Automatic chromosome classification and chromosome abnormalities identification based on dynamic time warping

AUTOMATIC CHROMOSOME CLASSIFICATION AND CHROMOSOME
ABNORMALITIES IDENTIFICATION BASED ON DYNAMIC TIME WARPING

BENOIT LEGRAND
(B. E., Supélec, France)

A THESIS SUBMITTED
FOR THE DEGREE OF MASTER OF ENGINEERING
DEPARTMENT OF ELECTRICAL & COMPUTER ENGINEERING
NATIONAL UNIVERSITY OF SINGAPORE
2006


Acknowledgments

First and foremost, I would like to express my gratitude to my supervisor, Dr Chang
Che Sau, for his invaluable time and support during my two years of study in Singapore.

I also would like to express my deep gratitude to Dr Ong Sim Heng, for his
pertinent and useful advice.

I am grateful to Dr Nallasivam Palanisamy and Dr Neo Soek Ying of the Genome
Institute of Singapore. The chromosome images and their expertise in cytogenetics were
essential to finalize my project.

I would like to thank Mr. Seow Hung Cheng, the Power System Lab officer, for his
warm welcome and his continuous support.

This work was supported by a research scholarship from the National University of
Singapore (NUS). I am extremely thankful to NUS for the financial support.

My study in Singapore was made possible through the double-degree program
between the National University of Singapore (NUS) and Ecole Supérieure d’Electricité
(Supélec). I am particularly grateful to the international relation offices of both Supélec
and NUS for their advice before and during my stay in Singapore.

i


Table of contents

Table of contents
Acknowledgments................................................................................................................ i
Table of contents................................................................................................................. ii
Summary ............................................................................................................................ ix
List of publications related to this study ............................................................................ xi
List of Tables .................................................................................................................... xii
List of Figures .................................................................................................................. xiii
List of Symbols ................................................................................................................. xv
List of Symbols ................................................................................................................. xv
List of Abbreviations ....................................................................................................... xix
CHAPTER 1

INTRODUCTION ............................................................................................ 1

1.1

Motivations and objectives for the research ....................................................... 1

1.2

Major steps of chromosome classification.......................................................... 3

1.3

Main contributions .............................................................................................. 4

1.3.1

Normal chromosome classification............................................................. 5

1.3.2

Mapping of chromosome abnormalities ..................................................... 6

1.4

Thesis organization ............................................................................................. 7

CHAPTER 2

CHROMOSOMES AND IMAGING TECHNIQUES ................................................ 9

2.1

Chromosome structure ........................................................................................ 9

2.2

Karyotyping and chromosome banding ............................................................ 12

2.3

Chromosome abnormalities .............................................................................. 14

2.3.1

Extra or missing chromosomes................................................................. 14

2.3.2

Duplication................................................................................................ 15

ii


Table of contents

2.3.3

Deletion..................................................................................................... 15

2.3.4

Translocation............................................................................................. 16

2.3.5

Inversion ................................................................................................... 17

2.3.6

Insertion .................................................................................................... 18

2.3.7

Ring........................................................................................................... 18

2.4

Modern chromosome imaging techniques ........................................................ 19

2.4.1

Fluorescence in situ hybridization (FISH)................................................ 19

2.4.2

Comparative genomic hybridization (CGH)............................................. 20

2.4.3

Multicolor karyotyping (M-FISH, SKY and CCK).................................. 21

2.5

Conclusion ........................................................................................................ 22

CHAPTER 3
3.1

LITERATURE SURVEY ON AUTOMATIC CHROMOSOME CLASSIFICATION ...... 23

Image quality improvement .............................................................................. 24

3.1.1

Noise removal ........................................................................................... 24

3.1.2

Debris removal.......................................................................................... 24

3.1.3

Image enhancement .................................................................................. 25

3.2

Chromosome segmentation............................................................................... 26

3.2.1

Threshold .................................................................................................. 26

3.2.2

Variance .................................................................................................... 27

3.2.3

Entropy...................................................................................................... 27

3.3

Touching and overlapping chromosomes ......................................................... 27

3.3.1

Skeleton..................................................................................................... 28

3.3.2

Curvature................................................................................................... 29

3.3.3

Band analysis ............................................................................................ 29

iii


Table of contents

3.4

Feature extraction.............................................................................................. 30

3.4.1

Skeleton..................................................................................................... 30

3.4.2

Length ....................................................................................................... 31

3.4.3

Centromere................................................................................................ 31

3.4.4

Density profile .......................................................................................... 32

3.4.5

Border ....................................................................................................... 32

3.4.6

Area........................................................................................................... 33

3.4.7

Signature ................................................................................................... 33

3.5

Feature reduction .............................................................................................. 33

3.6

Classification..................................................................................................... 35

3.6.1

Neural Network......................................................................................... 36

3.6.1.1

Multi Layer Perceptron ......................................................................... 36

3.6.1.2

Probabilistic Neural Networks.............................................................. 37

3.6.1.3

Higher Order Neural Networks............................................................. 37

3.6.2

Expert system............................................................................................ 37

3.6.3

Bayesian classifier .................................................................................... 38

3.6.4

Classification in two steps ........................................................................ 38

3.6.5

Hybrid method .......................................................................................... 39

3.7

Implementation ................................................................................................. 39

3.8

Conclusion ........................................................................................................ 39

CHAPTER 4

CHROMOSOME SEGMENTATION AND FEATURE EXTRACTION ...................... 40

4.1

Noise filtering ................................................................................................... 41

4.2

Segmentation..................................................................................................... 41

iv


Table of contents

4.2.1

Threshold .................................................................................................. 42

4.2.2

Border detection........................................................................................ 43

4.2.3

Filling........................................................................................................ 43

4.2.4

Thinning.................................................................................................... 43

4.2.5

Pruning...................................................................................................... 44

4.2.6

Skeleton extension .................................................................................... 44

4.3

Feature extraction.............................................................................................. 44

4.3.1

Length ....................................................................................................... 45

4.3.2

Density profile .......................................................................................... 46

4.3.3

Centromere location.................................................................................. 47

4.4

Feature normalization: ...................................................................................... 48

4.4.1

Length normalization ................................................................................ 48

4.4.2

Density profile .......................................................................................... 48

4.4.3

Centromeric index..................................................................................... 50

4.5

Conclusion ........................................................................................................ 51

CHAPTER 5

NORMAL HUMAN CHROMOSOME CLASSIFICATION ..................................... 52

5.1

Introduction....................................................................................................... 52

5.2

Data base........................................................................................................... 53

5.3

DTW-based classifier........................................................................................ 55

5.3.1

Motivation for using the DTW algorithm................................................. 55

5.3.2

DTW-based chromosome classifier.......................................................... 58

5.3.2.1

Reference chromosomes ....................................................................... 60

5.3.2.2

Length feature ....................................................................................... 62

v


Table of contents

5.3.2.3
5.3.3

Dimension of the density profile........................................................... 63

5.4

Classification results ................................................................................. 64
Comparison with the Bayesian classifier.......................................................... 66

5.4.1

Bayesian classifier .................................................................................... 66

5.4.2

Classification performances...................................................................... 67

5.4.3

Size of the training set............................................................................... 68

5.5

Conclusion ........................................................................................................ 70

CHAPTER 6

ABNORMAL CHROMOSOME ANALYSIS ........................................................ 71

6.1

Introduction....................................................................................................... 71

6.2

Image acquisition .............................................................................................. 73

6.2.1

G-Banding images .................................................................................... 74

6.2.2

SKY Images .............................................................................................. 74

6.3

Reference density profiles................................................................................. 75

6.4

Translocation analysis procedure...................................................................... 76

6.4.1

Overall principle ....................................................................................... 76

6.4.2

Scale ratio.................................................................................................. 81

6.4.3

Chromosome sections comparison ........................................................... 84

6.4.4

Recombination point identification........................................................... 88

6.5

Performances and results .................................................................................. 89

6.6

Conclusions....................................................................................................... 95

CHAPTER 7

CONCLUSION.............................................................................................. 96

Bibliography ..................................................................................................................... 98

APPENDIX A

IMAGE PROCESSING ALGORITHMS ............................................................ 104

vi


Table of contents

A.1

Median filter.................................................................................................... 104

A.2

Border follower algorithm .............................................................................. 105

A.3

Filling algorithm ............................................................................................. 107

A.4

Thinning algorithm ......................................................................................... 108

APPENDIX B

DYNAMIC TIME WARPING ALGORITHM ..................................................... 112

APPENDIX C

BAYESIAN CLASSIFIER ............................................................................. 116

C.1

Theory ............................................................................................................. 116

C.2

Feature dimension reduction........................................................................... 118

APPENDIX D

MAPPING OF CHROMOSOME ABNORMALITIES........................................... 120

D.1

Cell line HCT116............................................................................................ 121

D.2

Cell line HEPG2 ............................................................................................. 123

D.3

Cell line HEP3B.............................................................................................. 125

APPENDIX E

IDEOGRAMS AND REFERENCE DENSITY PROFILES ..................................... 126

E.1

Chromosome 1 ................................................................................................ 126

E.2

Chromosome 5 ................................................................................................ 127

E.3

Chromosome 6 ................................................................................................ 128

E.4

Chromosome 7 ................................................................................................ 129

E.5

Chromosome 8 ................................................................................................ 130

E.6

Chromosome 10 .............................................................................................. 131

E.7

Chromosome 11 .............................................................................................. 132

E.8

Chromosome 16 .............................................................................................. 133

E.9

Chromosome 17 .............................................................................................. 134

E.10

Chromosome 18 .............................................................................................. 135

vii


Table of contents

E.11

Chromosome 21 .............................................................................................. 136

viii


Summary

Summary
Studies on chromosomes are essential to understand chromosome aberrations and
their corresponding diseases. Nowadays, human chromosome analysis is commonly used
to identify genetic defects in prenatal screening and in cancer pathology research.

Manual chromosome analysis is a slow and laborious operation that requires
qualified personnel. Modern imaging techniques and image analysis software have
recently greatly facilitated the study on chromosome. However, there is a lack of
software able to deal with abnormal chromosomes and to combine information from
different imaging methods. This study proposes a procedure to automate the chromosome
banding analysis. The proposed system is able to automatically map some common
chromosome abnormalities by combining a banding analysis with the information
provided by the spectral karyotyping imaging techniques. The dynamic time warping
(DTW) algorithm, traditionally used in speech recognition applications, has been adapted
to the identification of chromosome banding patterns in order to overcome the problems
due to the non-rigid nature of chromosomes.

At first, a pattern classifier has been developed to perform the classification of
normal human chromosomes. This classifier is based on the dynamic time warping
algorithm in order to be able to compare unknown banding patterns with some precomputed references. This classification method has the main advantage to require only a
few training samples in comparison with the traditional chromosome classifiers based on
Neural networks or Bayesian classifiers. This property is due to the transfer of the
ix


Summary

knowledge that chromosomes can have different elongations from the training set to the
classifier itself. For the same classification accuracy, the DTW-based classifier achieves a
large reduction of 88% of the number of training samples in comparison with the
Bayesian classifier. This performance is particularly useful for dealing with species that
are seldom studied or to classify some abnormal chromosomes with only a few training
samples.

The second part of the study is dedicated to the development of an automatic

system to map abnormal chromosomes. Several modern cytogenetics imaging techniques
facilitate the study of chromosome defects, but are individually unable to completely map
chromosome aberrations. Comparative genomic hybridization (CGH) is an imaging
method that has the ability to reveal chromosome amplifications and deletions, but not to
determine the amplified or deleted sections. Multiplex fluorescence in situ hybridization
(M-FISH) and spectral karyotyping (SKY) detect inter-chromosomal rearrangements, but
are not able to discover intra-chromosomal rearrangement and to identify the exchanged
chromosome sections. For all these reasons, the traditional banding analysis method is
essential to obtain a complete identification of chromosome aberrations. Nowadays,
banding analysis and combination of information from different imaging systems are
done manually by cytogenetics experts. This study proposes a procedure to automate the
banding analysis operation by using the information given by the spectral karyotyping
images. The proposed system has been successfully applied to the mapping of
chromosome translocations, duplications and deletions in solid tumor cells. This new tool
is useful to better understand chromosome disorders.

x


List of publications related to this study

List of publications related to this study
1. B. Legrand, C.S. Chang, S.H. Ong, S.Y. Neo and N. Palanisamy, “Chromosome
classification using dynamic time warping”, submitted to Pattern and Recognition
Letters.

2. B. Legrand, C.S. Chang, S.H. Ong, S.Y. Neo and N. Palanisamy, “Mapping of
chromosome abnormalities by automated banding analysis”, submitted to IEEE
Transactions on Biomedical Engineering.

xi


List of Tables

List of Tables
Table 2.1: Description of the seven Denver groups

12

Table 5.1: Normal human chromosome data base

54

Table 5.2: Confusion matrix

65

Table 5.3: Classification accuracy (DTW-based classifier)

65

Table 5.4: Number of samples in the training and test set

67

Table 5.5: Relation between classification accuracy and features

68

Table 5.6: Classification accuracy (Bayesian classifier)

68

Table 6.1: Cell lines

73

Table 6.2: Chromosome abnormalities

89

Table 6.3: Translocation t(5;7) in the cell line HCT116

93

xii


List of Figures

List of Figures
Figure 1.1: Major steps of chromosome identification

4

Figure 2.1: Normal human female G-banded karyotype

10

Figure 2.2: Silhouette of a chromosome

10

Figure 2.3: Mitotic cell division

11

Figure 2.4: Example of metacentric, submetacentric and acrocentric chromosomes

12

Figure 2.5: Ideograms of the normal human chromosome 16 in different resolutions [50]

14

Figure 2.6: Example of chromosome abnormalities (duplication, deletion, translocation)

17

Figure 2.7: Example of chromosome abnormalities: (inversion, insertion, ring)

19

Figure 2.8: Example of FISH image

20

Figure 2.9: Example of SKY image (cell line HEP3B - liver tumor)

22

Figure 3.1: Major steps of automatic chromosome classification

23

Figure 3.2: Debris on a metaphase image [36]

25

Figure 3.3: Example of touching and overlapping chromosomes on a metaphase image

28

Figure 3.4: Skeleton of overlapping chromosome

29

Figure 3.5: Possible cut points on overlapped chromosomes [13]

29

Figure 4.1: First step: Image quality improvement

41

Figure 4.2: Second step: Chromosome segmentation

41

Figure 4.3: Chromosome segmentation

42

Figure 4.4: Third step: Feature computation

45

Figure 4.5: Chromosome length computation

45

Figure 4.6: Skeleton between the two sister chromatids

46

Figure 4.7: Section of a chromosome

47

Figure 4.8: Normalized density profile of a normal human G-banded chromosome 1

50

Figure 4.9: Centromeric index

51

Figure 5.1: Normalized density profiles of two normal chromosomes 1

56

Figure 5.2: Alignment found by the DTW algorithm for the density profiles of Figure 5.1

57

xiii


List of Figures

Figure 5.3: Normal chromosome DTW-based classifier

58

Figure 5.4: Overall principle of the DTW-based chromosome classifier

59

Figure 5.5: Classification accuracy versus the number of training samples

61

Figure 5.6: Classification accuracy versus the number of pre-selected reference chromosomes

63

Figure 5.7: Classification accuracy versus the dimension of the density profile

64

Figure 5.8: Density profiles of chromosomes 19 and 21

66

Figure 5.9: Classification accuracy versus the number of training samples (Bayesian classifier)

69

Figure 6.1: Necessity of the banding analysis (translocation t(1,11) in the cell line HEP3B)

72

Figure 6.2: Overall flowchart of the proposed algorithm

78

Figure 6.3: Find the normal section that best match an abnormal section

80

Figure 6.4: Comparison of an abnormal section with a reference chromosome

86

Figure 6.5: Similarity between the translocation and sections of a reference chromosome

87

Figure 6.6: Example of translocation t(5;7) from the cell line HCT116

90

Figure 6.7: Section of reference chromosome 5

91

Figure 6.8: Section of reference chromosome 7

92

Figure 6.9: Complete mapping of translocation t(5;7) from the cell line HCT116

94

Figure A.1: Border follower algorithm

105

Figure A.2: Flow chart of the border follower algorithm

106

Figure A.3 Flow chart of the filling algorithm

108

Figure A.4: Flow chart of the thinning algorithm

109

Figure A.5: Flow chart of the first step of the thinning algorithm

111

Figure B.1: Unknown and reference density profiles

112

Figure B.2: Local cost matrix

113

Figure B.3: DTW path

114

Figure B.4: DTW step

115

xiv


List of Symbols

List of Symbols
Chapter 4: Chromosome segmentation and feature extraction
pi

Gray level value of the ith point of the raw density profile

pi′

Gray level value of the ith point of the re-sampled density profile

pi′′

Normalized value of the ith point of the re-sampled density profile

M

Number of points in the raw density profiles

N

Number of points in the re-sampled density profiles

Chapter 5: Normal human chromosome classification
Li , p

Normalized length of the ith training sample of chromosome p

Lp

Average of the normalized lengths of chromosome p

T

Number of training samples used to create each reference template

Chapter 6: Abnormal chromosome analysis
l

Length of a normal chromosome in a reference karyotype

lˆ

Length of a normal chromosome in the abnormal karyotype

li

Length of the reference chromosome i

lˆi

Length of a normal chromosome i in the abnormal karyotype

Li ,k

Normalized length of the chromosome i in the normal karyotype k

Li

Average of the normalized lengths of chromosome i

K

Total number of normal karyotypes used to compute the average of the
normalized chromosome lengths

xv


List of Symbols

CLR (i , j )

Chromosome length ratio: Average of the normalized lengths of
chromosome i divided by the average of the normalized lengths of
chromosome j

Γ

Scale ratio: Length of a normal chromosome in the reference karyotype
divided by the length of the same normal chromosome in the abnormal
karyotype. Ratio between the scales of the reference and abnormal
karyotypes

Γi

Scale ratio for chromosome i: Ratio between the scale of the reference
chromosome i and the scale of the abnormal karyotype

R

Total number of chromosome used as length patterns to compute the
improved scale ratio

sˆ

Length of the abnormal section that are analyzed

si

Length of the sections of the reference chromosome i that are compared
with the abnormal section

n

Abscissa of the beginning of the reference section along the reference
chromosome

d

Direction of the comparison of the abnormal section with the reference
sections

A

Density profile sequence of the abnormal section

R id (n )

Density profile sequence of the section of the reference chromosome i
that starts at the abscissa n, in the direction d

DTWCost id (n)

Result of the DTW comparison, in the direction d, between the

abnormal section and the section of reference chromosome i that starts

xvi


List of Symbols

at the abscissa n
ni (d )

Abscissa of the beginning of the best reference section along the
reference chromosome i, for each comparison direction d

dˆi

Best comparison direction

nˆi

Best abscissa of the beginning of the reference section along the
reference chromosome i

r

Abscissa of the recombination point along the abnormal chromosome

nˆi ,r

Best abscissa of the beginning of the reference section along the
reference chromosome i, for the abscissa r of the recombination point

RPCost (r )

Total DTW cost corresponding to the abscissa r of the recombination
point

rˆ

Abscissa of the best recombination point

Appendix B Dynamic time warping algorithm
Unk

Sequence of the unknown density profile

Ref

Sequence of the reference density profile

ui

Value of the ith of the unknown density profile

ri

Value of the ith of the reference density profile

n

Number of point in the unknown density profile

m

Number of point in the reference density profile

ci , j

Element of the local cost matrix

G path

Global cost associated to the DTW path

N path

Length of the DTW path

xvii


List of Symbols

Ci , j

Cumulative cost in the local DTW matrix

Appendix C Bayesian classifier
x

Feature vector

ωi

Class ω i : the unknown sample is a chromosome i

P(ωi )

Prior probability of class ω i

p( x | ω i )

Probability density function

P(ω i | x)

Conditional probability of class ω i given a feature vector x

Σi)

d by d covariance matrix of the class ω i

N (i )

Number of training samples of the class ω i

J

Criterion for class separability

x(i ) k

kth training feature vector for chromosome i

SWithin

Within-Class Scatter Matrix

S Between

Between-Class Scatter Matrix

μ (i )

Mean of the feature vectors of the class ω i

μ(0)

Expected mean vector

d

Dimension of the feature vector before the dimension reduction process

nd

Reduced dimension of the feature vectors

xd

d -components column feature vector

xnd

nd -components column feature vector

xviii


List of Abbreviations

List of Abbreviations
ATCC

American type culture collection

BP

Break point

CGH

Comparative genomic hybridization

CCK

Color changing karyotype

CI

Centromeric index

DNA

Deoxyribonucleic acid

DTW

Dynamic time warping

DP

Density profile

FIR

Finite Impulse Response (filter)

FISH

Fluorescence in situ hybridization

GIS

Genome institute of Singapore

ISCN

International system for human cytogenetic nomenclature

MAT

Medial axis transform

MLP

Multi layer perceptron

M-FISH

Multiplex fluorescence in situ hybridization

NUS

National university of Singapore

RP

Recombination point in an abnormal chromosome

SKY

Spectral karyotyping

xix


Chapter 1 Introduction

Chapter 1

INTRODUCTION

In this introductory chapter, the motivation for the work done is presented. Then,
major contributions, and structure of the thesis are summarized.

1.1

Motivations and objectives for the research
Chromosomes are located in the nuclei of eukaryote cells and contain the DNA

double helices. They carry the genetic instructions for making living organisms [1].
Genetic defects that affect the chromosome structures are numerous, including
chromosome rearrangements, duplications and deletions [6]. Identification of such
chromosome aberrations has enormous impact on clinical diagnosis, medicine
development and basic research [56]. Nowadays, human chromosome analysis is
particularly used to identify genetic disorders in prenatal screening and in cancer
pathology research [55].

Chromosome banding analysis is the most commonly used method to check the
integrity of chromosomes. This method was discovered in 1969 and allows the
identification of the chromosomes by assigning a unique banding pattern on each
chromosome [3]. Banding analysis is routinely used in chromosome classification and
abnormalities detection. However, manual banding analysis is a slow and laborious
operation, and complex chromosome rearrangements are hard to identify. As a
consequence, there are medical and economic motivations to automate this process [2].

1


Chapter 1 Introduction

Automatic image analysis software has recently been developed to classify
chromosomes by using features including the length and the banding patterns [2] [9] [22]
[24] [27] [28] [31] [33] [34] [35] [36] [37] [39] [42] [43]. They are commonly used in
laboratories to classify normal human chromosomes. However, these systems are too task
specific and limited to the identification of normal chromosomes of a few species [3].
Most of these systems are based on neural networks or Bayesian classifiers because of
their ability to learn expert knowledge through a training process [35]. These methods
require the creation of sizeable training sets for each species.

Discovered in the nineties, new cytogenetic imaging techniques have greatly
improved the study of chromosome defects [55] [56]. These methods are based on in-situ
hybridization of the tumor DNA with colored DNA probes. They facilitate the detection
of complex chromosome abnormalities. However, these methods often require to be
combined with a traditional banding analysis to completely identify the various
chromosome segments [3]. Comparative genomic hybridization (CGH) has the ability to
reveal chromosome amplifications and deletions, but not to determine the amplified or
deleted sections. Multiplex fluorescence in situ hybridization (M-FISH) and spectral
karyotyping (SKY) detect inter-chromosomal rearrangements, but are not able to
discover intra-chromosomal rearrangement and to identify the exchanged chromosome
parts. As a consequence, chromosome banding analysis is essential for mapping
chromosome aberrations by identifying the involved abnormal sections. This issue is
currently not addressed by banding analysis imaging software that is mostly unable to

2


Chapter 1 Introduction

deal with abnormal chromosomes and to combine information from different imaging

techniques.

Chromosome aberrations are numerous and it is difficult to create large training set
for each abnormality. As a consequence, the ability to recognize chromosome sections
with only a few training samples is essential. This study proposes to use the dynamic time
warping (DTW) algorithm to compare unknown chromosome sections with some
reference banding patterns. DTW systems are mainly used in speech recognition
applications [4] to compare an unknown word with a stored reference word. The main
advantages of this method are that they require a small training set and are able to
recognize chromosome with different elongations. Based on this DTW algorithm, the
proposed study has developed a classifier for normal human chromosomes, and a banding
analysis procedure to map chromosome abnormalities.

1.2

Major steps of chromosome classification
The process of chromosome classification can be divided in 4 major steps (Figure

1.1). The input of the system is a gray level digital image of a banded chromosome. The
first 3 steps are image processing operations that are necessary to extract some features
from the raw chromosome image. The quality of the initial chromosome image and the
efficiency of the whole feature extraction process are essential to provide good features to
the chromosome identification step. This fourth step uses the previously computed
features to identify the unknown chromosome. If the unknown chromosome is known to
be normal, then the chromosome identification box is a classifier, and its output is one of

3


Chapter 1 Introduction

the 24 possible chromosome indexes (1,..., 22, X, Y). The classifier use a training set to
perform this classification. On the contrary, if the unknown chromosome is possibly
abnormal, then additional information about the abnormality is provided to the
chromosome identification box, and its output is a mapping of the abnormal
chromosome. This additional information is obtained from a modern color imaging
system, which is able to detect the existence of some abnormalities, but requires an
analysis of the chromosome bands to perform a complete identification.

Normal
training
samples

1

Unknown
banded
chromosome
image

Image quality
improvement

2
Chromosome
segmentation

Information
about a possible
abnormality

3
Feature
extraction

4
Chromosome
identification

Identified
chromosome
segments

Figure 1.1: Major steps of chromosome identification

1.3

Main contributions
Contributions have been made to the steps 2 and 3 of Figure 1.1 in order to perform

an initial efficient feature extraction process with the banded chromosome images
provided by the Genome Institute of Singapore. Then, two major contributions are
proposed to the fourth step in order to identify normal and abnormal chromosomes.
Firstly, a normal chromosome classifier has been developed in order to strongly reduce
the number of necessary training samples. Secondly, a new procedure is proposed to map
4


Chapter 1 Introduction

abnormal chromosomes by combining the classical banding analysis with information
obtained with the Spectral Karyotyping (SKY) system.

1.3.1

Normal chromosome classification

Chromosomes are non-rigid objects, and their banding patterns are difficult to
compare because of their different elongations. A section of a chromosome can be more
contracted or on the contrary more elongated than the rest of the same chromosome. As a
consequence, the comparison of chromosome banding patterns with some pre-computed
reference patterns is not straightforward because the bands do not line up correctly. This
study proposes to solve the problem by using the dynamic time warping (DTW)
algorithm in order to find an optimum match by stretching the banding patterns. DTW
systems are well known for their application in automatic speech recognition systems [4]
to compare an unknown word with a reference word that can be pronounced with
different speeds. DTW algorithm can effectively be applied to the problem of normal
chromosome classification by comparing unknown samples with reference patterns. The
main advantage of this method is that it requires a much smaller training set in
comparison with the conventional methods based on Bayesian classifiers or Neural
Networks. This is because they have to learn from the training samples all the possible
elongations that can appear for each chromosome. On the contrary, the reference patterns
necessary to the DTW-based classifier can be computed from a few training samples. In
other words, the DTW algorithm transfers the knowledge that chromosomes are non-rigid
objects from the training set to the classifier itself. In this study, a classifier based on
dynamic time warping (DTW) has been developed to perform the classification of human

5