason. Table II shows lowest standard
acceptable for a given reason for referral in constitutional
TABLE I G -BANDING RESOLUTION EVALUATION SCORE [10].
Band
resolution
Land marks Criteria
300 band 2 dark bands on 8p (8p12 & 8p22)
3 dark bands on 10q (10q21, 10q23, 10q25)
20p12 visible
22q12 distinct
400 band 3 dark bands on mid-4q (q22-28)
3 dark bands mid-5q (5q14, 5q21, 5q23)
2 dark bands on 9p (9p21 & 9p23)
13q33 distinct
500 band 7q33 & 7q35 distinct
3 dark bands on 11p (11p12, 11p14, 11p15.4)
14q32.2 distinct
4 dark bands on 18q (18q12.1, 18q12.3, 18q21.2,
18q22)
550 band 5q31.2 distinct
8p21.2 visible
2 dark bands on 11pter (11p15.2 & 11p15.4)
22q13.2 distinct
700 band 2p25.2 distinct
2q37.2 distinct
10q21.1 and 10q21.3 resolve
17q22-q24 resolves into 3 dark bands
TALBE II MINIMUM G -BANDING RESOLUTION FOR REGERRAL REASON
[10].
Reason for referral Minimum level of
resolution of G-
banding images
quality
Confirmation of aneuploidy e.g. direct
lymphocyte, direct CV or solid tissue culture
preparation
< 300 band
Exclusion of known large structural
rearrangements e.g. lymphocyte, solid tissue, CVS
direct preparation or amniotic fluid cell
preparation
300 band
Identification and exclusion of small expected
structural rearrangements e.g. lymphocyte, solid
tissue, CVS culture or amniotic fluid preparation
400 band
Routine amniotic fluid and CV culture
preparations
400 band
Abnormal ultrasound scan associated with AF, CV
and solid tissue referrals
500 band
Blood referrals, not covered by exclusion criteria 550 band
For microdeletion syndroms (when no FISH probe
is available)
700 band
Figure 1 Cell cycle and stages in cell division
Figure 2 Ideograms of G-banding patterns for normal human chromosome
number 1 at five different levels of resolution, approximately 300-,400-,
550-, 700-, and 850-band levels [6].
277
analysis as recommended by the Association for Clinical
Cytogenetics [10].
The standard ideograms of chromosomes class 1 shown in
Figure 2 provide schematic representations of chromosomes
corresponding to approximately 300, 400, 550, 700 and 850
bands [6]. It is clearly noticeable that different resolution
yielding different banding or intensity profile.
Apart from this, different levels resolution of G-banding
images acquired from different staining process creates a
tedious problem to most laboratories. Sometime, a slightly
different of period when capturing the metaphase moment
could cause a different in level resolution of banding patterns.
Figure 3 shows the different banding levels of chromosome
number 1, acquired at the different moments during the cell
division process. The longer chromosomes give many more
bands than the shorter one [11].
Since the shape and the number of bands of the same class
chromosomes are varied according to their relative mitosis
stages, classification of chromosomes at the level resolution
different from the resolution of trained database is prone to a
huge number of error rates. Estimation the band levels
resolution of chromosomes prior to the classification stage is a
promising approach to enhance the accuracy of the
classification when incorporate this module into the system to
help selecting the proper band level database to the analyzing
subject. In the next section, automatic band level resolution
estimation algorithm is described in detail.
III. PROPOSED METHODOLOGY
Figure 4 shows the proposed automatic system for band
level resolution estimation of human chromosome images.
There are several steps as described below.
A. Image Enhancement
The first step is to eliminate noises, enhance image
contrast and adjust the brightness of the images by using the
following methods:
• Median filtering – to get rid of pepper noises.
• Contrast enhancement – to adjust a low-contrast gray-
scale image.
Figure 4 Diagram shows the proposed automatic chromosomes band counting. system.
Figure 3 G-banded chromosomes of the class no.1 arranged in increasing
order of resolution from approximately the 400- to the 700-band levels
G-bands per haploid karyotype
400 450 500 550 600 650 700
Metaphase Prometaphase Prophase
278
• Intensity enhancement – to compensate non-uniform
brightness and contrast behavior of the images by using
histogram equalization.
• Segmentation – to convert the image to binary by using
Otsu’s thresholding algorithm. Then, closing, opening
and filling image operations are performed to mask out
the individual chromosomes region and smooth the
image boundary.
B. Medial axis determination
A simple thinning algorithm is applied here to determine an
axis of an individual chromosome, as shown in Figure 6(a).
However, the length of thinning line is shorter than that of the
chromosome, thus requiring line extension. Bresenham’s line
algorithm is applied to extend the tip points of the thinning
lines to tip point of the chromosome, as illustrated in Figure
6(b). Finally, as displayed in Figure 6(c), the medial axis of
individual chromosome image is formed.
C. Intensity profile extraction
Intensity profile is extracted from the individual
chromosome image on the perpendicular lines of the medial
axis. In Figure 7(a), the perpendicular lines So are generated
from the local slopes of the medial axis. However these
perpendicular lines might overlap and therefore result in
incorrect intensity profile. The orientations of the perpendicular
lines So are rearranged to eliminate the overlapping lines as
shown in Figure 7(b).
Let is the 1 vector of rearranged slopes. The So is
rearranged by using a moving mean value as:
1
·1,,·1 2,3,4, …, 1
(1)
The and are the smoothing weights calculated from
the difference between the current slope and previous slope and
the current slope and the next slope respectively.
The sampling is done by collecting all the intensity value on
each perpendicular line and calculating the average value along
this line except three pixels that closed to the medial axis. This
is because most of the chromosome images appeared to be
“white” in the middle of the chromosome body along its medial
axis as noticeable in Figure 7 (b). The original intensity profile
as shown Figure 8 (a) is then smoothed by using a moving
average filter. Let is the original intensity profile and is the
smoothed intensity profile.
The span of the moving average of this work is set to 5
(experimentally selected). Transformation from the original to
the smoothed profile can be expressed as;
1
1,,1 2
2,1,,1,2 3,4,5, . ., 2
1,,1 1
(2)
D. Intensity gradient determination
The intensity gradient profile is the rate of changes in
intensity value along the medial axis line defined as:
(3)
(a) (b)
Figure 5 Preprocessing process; (a) original chromosome image,
(b) binary image.
(a) (b) (c)
Figure 6 Finding medial axis; (a) Thinning, (b) Extending tips, (c) Medial
axis.
(a) (b)
Figure 7 Finding the perpendicular lines along the medial axis;
(a) perpendicular lines from the local slopes, (b) rearranged perpendicular
lines.
279
where, is a vector returns the one-dimensional numerical
gradient of . corresponds to
, the differences in x
(chromosome medial axis) direction. Figure 8(c) shows a
sample of intensity gradient profile derived from the intensity
profile.
E. Band counting
The bands are counted by detecting the changes in direction
of the gradient intensity profile. The band is divided into 2
types; white bands and dark bands. The rate of change in
intensity between the edge of white and dark band should be
high. The algorithm is set to count the changes in sign (+/-) of
the gradient profile along the medial axis. The gradient profile
is converted to the sign profile as
1, 0
0, 0
1, 0 (4)
where, is the sign profile calculated from a gradient profile. Figure
8(d) shows the sample of sign profile.
F. K-Nearest Neighbors
The k-NN is one of the supervised data mining techniques
for classifying data into classes. The method is based on the
majority of the k (k = number of neighbors) nearest distances
between the features of the training dataset to the features of
the testing data [13]. Four features which are number of bands,
length, width and aspect ratio are used in the proposed system.
The number of neighbors was set to vary from k = 1 to 15. The
features of test dataset are then compared to the nearest trained
features.
IV. EXPERIMENTS AND RESULTS
The dataset of chromosome images are from the Center for
Medical Genetics Research, Rajanukul institute, Thailand. In
this research, 210 karyogram images are collected according to
the availability of information of the band level resolution. All
chromosomes were stained using G-banding technique. These
chromosomes were cultured, stained, captured by cytogenetic
specialists. Karyograms were also prepared by using the current
analysis system in the laboratory. However, image intensity
and scaling are very diverse. The band level resolution of each
images were identified by experienced staff. Three band levels
resolutions are investigated, i.e. 400, 500 and 550 band levels.
TABLE III SUMMARY OF THE DATASETS
DESCRIPTION BAND RESOLUTION TOTAL
400 500 550 cell %
Training 56 56 56 168
80
Testing 14 14 14 42 20
Total 70 70 70 210 100
Seventy metaphase cell images for each band level
resolution are employed in experiment, or 210 images in total.
Each metaphase cell image consists of 46 individual
chromosomes. These samples are divided into two sets, training
and testing dataset as shown in table III. All data in each band
level resolution are randomly selected and divided into 80% for
training and 20% for testing, respectively. The experimentation
is divided into 2 sections, band counting and band estimation.
A. Band counting
The test is conducted to count number of bands of
individual chromosomes from the training dataset of 400, 500
and 550 band level resolutions, respectively. The output of each
run contains the following features.
1) The computed bands:
This feature is computed by the technique as described in
the previous section.
2) The Aspect ratio (AR):
This feature is defined by the ratio of length and the width
of each chromosome.
Figure 8 Output profiles from the band estimation algorithm, (a) intensity and smoothed intensity profiles, (b) original input, (c) gradient intensity profile, and
(d) sign profile.
(a)
(b)
(c) (d)
280
(5)
This feature indicates the morphological shape of the
chromosome. The high value of AR means long and/or narrow
chromosome, thus implying the high value of band level
resolution. On the other hand, the low AR implies the low value
of band level resolution.
The results are plotted as shown in figure 9, which is the
plot of counted number of band vs. the aspect ratio of all
chromosomes from each band level datasets. This graph shows
the three clusters of these dataset in a 2D feature space spanned
by the relation between numbers of bands counted from the
proposed algorithm and theirs relative aspect ratios. The border
areas of these clusters are distributed dispersedly but the center
of populated areas of each cluster can be observed.
B. Estimation of Band Level Resoultion
The band level resolution is estimated by implementing k-
nearest neighbor (k-NN) classification model. The features are
number of bands, length, width and aspect ratio, respectively.
The number of neighbors was set to vary from k = 1 to 15. For
simple image representation, Figure 10 displays the
relationship of the target node and the neighbor nodes in the 2D
feature space of the Count Number and the Aspect Ratio
features.
Figure 11 shows the accuracy percentage of the
classification results of each band level and the overall test set.
The maximum accuracy percentage is 72%, 93%, and 79% for
classification of 400-, 500-, and 550-band level resolutions.
The overall accuracy is 76% at k = 3, 9 and 11, respectively.
V. DISCUSSION AND FUTURE WORK
This work has shown the method for extracting features
from individual chromosomes, i.e. intensity profile, gradient
intensity profile, sign profile and aspect ratio. By using k-NN
technique, the classification of human chromosome band level
resolution can be identified with the maximum accuracy of
76%. The main issues that affect the accuracy of the system are
the chromosome image quality and the accuracy of band level
resolution from experts.
The actual chromosomes in the sample glass slides are in
three dimensions. When they are captured by a camera, they
are projected to two-dimensional images. Some of them are
overlapped or crossed by others. The more of chromosomes are
overlapped, the lower quality of the chromosome images is.
The low quality of chromosome image affects the sign profile
and therefore decreasing the classification accuracy of the
proposed system.
Another effect on the system accuracy is the classification
results from experts. Experts do not count the actual number of
bands in the chromosome image since actual counting takes too
much time for human. Instead, they notice the landmark of
some chromosomes to identify the band level resolution. This
is an estimation approach. Unfortunately, the chromosome
images have wide variation in terms of landmarks, length,
width, and band intensity.
To account for the aforementioned issues, our future work
is considering more features from the chromosome images, e.g.
the total number of bands in one cell (instead of individual
chromosome), the normalized band intensities, scaling effect
and compensation of band counting in the overlapped areas.
Other classification techniques are also under investigation. In
addition, the interactive cross validation with the experts will
be conducted to reduce human errors.
Figure 9 Aspect Ratio VS number of bands counted from chromosome
number 1 from 400-,500-, and 550- band levels.
Target node
5 neighbors
Figure 10 The 5 neighbor nodes that have distances of features closet to the
target node.
Figure 11 The classification results when vary the k number from 1 to 15
281
ACKNOWLEDGMENT
Special thanks to Dr. Verayuth Praphanphoj, M.D., Center
for Medical Genetics Research, Rajanukul Insititute, Thailand
and all his staff for supporting data and information.
REFERENCES
[1] X. Wang, B. Zheng, S. Li, J.J. Mulvihill, and H. Liu, “Automated
classification of metaphase chromosomes: optimization of an adaptive
computerized scheme,” Journal of Biomedical Informatics, vol. 42, pp.
22-31, 2009.
[2] F.C.A. Groen, T. K. Kate, A.W.M. Smeulders and I.T. Young, “Human
chromosome classification based on local band descriptors”, Pattern
Recognition Letters, vol. 9, pp. 211-222, April 1989.
[3] B. Legrand, C.S. Chang, S.H. Ong, S.Y. Neo and N. Palanisamy,
“Chromosome classification using dynamic time warping”, Pattern
Recognition Letters, vol. 29, pp. 215-222, 2008.
[4] J.H. Kao, J.H. Chuang and T. Wang, “Chromosome classification based
on the band profile similarity along approximate medial axis”, Pattern
Recognition, vol. 41, pp. 77-89, 2008.
[5] J. Piper and E. Granum, “On Fully Automatic Feature Measurement For
Banded Chromosome Classification”, Cytometry, vol. 10, pp. 242-255,
1989.
[6] ISCN2009: An internationalsystem for human cytogenetics
nomenclature in D.G. Harnden, H.P. linger (Eds.), Cytogenetics and Cell
Genetics, Karger, Basel, 2009
[7] H.C. Macgregor, An Introduction to Animal Cytogenetics, Chapman &
Hall, 1993.
[8] C. Starr and B. McMillan, Human Biology, 5th ed., Books/Cole,
Thomson Learning, 2003.
[9] G. Karp, Cell Biology, International edition, 2nd ed., McGraw-Hill
International Book Co., 1984.
[10] Association for Clinical Cytogenetics, Professional Guidelines for
Clinical Cytogenetics; General Best Practice Guidelines (2007) v1, 04,
March. 2007.
[11] A. Carothers and J. Piper, “Computer-aided classification of human
chromosome: a review”, Statistics and Computing, vol. 4, pp. 161-171,
1994.
[12] B. Lerner, H. Guterman, I. Dinstein and Y. Romem, “Medial axis
transform-based features and neural network for human chromosome
classification”, Pattern Recognition, vol. 28, no. 11, pp. 1673-1683,
1995.
[13] M.S. Serbanescu, “A k-nearest neighbor approach for chromosome
shape classification”, Annals of the University of Craiova, Mathematics
and Computer Science Series, vol. 37(3), pp. 142-146, 2010.
282
