CHROMOSOME REGION RECOGNITION

WITH LOCAL BAND PATTERNS

Toru Abe

Cyberscience Center, Tohoku University, Sendai, Japan

Chieko Hamada

Graduate School of Information Sciences, Tohoku University, Sendai, Japan

Tetsuo Kinoshita

Cyberscience Center, Tohoku University, Sendai, Japan

Keywords:

Chromosome image analysis, Region extraction, Region classiﬁcation, Local band pattern, Subregion search.

Abstract:

To make the visual examination of a chromosome image for various chromosome abnormalities, individual

chromosome regions have to be extracted from the subject image and classiﬁed into the distinct chromosome

types. To improve the accuracy and ﬂexibility in this process, we propose a subregion (local band pattern)

based method for recognizing chromosome regions in the image. This method regards each chromosome

region as a series of subregions, and iterates a search for subregions in the image consecutively. Consequently,

chromosome region classiﬁcation is performed simultaneously with its extraction for each subregion. Since the

dimensions and intensities of chromosome regions vary with every image, effective subregion searches require

templates whose dimensions and intensities correspond with those of chromosome regions in the image. To

develop an effective subregion search, we also propose a method for adjusting the dimensions of templates to

those of chromosome regions in the image and adapting the intensities in the image to those of the templates.

1 INTRODUCTION

The examination of chromosome images for various

chromosome abnormalities plays an important role in

many clinical practices, including treatment and pre-

vention of genetic disorders, radiation dosimetry, tox-

icology, etc (Carothers and Piper, 1994). To make the

visual examination of a chromosome image, individ-

ual chromosome regions have to be extracted from the

subject image and classiﬁed into the distinct chromo-

some types in advance.

To improve the accuracy and ﬂexibility in this

process, we propose a subregion (local band pattern)

based method for recognizing individualchromosome

regions in an image. This method regards each chro-

mosome region as a series of subregions, and iterates

a search for subregions in the subject image consec-

utively. As a result, chromosome region classiﬁca-

tion is performed simultaneously with its extraction

for each subregion. Since the dimensions and intensi-

ties of chromosome regions vary with every image, to

achieve effective subregion searches, the dimensions

and intensities of templates for subregion searches

are required to correspond with those of chromosome

regions in the subject image. To develop an effec-

tive chromosome subregion search, we also propose

a method for adjusting the dimensions (widths and

lengths) of templates to those of chromosome regions

in the image and adapting the intensities in the image

to those of the templates. Furthermore, to show the

effectiveness of the proposed method, we also present

the results of subregion search experiments on chro-

mosome images.

2 CHROMOSOME IMAGE

EXAMINATION

This section explains the general procedures for ex-

amining chromosome images and the difﬁculties in

those procedures.

Abe T., Hamada C. and Kinoshita T. (2009).

CHROMOSOME REGION RECOGNITION WITH LOCAL BAND PATTERNS.

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 49-56

DOI: 10.5220/0001534500490056

 SciTePress

2.1 Procedures for Examining

Every cell nucleus in a normal human being contains

46 chromosomes consisting of 44 autosomes and two

sex chromosomes. The autosomes are composed of

22 homologous pairs of chromosomes, and by con-

vention, numbered from 1 to 22. The sex chromo-

somes are referred to as X and Y. A normal human fe-

male has two X chromosomes, while a normal human

male has an X and a Y chromosome. Each chromo-

some has a narrow part, which is called a centromere,

and it divides the entire region into two parts. The

shorter part is called a short arm and the longer part is

called a long arm. With proper staining methods, such

as Giemsa staining (G-staining) method, a character-

istic series of light and dark bands appears along the

longitudinal axis of a chromosome (Figure 1 (a)). The

band appearance on a chromosome is called a band

pattern, and it is unique to each type of chromosome.

Usually the examination of a chromosome image

requires the following procedures (Graham and Piper,

1994):

1. Staining a set of chromosomes and capturing its

image.

2. Extracting individual chromosome regions from

the image.

3. Classifying the chromosome regions into the 24

types (1, 2, ..., 22, X, and Y).

4. Inspecting the region appearances for chromo-

some abnormalities.

To make the visual examination of a chromosome

image, individual chromosome regions are extracted

from the subject image, and the extracted regions

are classiﬁed into the 24 distinct chromosome types

(Figure 1 (b)). The dimensions of a chromosome

change with the stage in a cell division, and the inten-

sities of it change with staining conditions, therefore

the dimensions and intensities of a chromosome re-

gion vary with every image. Meanwhile, the relative

length, the relative centromere position, and the band

(a) (b)

Figure 1: (a) chromosome image, (b) classiﬁcation re-

sult (ZooWeb, 2003).

pattern of each chromosome type vary little with ev-

ery image. For this reason, the latter features are used

for the classiﬁcation (Harnden and Klinger, 1985).

According to the classiﬁcation result, abnormal-

ities of number, where there are one or more entire

chromosomes additional to or missing from the nor-

mal complement, can be detected. From the region

appearances (the band pattern on each chromosome

region), abnormalities of structure, where part of the

bands are lost (deletion), repeated (duplication), or

shifted (translocation), can be examined visually.

2.2 Difﬁculties in Examining

The existing methods perform chromosome region

extractions apart from chromosome region classiﬁca-

tions, and their classiﬁcation procedures suppose that

individual chromosome regions are extracted accu-

rately from a subject image beforehand (Groen et al.,

1989; Wu et al., 2005). However, chromosome re-

gions in the image frequently touch or overlap each

other, and have some parts difﬁcult to distinguish

them from the background. Consequently, the ac-

curate extraction of individual chromosome regions

from the image is not an easy procedure.

Although extracted regions can be classiﬁed into

several chromosome groups according to the relative

lengths and the relative centromere positions of them,

to discriminate between all 24 chromosome types,

the use of band patterns is required in the classiﬁ-

cation. The classiﬁcation methods using band pat-

terns are generally categorized into two approaches:

one is a global approach, and the other is a local

approach (Graham and Piper, 1994; Carothers and

Piper, 1994; Wu et al., 2005). In the global approach,

the band pattern on an entire region (the longitudi-

nal proﬁle of intensity in an extracted region) is de-

termined, and a chromosome type is assigned to the

region by comparing its band pattern with reference

band patterns (Piper and Granum, 1989; Wu et al.,

2005). Therefore, when aberrant bands appear partly

on a region because of various reasons (region extrac-

tion failure, region overlap, chromosome abnormali-

ties, etc.), it is difﬁcult to assign a chromosome type

correctly. In the local approach, local features such

as particular bands are determined in a region, and

they are used for the classiﬁcation. This approach

can partially reduce the aberrant band inﬂuence on

the classiﬁcation accuracy (Groen et al., 1989; Gra-

ham and Piper, 1994; Moradi and Setarehdan, 2006).

However, it is reported that the local approaches are

inferior to the global approaches in the classiﬁcation

accuracy (Wu et al., 2005). The conceivable reasons

for that are as follows:

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

• It is difﬁcult to determine local features accurately

in a chromosome region.

• Compared to the global approaches, the local ap-

proaches use fewer features for the classiﬁcation.

3 CHROMOSOME REGION

RECOGNITION WITH LOCAL

BAND PATTERNS

To overcome the problems in the existing methods,

we propose a subregion based method for recogniz-

ing individual chromosome regions in an image. This

method regards each chromosome region as a series

of subregions, and iterates a search for subregions in

the subject image consecutively.

In the method, a reference band pattern is prepared

for every chromosome type. Each reference band pat-

tern for an entire chromosome region is divided into

several parts, and they are used as the templates for

extracting and classifying the chromosome region. In

the following, the divided parts are referred to as lo-

cal band patters, and the m th local band pattern on the

chromosome type i is denoted by lbp

(i)

(Figure 2 (a)).

Firstly, the subject image is searched for the subre-

gion corresponding to a local band pattern (template).

If a subregion corresponding to lbp

(i)

is detected, sec-

ondly, the neighborhood of the detected subregion is

searched for the next subregion corresponding to the

adjacent lbp

(i)

m−1

or lbp

(i)

m+1

(Figure 2 (b)). By iterat-

ing the search for subregion consecutively, with the

ﬁrst detected subregion as the starting point, one sub-

region after another is detected, and the entire region

of a chromosome is determined in the image.

When a subregion corresponding to lbp

( j)

n+1

is de-

lbp

(i)

lbp

(i)

m+1

lbp

(i)

m−1

reference band pattern on chromosome i

local band patterns

(a)

lbp

(i)

lbp

(i)

m+1

lbp

(i)

m−1

(b)

lbp

(j)

lbp

(j)

n+2

lbp

(j)

n−2

lbp

(j)

n+1

lbp

(j)

n−1

lbp

(j)

n+3

NH1

NH2

(c)

Figure 2: (a) local band patterns, (b) extraction and classiﬁ-

cation with local band patterns, (c) control of template and

search area.

tected and lbp

( j)

n+2

cannot be found in the neighbor-

hood NH1, it is surmised that the aberrations (chro-

mosome region overlaps, chromosome abnormalities,

etc.) occur in NH1 (Figure 2 (c)). To deal with this

difﬁculty and complete the search for the entire chro-

mosome region, if the adjacent local band pattern can-

not be found in the neighborhood, the template and

search area are changed. For example, in Figure 2 (c),

the template is changed from lbp

( j)

n+2

to lbp

( j)

n+3

, and

the search area is extended from NH1 to NH2.

By taking these approaches, the following advan-

tages are expected in this method:

• As the consecutive search for a subregion, simul-

taneously with the extraction, the classiﬁcation is

performed on part of a chromosome region, and

the results of preceding searches are utilized for

the following searches.

• By controlling the template and search area, the

consecutive searches integrate features in the sub-

regions while reducing aberrant band inﬂuence.

4 ADJUSTING TEMPLATE

DIMENSIONS AND ADAPTING

SUBJECT IMAGE INTENSITIES

A subregion search is made by scanning a subject

image with a template and seeking in the image for

subregions where the mean-squared-error (MSE) to

the template are sufﬁciently small. To achieve ef-

fective subregion searches, the dimensions and inten-

sities of templates are required to correspond with

those of chromosome regions in the image. This sec-

tion presents a methodfor adjusting the dimensions of

templates to those of chromosome regions in the im-

age and adapting the intensities in the image to those

of the templates.

4.1 Adjusting Template Dimensions

While the dimensions of chromosome regions vary

with every image, the relative length of each chro-

mosome type varies little from one image to another

and the widths of chromosome regions are similar in

each image. The proposed method binarizes an im-

age by the intensities of pixels, and then determines

the width W of chromosome regions and the sum of

chromosome region lengths (total length L) in the bi-

narized image. The determined W and L are used for

adjusting the dimensions of templates.

Let p

and p

represent pixels corresponding to

the chromosome regions and the background in the

CHROMOSOME REGION RECOGNITION WITH LOCAL BAND PATTERNS

π/4

chromosome

region

(a)

= w

’

(b)

w’

(c)

Figure 3: Region width at boundary p

binarized image, respectively. To determine W, as

shown in Figure 3 (a), from every p

bordered on

the background, the Euclidean distances d

to the

other boundary are measured in eight directions (t =

0,1,... ,7). With d

and d

t+2

, the estimated width w

and its direction θ

are calculated at p

= d

× d

t+2



+ d

t+2

, (1)

= tan

−1

t+2

) + t × π/4, (2)

where d

= d

and d

= d

. For calculating stabil-

ity, d

and d

t+2

greater than a threshold Td are used

for computing (1) and (2). In addition to w

, from p

to the other boundary in the direction θ

, the actual

measurement w

′

is taken. As shown in Figure 3 (b),

where both boundaries are straight and parallel to

each other, w

and w

′

are both equal to the true width

of the chromosomeregion. However, as shown in Fig-

ure 3 (c), where boundaries curve or they aren’t par-

allel to each other, w

and w

′

are different and they

may differ from the true width. Therefore, the pro-

posed method accepts w

′

as the reliable width at p

only when e

= |w

− w

′

| is less than a threshold Te.

If more than one reliable w

′

is accepted at a pixel p

the width w

′

with the smallest e

is chosen as the re-

gion width at p

. By choosing w

′

at each p

bordered

on the background and counting the occurrence fre-

quency for every value of chosen w

′

, the most fre-

quently occurred value is determined as the width W

of chromosome regions in the image. Thus, by choos-

ing reliable w

′

and using them for counting the occur-

rence frequency, the proposed method can determine

the chromosome region width W stably.

The sum of chromosome region areas in the im-

age can be estimated as the total number S of pix-

els p

in its binarized image, and it is approximated

by the product of the region width W and the to-

tal region length L. Therefore, L can be determined

by L = S/W. Suppose that templates for subregion

searches are made from an reference image I

, and

let I

represent the subject image. In the proposed

method, to adjust the dimensions of templates:

• the width and total length of chromosome regions

subject image: I

template: T

(x, y)

(a)

overlap with p

(chromosome

regions)

chromosome

region

template

overlap with p

(background)

(b)

Figure 4: Subregion search with a template.

are determined in both I

and I

(they are denoted

by W

, L

, W

, and L

, respectively),

• the width of the templates are set to W

, and the

length of each template is multiplied by L

4.2 Adapting Subject Image Intensities

Let I

(x,y) and T(u, v) represent the intensities at

(x,y) in the subject image and at (u,v) in the tem-

plate, respectively. As shown in Figure 4 (a), when

the template is set at (x,y) and rotated by θ in the sub-

ject image, the MSE e

at (x,y) is computed by

(x,y) =

U−1

∑

u=0

V−1

∑

v=0



′

) − T(u,v)



, (3)

′

= x+ ucosθ− vsinθ, (4)

′

= y+ usinθ + vcosθ, (5)

where U and V are the width and length of the tem-

plate, respectively. The rotation angle θ is set to mini-

mize e

(x,y). The intensities of chromosome regions

change with every image, and they vary locally in an

image according to staining conditions. To achieve

effective subregion searches, the proposed method

adapts the intensities in the subject image so that the

MSE e

(x,y) to the template is reduced, and then uses

the adapted MSE ˜e

(x,y) for the subregion search.

As shown in Figure 4 (b), the region of a template

set in the subject image consists of two parts: one

part O

overlaps with the background, and the other

part O

overlaps with the chromosome regions in the

image. O

and O

can be determined from the pix-

els corresponding to the background p

and chromo-

some regions p

in the binarized subject image. The

adapted MSE ˜e

at (x,y) is determined by

˜e

(x,y) =



(x,y) +

(x,y)



, (6)

(x,y) =

∑

(u,v)∈O



′

) − T(u,v)



, (7)

(x,y) =

∑

(u,v)∈O



αI

′

) + β − T(u,v)



, (8)

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

where the sums of squared-error

(x,y) and

(x,y)

are computed in O

and O

, respectively. The in-

tensities are similar almost everywhere in the back-

ground, and it is necessary that

(x,y) is supplied

to ˜e

(x,y) as a penalty. Accordingly,

(x,y) is com-

puted from raw intensities I

′

) in the subject im-

age, although

(x,y) is computed from adapted in-

tensities αI

′

) + β. For every subregion, con-

stants α and β are set to minimize

(x,y), and they

are determined by

α =

∑

(u,v)∈O

′

)T(u,v)

−

∑

(u,v)∈O

′

)

∑

(u,v)∈O

T(u, v)

∑

(u,v)∈O

′

) −

∑

(u,v)∈O

′

)

(9)

β =

∑

(u,v)∈O

T(u,v) − a

∑

(u,v)∈O

′

)

(10)

where |O

| is the number of pixels in O

. If α not ex-

ceeding 0 is determined for any subregion, such sub-

region is excluded from the subregion search because

the band pattern of it is reverse to that of the template.

5 EXPERIMENTS

To demonstrate the effectiveness of the proposed

method for adjusting template dimensions and adapt-

ing subject image intensities, we have carried out sub-

region search experiments on chromosome images.

5.1 Chromosome Images

Experiments were carried out on the chromosome

images that are opened to public by the website

of the Wisconsin State Laboratory of Hygiene and

ZooWeb (ZooWeb, 2003). This site provides not only

(a) (b)

Figure 5: (a) classiﬁcation result, (b) binarized image.

original chromosome images but also their classiﬁca-

tion results. Examples of the original chromosome

image and its classiﬁcation result are shown in Fig-

ure 1. Although the proposed method can be applied

to the original chromosome images, it is difﬁcult to

evaluate the subregion search results in them. There-

fore, the experiments were conducted on the classiﬁ-

cation results, where every chromosome region was

extracted, classiﬁed, and arranged in standard order.

Figure 5 (a) and (b) show examples of the clas-

siﬁcation result and its binarized image. Thirty-one

classiﬁcation results were used in the experiments.

They consist of 19 female and 12 male chromosome

images. This set includes 9 normal chromosome im-

ages (46 chromosomes in each image) and 22 numer-

ical abnormal chromosome images (2 images with 45

chromosomes and 20 images with 47 chromosomes).

Each image is 768×576 pixels in size, and charac-

ters in it are removed beforehand. To conduct cross-

validation, the images were divided into two sets A

(16 images) and B (15 images) (when one set was

used as subject images, the other set was used as ref-

erence images and employed for making templates).

5.2 Templates

To make templates for subregion searches, ﬁrstly,

chromosome regions were extracted from the refer-

ence images in a chosen set, and the intensity proﬁle

was acquired in each extracted region. Secondary, for

each chromosome type, the average intensity proﬁle

was made from the acquired proﬁles, and it was used

as the reference band pattern. Finally, templates were

made by dividing the reference band patterns.

As shown in Figure 6 (a) and (b), to acquire the

intensity proﬁle in a chromosome region, the medial

axis is determined in each extracted region. On the

determined medial axis, average intensities are taken

perpendicularly to the medial axis (Figure 6 (c)), and

they are used as an intensity proﬁle (Figure 6 (d)).

For each chromosome type i, intensity proﬁles

(a) (b) (c) (d)

Figure 6: (a) chromosome region, (b) extracted region and

medial axis, (c) average intensities perpendicularly to the

medial axis, (d) intensity proﬁle of the chromosome region.

CHROMOSOME REGION RECOGNITION WITH LOCAL BAND PATTERNS

chromosome type 1

lbp

(1)

chromosome type 2

lbp

(2)

chromosome type 3

lbp

(3)

chromosome type 4

lbp

(4)

chromosome type 5

lbp

(5)

Figure 7: Examples of templates used in the experiments.

(i)

(k = 1,2,. .., K

(i)

) are made. Since region ex-

traction and medial axis determination may fail for

some chromosome regions, the number K

(i)

of in-

tensity proﬁles differs depending on the chromosome

type i. In each i, the longest proﬁle P

(i)∗

= P

(i)

determined, the lengths of other proﬁles P

(i)

(k 6= l)

are extended to that of P

(i)∗

, and the average proﬁle is

made from all P

(i)

. The average proﬁle is used as a

reference band pattern of chromosome type i.

By dividing the reference band patterns into local

band patterns (lbp), templates for subregion searches

are made. To adjust the dimensions of a template for

the chromosome type i, the width W

and the total

length L

of chromosome regions are determined in

the reference image where P

(i)∗

is acquired. Figure 7

shows examples of the templates made from set A and

used in the experiments. In the experiments:

• For the chromosome types 1, 2, ..., 5, a single

template was made each in every set.

• Thresholds were set as Td = 3pixels and Te =

2pixels in estimating W

• The mean and variance of intensity in each tem-

plate were set to 100 and 50

, respectively.

5.3 Experimental Results

The following four type methods of searching subre-

gion were applied the subject images:

SRS1 without adjusting template dimensions and

without adapting subject image intensities,

SRS2 without adjusting template dimensions and

with adapting subject image intensities,

SRS3 with adjusting template dimensions and

without adapting subject image intensities,

Figure 8: Examples of the correct subregions in a subject

image.

(a)

(b)

Figure 9: Examples of the subregion search results: (a) with

SRS1, (b) with SRS4.

SRS4 (the proposed method)

with adjusting template dimensions and

with adapting subject image intensities.

To evaluate their results, precision P and recall R were

used. They are deﬁned by

P = |D∩C|/|D|, (11)

R = |D∩C|/|C|, (12)

where D is a set of detected subregions and C is a

set of correct subregions (subregions corresponding

to a template) in an subject image. |D| and |C| denote

the number of subregions in D and C, respectively. P

and R were computed for each subject image, and the

averages of them were calculated for each method.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

0.6

0.8

1.0

ll (R)

SRS1

SRS2

SRS3

SRS4 (proposed method)

0.0

0.2

0.4

0 100 200 300 400 500 600 700 800 900 1000

rec

MSE

Figure 10: The average R at different thresholds of MSE.

0.6

0.8

1.0

ll (R)

0.0

0.2

0.4

0 20 40 60 80 100 120 140 160 180 200

rec

order

SRS1

SRS2

SRS3

SRS4 (proposed method)

Figure 11: The average R at different thresholds of order.

In the experiments, detected subregions D for a

template were deﬁned as follows:

• subregions in a subject image are sorted by their

MSEs to the template in ascending order,

• if the MSE or the order of a subregion is less than

or within a speciﬁed threshold, this region is de-

cided as ‘detected.’

For each template, the correct subregions C were set

manually in every subject image (Figure 8).

Figure 9 (a) and (b) show examples of the subre-

gion search results with SRS1 and SRS4, respectively.

The numbers on the ﬁgures denote the order of each

subregion. These results were obtained by using the

same template whose correct subregions correspond

to those on Figure 8.

The averages of P and R were computed by vary-

ing the speciﬁed thresholds. Figure 10 shows the av-

erage R for all methods (SRS1, 2, 3, 4) at different

thresholds of MSE, and Figure 11 shows the average

R at different thresholds of order. Figure 12 shows P

at different R, where R was changed by varying the

threshold of MSE.

These results show that adjusting the dimensions

of templates and adapting the intensities in a subject

image improve the accuracy (precision and recall) in

0.6

0.8

1.0

ision (P)

SRS1

SRS2

SRS3

SRS4 ( d th d)

0.0

0.2

0.4

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

pre

recall (R)

SRS4

(

propose

Figure 12: The average P at different the average R.

subregion searches, especially the proposed method,

which uses both the approaches, improves the accu-

racy considerably. Consequently, it is expected that

the proposed method can achieve subregion searches

effectively.

6 CONCLUSIONS

In this paper, to improve the accuracy and ﬂexibility

in this process, we have proposed a local band pattern

based method for recognizing individual chromosome

regions in an image, and to develop an efﬁcient chro-

mosome subregion search, we also have proposed the

method for making the dimensions and intensities of

templates correspond with those of chromosome re-

gions in a subject image. By adjusting the dimen-

sions of the templates to those of chromosome regions

in the subject image and adapting the intensities in

the subject image to those of the templates, the pro-

posed method can improve the accuracy in subregion

searches.

To achieve an effective recognition of individual

chromosome regions in the subject image, we plan to

develop following methods:

• A method for determining effective subregion

search templates in each reference band pattern.

• A method for extracting and classifying a chro-

mosome regions in the subject image efﬁciently.

• A method for recognizing a complement of chro-

mosome regions in the subject image effectively.

REFERENCES

Carothers, A. and Piper, J. (1994). Computer-aided classi-

ﬁcation of human chromosomes: a review. Statistics

Computing, 4:161–171.

Graham, J. and Piper, J. (1994). Automatic karyotype anal-

ysis. Meth. Mol. Biol., 29:141–185.

CHROMOSOME REGION RECOGNITION WITH LOCAL BAND PATTERNS

Groen, F. C. A., ten Kate, T. K., Smeulders, A. W. M., and

Young, I. T. (1989). Human chromosome classiﬁca-

tion based on local band descriptors. Pattern Recognit.

Lett., 9:211–222.

Harnden, D. G. and Klinger, H. P., editors (1985).

ISCN1985: An international system for human cyto-

genetic nomenclature (1985). S. Karger AG.

Moradi, M. and Setarehdan, S. K. (2006). New features

for automatic classiﬁcation of human chromosomes:

a feasibility study. Pattern Recognit. Lett., 27:19–28.

Piper, J. and Granum, E. (1989). On fully automatic feature

measurement for banded chromosome classiﬁcation.

Cytometry, 10:242–255.

Wu, Q., Liu, Z., Chen, T., Xiong, Z., and Castleman, K. R.

(2005). Subspace-based prototyping and classiﬁcation

of chromosome images. IEEE Trans. Pattern Anal.

Machine Intell., 14:1277–1287.

ZooWeb (2003). — Karyotypes home page. http://worms.

zoology.wisc.edu/zooweb/Phelps/karyotype.html.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing