M-FISH KARYOTYPING - A NEW APPROACH BASED ON WATERSHED TRANSFORM

International Journal of Computer Science, Engineering and Information Technology (IJCSEIT), Vol.2, No.2, April 2012
DOI : 10.5121/ijcseit.2012.2210 105
M-FISH KARYOTYPING - A NEW APPROACH BASED
ON WATERSHED TRANSFORM
Sreejini K S1
, Lijiya A2
and V K Govindan3
Department of Computer Science and Engineering, National Institute of Technology,
Calicut, Kerala, India
1
sreejini_k_s@yahoo.com, 2
lijiya@nitc.ac.in, 3
vkg@nitc.ac.in
ABSTRACT
Karyotyping is a process in which chromosomes in a dividing cell are properly stained, identified and
displayed in a standard format, which helps geneticist to study and diagnose genetic factors behind various
genetic diseases and for studying cancer. M-FISH (Multiplex Fluorescent In-Situ Hybridization) provides
color karyotyping. In this paper, an automated method for M-FISH chromosome segmentation based on
watershed transform followed by naive Bayes classification of each region using the features, mean and
standard deviation, is presented. Also, a post processing step is added to re-classify the small chromosome
segments to the neighboring larger segment for reducing the chances of misclassification. The approach
provided improved accuracy when compared to the pixel-by-pixel approach. The approach was tested on
40 images from the dataset and achieved an accuracy of 84.21 %.
KEYWORDS
Bayes classifier, Chromosome image segmentation, Karyotyping, M-FISH, Watershed transform.
1. INTRODUCTION
In clinical and research cytogenetic studies, automated computerized systems for human
chromosome analysis are very essential since a small deviation from the usual number of
chromosomes will result in physical abnormalities. Chromosomes are structures located in nuclei
of eukaryote cells that carry all the genetic instructions for making living organisms. Normal
human metaphase spread contains 46 chromosomes, 22 pairs of autosomes and sex chromosomes
(XY: Male, XX: Female). Chromosomes are present in every cell except red blood cells.
Chromosome analysis is done on dividing cells in their metaphase stage (different phases of cell
division: metaphase, anaphase, and telophase). During metaphase, chromosome can be stained to
become visible and can be imaged by a microscope. Cells used for chromosome analysis are
usually taken from amniotic fluid or from blood samples.
Karyotype is the tabular representation of human chromosomes in a cell. In this representation,
the chromosomes are ordered by length from largest (chromosome 1) to smallest (chromosome 22
in humans), followed by sex chromosomes. Karyotypes are very useful for accurately diagnosing
the genetic factors behind various diseases. Manual karyotyping is time-consuming, expensive
and need well trained personnel. During the early period of chromosome analysis, researchers
used grayscale images and features such as size, shape, centromere position and banding pattern
for classification.

106
Since 1996, a staining method called M-FISH introduced by Speicher et al. [1] produces color
images. This simplifies the karyotyping and detection of subtle chromosome aberrations. Images
are captured with a fluorescent microscope with multiple optical filters. Combinatorial labelling
of 5 fluorophores is used to assign a specific fluor combination to each of the chromosomes, so
that each chromosome type can be visualized in a unique color. A sixth fluorophore, DAPI (4 in,
6-diamidino-2-phenylindole), is counterstained to all chromosomes. Figure 1 shows the five
channel M-FISH image data of a VYSIS probe.
Figure 1. Five channel M-FISH image data. (a) Aqua fluor. (b) Red fluor
(c) Far red fluor. (d) Green fluor. (e) Gold fluor. (f) DAPI image.
This paper proposes an approach for M-FISH karyotyping based on naive Bayes classification of
M-FISH image segments obtained by applying watershed segmentation. Here, classification is
based on the features, mean and standard deviation, followed by a post-processing to reduce the
misclassification.
The paper is organized as follows: Section 2 discusses some of the major existing work in the
literature. Image segmentation and classification processes are given in Section 3. The
comparative results obtained on standard database for the proposed approach and existing
approaches are presented in Section 4, and the Section 5 concludes the paper.
2. LITERATURE SURVEY
Active research on karyotyping started since when the number of chromosome in human is found
to be 46 in 1956. There are already a number of attempts proposed by various researchers to
automate the process of karyotyping. We briefly review some of the major such works in this
section:
The first M-FISH system developed by Speicher et al. [1] achieved semi-automated analysis of
M-FISH image. In this approach, a mask is created to segment the chromosome from DAPI
channel, and a threshold is applied to each pixel in the mask in order to detect the presence or
absence of fluor in that pixel. The pixels are then classified by comparing its response with fluor

107
labelling table. This unsupervised classification method, though simple requires some manual
corrections of the segmentation map.
Presently used classification methods are based either on pixel-by-pixel or on region based
classification algorithms. Pixel-by-pixel methods either classify each pixel of the M-FISH image
or create a binary mask of the DAPI image using edge detection algorithms, and then classify
each pixel of the mask. In region based methods, the regions obtained by decomposing the image
are classified.
Automatic pixel-by-pixel classification approach proposed by Sampat et al. [2] modelled the
karyotyping as a 25 class 6 feature pattern recognition problem and classified by using Bayes
classifier. The 25 classes are the 24 chromosome types and the background, and the 6 features
used are the gray scale values from 6 color channels. The classifier was trained, and tested only
on a small set of non-overlapping images.
Another work of Sampat et al. [3] proposes supervised parametric and nonparametric
classification techniques for pixel-by-pixel classification of M-FISH images. In supervised
parametric technique, they modelled the problem as a 6-feature 25-class maximum likelihood
pattern recognition task, and in the supervised non-parametric approach, they employed nearest
neighbor and the k-nearest neighbor methods. Non-parametric approach performed better than the
parametric approach. The highest classification accuracy was obtained with the k-nearest
neighbor method and k = 7 is an optimal value for this classification task. The approaches do not
handle overlapping images and they used only a small number of test images.
Unsupervised classification method based on fuzzy logic classification and a prior adjusted
reclassification that corrects misclassifications is discussed in [4]. First, the separation of
foreground and background is carried out by majority voting among k-means clustering, adaptive
thresholding, LoG edge detection, and global thresholding methods. It requires spectral
information, obtained from color table and then classifies the pixels by using fuzzy logic classifier
and a prior adjusted reclassification was performed by adjusting the prior for each chromosome,
so most likely class for each chromosome was found. It does not require training. High average
accuracy is reported, however only a small number of test images were used.
Various pre-processing methods such as image registration, dimension reduction and background
flattening are discussed in [5, 6]. Color compensation techniques are discussed in [7]. The
authors report that these techniques are useful for improving the accuracy of karyotyping.
Mohammed et al. [8] presented an automated method for segmentation and classification of
multispectral chromosome images. They used adaptive thresholding and valley searching for
background cancellation. Discrete Wavelet Transform (DWT) is used to extract suitable number
of features. The approximate normalised DWT coefficients are used to reduce the size of the
image to 325 X 260 pixels from 645 X 517 pixels. Bayes decision theory is used to classify each
pixel in the normalized approximation image. After classification, expands the size of image to its
original, by adding zeros between each two neighboured pixels. Majority filtering is used for
removing the noise introduced during the expansion process. Each chromosome is segmented by
collecting all the pixels belonging to this chromosome. The overlapping problem can be solved by
use of the medial axis transform. High classification rate was reported.
Use of Gaussian mixture model (GMM) classifier for M-FISH images classification is presented
in [9]. They modelled karyotyping as 26 class 6 feature pixel-by-pixel classification problem. The
26 classes are the 24 types of chromosomes, the background and the chromosome’s overlap; the 6
features are the brightness of dyes at each pixel in six color channels. The overall classification

108
accuracy achieved is reported as 89.18% and is found to be better than pixel-by-pixel method, but
they used only a small number of images.
A region based watershed segmentation method applied to DAPI channel for multispectral
chromosome image classification is presented in [10]. In this, marker controlled watershed
transform is used to control over segmentation. A binary mask of the DAPI channel is computed
in order to further reduce unwanted areas. Finally, a vector containing 5 features, each feature
representing the average intensity value of each channel, is computed from each segmented area,
and the vectors are classified using Bayes classifier. Good overall accuracy is reported. However,
only a small number of non-overlapping testing images were used. This work was further
extended with multichannel in [11]. They used gradient computed from all the channels.
Classification is performed using region based Bayes classifier and the neighboring regions are
then merged. This makes the detection of unhybridized regions simpler. Good overall accuracy is
reported.
Support Vector Machines (SVM) classifier with multichannel watershed transform to perform M-
FISH karyotyping was described in [12]. They constructed RB-SVM by using radial basis
function as the kernel function. The method tested on images from normal cells and reported
10.16% increase in classification accuracy than Bayesian classification.
The work by Wang [13] deals with Fuzzy c-means clustering algorithm (FCM) based
classification of M-FISH images. This uses 24 different cluster centers, which are formed from 24
classes of chromosomes and a pixel is assigned to each individual cluster according to its nearest
distance to the center. Finding a cluster center is equivalent to minimizing the dissimilarity
function; here, Euclidian distance is used as dissimilarity measure. The advantage of FCM is that
it can locate centers more accurately because here the membership values are from 0 to 1. It
works better than k-means clustering and Bayes classifier. Use of image normalization techniques
such as image registration, dimension reduction and background subtraction are also used,
leading to improvements in accuracy.
Later, Cao and Wang [14] presented Segmentation of M-FISH images for improved classification
of chromosomes with an adaptive Fuzzy C-Means clustering. Adaptive FCM was done by
incorporating a gain field which models and corrects intensity homogeneity and also regulates
center of each intensity cluster. Intensity homogeneity is mainly caused by the image acquirement
and uneven hybridization. It provides lowest segmentation and classification error and is better
than FCM and AFCM.
Overlapping and touching chromosomes are still a problem in pixel-by-pixel classification. Many
researchers have attempted to resolve this issue. Some of the important work in this category is
[15, 16, 17]. In [15], minimum entropy is used as the main segmentation criterion to decompose
overlapping and touching chromosome images. It also uses multi-spectral information in
chromosome images. However, the computational time and complexity is very high;
performance is very sensitive to its parameters and the approach is tested only on small number of
images.
Extension of the above minimum entropy algorithm [15] is discussed in [16] which removes the
pixel classification requirement. It works by estimating entropy from the raw data using
differential entropy estimation technique, i.e., they used nearest neighbour estimation technique
rather than calculating entropy from the classified pixels. The approach leads to computational
complexity lower than that of minimum entropy approach. Still, the computational time required
is unacceptably high, performance is very sensitive to its parameters and the algorithm is tested
only on few images.

109
Another approach to resolve overlapping and touching of chromosomes is presented in [17]. The
approach utilizes the geometry of a cluster, pixel classification results and chromosome sizes.
First, the chromosomes are segmented from the background by majority voting among k-means
clustering, adaptive thresholding, LoG edge detection, and global thresholding methods and then
chromosome pixels are classified by using fuzzy logic classifier. A group of connected pixels is
defined as cluster. Three sets of basic elements of cluster are cross shape cluster, T shape cluster
and I shape cluster. For a given cluster, landmark on the boundary and skeleton are computed and
then the cluster is decomposed in to multiple hypotheses, and the likelihood of each hypothesis is
computed based on pixel classification results and chromosome size. Most likelihood hypothesis
is chosen as the correct decomposition of that cluster. Good results are reported.
From the above review work on the various approaches suggested by the researchers, one can
conclude that region based classification approaches are superior to pixel by pixel approach in
terms of accuracy and computational time. Many researchers have tested their approaches only on
small sets of selected images. For practical purposes, the systems must be capable of providing
high accuracy on large data set. So, there is a need for research to improve the performance
further on large data sets so that such automated systems for karyotyping can be acceptable for
commercial purposes.
3. METHODS
The present work employs basically two major processing steps, segmentation and classification.
Marker controlled watershed transform is used for segmentation and region based Bayes
classification is used for classification.
3.1. CHROMOSOME SEGMENTATION
The separation of each chromosome from the metaphase image is the major operation carried out
in this stage. Basic steps involved are: removal of cells from the DAPI images, gradient
computation and minima selection, computation of watershed transformation and binary mask
creation. These are briefly presented in the following subsections:
3.1.1. REMOVAL OF CELLS FROM THE DAPI IMAGE
Original DAPI chromosome image contains nuclei and debris along with chromosome. We must
remove them based on the size and circularity before segmentation. Figure 2 shows the image
before and after blob removal.
Figure 2. (a) Before blob removal (b) After blob removal

110
3.1.2. GRADIENT COMPUTATION AND MINIMA SELECTION
Gradient magnitude of the DAPI channel image after cell removal is computed. Sobel operators
are used. Since watershed algorithm produces over segmentation, can be control by reduce the
number of allowable minima in the gray scale.
3.1.3. COMPUTATION OF WATERSHED TRANSFORM
Watershed transform of the resulting image is computed which results in tessellation of the image
in to different regions. The watershed transform has the advantage that the lines produced are
always form closed and connected regions and these lines always correspond to obvious contours
of objects which appear in image.
A gray scale image can be considered as a topographic surface, where height of each point is
related to its gray level. If we punch a hole in each local minimum and immerse this surface in
water, the regions in the image will start filling up with water. Immersion will starts from the
points of minimum gray value. When water level in two or more adjacent basins will start
merging, dams are built in order to prevent this merging. The flooding process will continue up to
the stage at which only the top of dam is visible above the water line [18]. Watersheds are the
lines dividing two catchment basins, each basins corresponds to each local minimum. Figure 3
shows the watershed lines superimposed on DAPI channel.
Figure 3. Watershed lines superimposed on DAPI channel
3.1.4. BINARY MASK CREATION
Binary mask is created from DAPI image after cell removal and superimposing watershed regions
on it, in order to avoid the segmentation errors present due to unhybridization. Binary mask is
created by Otsu’s thresholding method [19]. Basic operation is, logical AND operation between
the watershed lines and blob removed DAPI image.
3.2. FEATURE EXTRACTION AND CLASSIFICATION
3.2.1. FEATURE EXTRACTION
This stage extracts the features used for classification. Here, mean and standard deviation of each
segmented area are the features used. Then the intensities of the pixels belonging to that region
are replaced with mean intensity of that region for each segmented area.

111
3.2.2. CLASSIFICATION
The segments are classified using naive Bayes classifier. A naive Bayes classifier is a simple
probabilistic classifier based on Bayes theorem with strong (naive) independence assumptions.
Our goal is to classify the 46 chromosomes in to 24 chromosomes type (C = 24).
Let x ∈ Rd
denotes the feature vector computed from each segmented area; d = 5 X 2 = 10. Here,
classification is done by using mean with standard deviation of the image under test. Let P(ci)
denote the probability that a feature vector belongs to class ci,, where i varies from 1 to 24, and is
called prior probability. Let p (x | ci) denotes the class conditional probability distribution
function for a feature vector x given that x belongs to class ci and P (ci | x) be the posterior
probability that the feature vector x belongs to class ci, given the feature vector x .
By using Bayes theorem,
Computed prior class probabilities from training samples are,
The general multivariate Gaussian density function [20] in d dimension is given by
where x is the d-dimensional feature vector from five channels and µi is the mean vector of each
class ci , ∑i is the d x d covariance matrix of the class ci , and |∑i | and ∑i
-1
are the determinant
and inverse. Also (x – µi)t
denotes the transpose of (x – µi).
For each class, we need to calculate P(ci | x), the class to which a feature x belongs, is decided by
Bayes decision rule.
3.2.3. NEIGHBOR REGION MERGING
In this stage, for each region all the neighboring regions that share the same class are connected
inorder to get meaningful class map. If regions are adjacent then those regions are connected in
Region Adjacency Graph (RAG) and have a common boundary. Figure 4.a shows the original
classmap and classmap obtained after neighbour region merging is shown Figure 4.b. Here each
type of chromosome is colored with different color.

112
Figure4. (a) Before merging (b) After merging
3.2.4. POST-PROCESSING
It is observed that small segments are usually misclassified. To overcome this, small segments are
reclassified to the most likely class of one of its neighbours by Bayes theorem, so that it becomes
the same class as one of these neighbours.
Figure5. (a) Before post-processing (b) After post-processing
Figure 5.a shows the results before post-processing. Circle denotes the misclassification that was
corrected by the use of post-processing method as shown Figure 4.b.

113
4. RESULTS
4.1 M-FISH CHROMOSOME IMAGE DATABASE
Dataset [21] consist of 200 Multispectral images of size 517 X 645 pixels. 17 images are marked
as extreme (EX), that are “difficult to karyotype”. ASI, PSI, Vysis are the probes used. Each M-
FISH image set consist of 5 monospectral images recorded at different wavelengths, DAPI and its
“ground truth” image according to ISCN (International System for Human Cytogenetic
Nomenclature) for each M-FISH image except for EX images. Ground truth image is labelled so
that the gray level of each pixel represents its class number (chromosome type); background pixel
values are zero; pixels in the overlapped regions values are 255. It is used to determine the
accuracy of M-FISH images classification. But translocations are marked such that the full
chromosome is labelled with the class which makes up the most of the chromosome. Images used
for training and testing are taken from this dataset.
4.1 MINIMA SELECTION VALUE
In this method, the minima selection value is very important and it was found heuristically after
several experiments and fixed to 5. As this value increases, area of each region is increased and
total number of regions is decreased. Figure 6 shows the watershed segmentation of an M-FISH
image with different minima selection values. White lines indicate the watershed lines, which are
overlaid on M-FISH image.
Figure 6: Watershed segmentation of an M-FISH image
(a) t = 5, #regions = 302 (b) t = 10, #regions = 194
4.1 SEGMENTATION AND CLASSIFICATION ACCURACY
To compare the performance of segmentation and classification, we need to define two figures of
merits, namely, segmentation accuracy and classification accuracy [11]
Segmentation accuracy is defined as

114
This method provides segmentation accuracy of 98.19% with standard deviation of 2.57% .
Classification accuracy, is defined as
Tables 1 show the comparison of classification accuracy obtained with various approaches,
namely, pixel-by-pixel [2], mean only [10] and proposed approaches - mean & standard
deviation, and mean & standard deviation with post-processing. Same training and testing images
are used for all methods. In the present work, 40 images from the dataset are used for testing and
the proposed approaches (mean & standard deviation, and mean & standard deviation with post-
processing) give improved results compared to other methods [2, 10]. For all of the methods,
classification accuracy can be further improved by proper pre-processing techniques [5 - 7]. The
summary of the classification performance with 10 existing works is given in Table 2. Average
classification accuracies of the proposed approach on best 5, 10, 15 and 40 images are also given
in the Table 2. The results demonstrate that the performance of the proposed post-processing
based approach is superior to most of the existing approaches.
Table 1: Classification Accuracy
No.
Classification Accuracy of Various Approaches
Pixel-by-pixel
[2]
Mean
[10]
Proposed
Mean &
Std. Dev.
Mean & Std. Dev. with
Post-processing
1 87.82 93.65 94.9 95.21
2 91.93 94.16 94.69 94.73
3 64.61 92.2 92.31 92.83
4 90 89.88 92.6 92.6
5 83.63 89.18 90.41 91.33
6 87.43 88.97 90.56 90.74
7 89.78 90.63 90.99 90.33
8 90.57 88.73 89.71 90.26
9 87.03 88.43 89.91 90.09
10 67.15 88.91 89.36 89.69
11 80.79 88.17 88.77 89.24
12 61.24 82.37 87.29 88.22
13 80 87.72 88.89 89.16
14 63.12 81.73 84.39 86.28
15 69.51 77.17 84.68 85.89
16 64.25 74.56 84.49 85.73
17 72.93 80.84 84.56 85.38
18 87.82 83.51 84.26 84.31
19 84.1 85.81 83.96 84.3
20 80.3 82.03 83.57 84.01
21 78.13 81.24 83.03 83.81
22 70.5 80.2 83.27 83.76

115
23 82.36 84.65 82.03 82.62
24 71.57 76.18 77.75 80.85
25 69.64 77.45 80.18 80.82
26 61.38 76.92 80.11 80.55
27 55.35 82.07 78.27 80.53
28 61.76 66.43 80.4 80.4
29 67.45 77.84 79.16 80.2
30 81.13 79.92 79.73 80.16
31 82.67 79.88 79.28 79.73
32 63.92 74.11 79.52 79.41
33 75.5 76.49 78.49 78.97
34 74.03 76.79 77.4 78.76
35 68.81 68.07 77.44 78.32
36 58.23 67.46 75.92 77.56
37 76.42 79.19 76.98 76.98
38 64.92 65.57 73.74 75.82
39 71.8 75.22 75.19 75.19
40 72.62 78.6 73.35 73.6
Avg 74.81 81.32 83.54 84.21
Table 2: Comparison of classification Accuracy of proposed and existing works
Researcher Average Chromosome
Classification Accuracy
Number of Images
Tested
Sampat et. al.[2] 91.40% n/a
Choi et.al.[22] 85.90% 5
Wang [6] 60.36% 6
Sampat et. al.[ 3] 90.5% 5
Wang et. al [5] 87.5% 5
Choi et. al.[ 4] 89.08% 9
Schwartzkopf [24] 68.00% 183
Karvelis P S [10] 89.53% 15
Fazel [9] 89.17% 6
Proposed
84.21% 40
93.34% Best 5
91.70% Best 10
90.44% Best 15

116
4.1. CLASSIFICATION MAP
Figure 7: Classmap obtained for various approaches (a) Ground Truth (b) Mean only method (c)
Mean and standard deviation method (d) Proposed method
The classification map obtained for different approaches are shown in Figure 7. Actual
ground truth is shown in Figure 7.a. The chromosomes pixels marked in circle 1, 2, 3, 4
and 5 are same chromosome type pixels. Chromosomes pixels marked in 1, 2, 3 and 4 are
misclassified by mean only method and as shown in Figure 7.b. The chromosome pixels
in 5 are misclassified, as shown in Figure 7.c, when classifying with mean and standard
deviation approach. This is correctly classified, as shown in Figure 7.d, when post-
processing technique was applied.
5. CONCLUSION
An automated method for M-FISH karyotyping employing watershed segmentation followed by
naïve Bayes classifier is presented. Mean and standard deviation are the features used for
classification. Improved accuracy is obtained by adding a post-processing method that re-
classifies the small segments to neighboring segments based on Bayes theorem. This method
works for all probes and the results are better than pixel-by-pixel classification, which always
produces noisy results. As the classification is done on the watershed regions, the computational
time needed is also much less than the pixels by pixel approach. Use of pre-processing techniques
[5, 6] and manually corrected ground truth [22] will further improve classification accuracy.
Future work is to extend this method for multichannel watershed and to test on much larger data
set.

117
REFERENCES
[1] Speicher M.R., S. G. Ballard, and D. C. Ward, “Karyotyping human chromosomes by combinatorial multi-fluor
FISH,” Nat. Genet., vol. 12, pp. 368 – 375, 1996.
[2] Sampat M. P., A. C. Bovik, J. K. Aggarwal, and K. R. Castleman, “Pixel-by-Pixel Classification of MFISH
Images,” Proc. 24th
IEEE Ann. Intern. Conf. (EMBS), Houston, TX, 2002, pp. 999-1000.
[3] Sampat M. P., A. C. Bovik, J. K. Aggarwal, and K. R. Castleman, “Supervised parametric and non-parametric
classification of chromosome images,” Pattern Recognit., vol. 38, pp. 1209–1223, Aug. 2005.
[4] Choi H., K. R. Castleman, and A. C. Bovik, “Segmentation and fuzzy-logic classification of M-FISH
chromosome images,” in Proc. IEEE Int. Conf. Image Process. (ICIP 2006), Atlanta, GA, pp. 69–72.
[5] Wang. Y. and K. R. Castleman, “Normalization of multicolor fluorescence in situ hybridization (M-FISH)
images for improving color karyotyping,” Cytometry, vol. 64, pp. 101–109, Apr. 2005.
[6] Wang Y., “M-FISH image registration and classification,” in Proc. IEEE Int. Symp. Biomed. Imag.: Nano to
Macro, Apr. 2004, vol. 1, pp. 57–60.
[7] Choi H., K. R. Castleman, and A. C. Bovik, “Color Compensation of Multicolor FISH Images”, In IEEE
Transaction on Medical Imaging, volume 28, January 2009.
[8] Mohammed Alhanjouri, Fatma E. Z. Abou - Chadi, and Nadder Hamdy, “Segmentation and Classification of
Multispectral Chromosome Images,” Twenty Second National Radio Science Conference, March 15-17, Cairo-
Egypt, 2005.
[9] Wang Y., Fazel A., and Derakhshani R, “Classification of multicolor fluorescence in situ hybridization images
using gaussian mixture models,” Proceedings of ANNIE Conference, 2006.
[10] Karvelis P. S., D. I. Fotiadis, M. Syrrou, and I. Georgiou, “A watershed based segmentation method for
multispectral chromosome images classification,” in Proc. 28th IEEE Ann. Intern. Conf. (EMBS), New York,
2006, pp. 3009–3012.
[11] Karvelis P. S., A. T. Tzallas, D. I. Fotiadis, and I. Georgiou, “A multichannel watershed-based segmentation
method for multispectral chromosome classification,” IEEE Trans. on Med. Imag., vol. 27, no. 5, pp. 697-708,
2008.
[12] Georgiou. I., P. Sakaloglou, P. S. Karvelis and D. I. Fotiadis, “Enhancement of the Classification of Multichannel
Chromosome Images using Support Vector Machines,” 31st
Annual International Conference of the IEEE EMBS,
Minneapolis, Minnesota, USA, September 2-6, pp- 3601-3604, 2009.
[13] Wang Y., “Classification of M-FISH Images using Fuzzy C-means Clustering Algorithm and Normalization
Approaches,” 38th
Asilomar Conference on Signals, Systems and Computers, vol. 1, Issue 7-10, Nov, pp. 41 – 44,
2004.
[14] Cao H. and Y. Wang, “Segmentation of M-FISH images for improved classification of chromosomes with an
adaptive Fuzzy C-Means Clustering Algorithm,” In IEEE international symposium on biomedical imaging: from
nano to macro, pp 1442–1445, 2011.
[15] Schwartzkopf W. C., B. L. Evans, and A. C. Bovik, “Minimum Entropy Segmentation Applied to Multi-Spectral
Chromosome Images,” Proc. IEEE Int. Conf. on Image Processing, Thessaloniki, Greece, vol. II, pp. 865-868,
Oct. 7-10, 2001.
[16] Schwartzkopf W. C., B. L. Evans, and A. C. Bovik, “Entropy Estimation for Segmentation of Multi-Spectral
Chromosome Images”, IEEE Southwest Symposium on Image Analysis and Interpretation, April 7-9, pp.234-
238, 2002.
[17] Choi H., A. C. Bovik, and K. R. Castleman, “Maximum-likelihood decomposition of overlapping and touching
M-FISH chromosomes using geometry, size and color information,” In Proceedings of the 28th
IEEE EMBS
annual international conference, 2006, pp 3130–3133, 2006.
[18] Vincent L., and P. Soille, “Watersheds in Digital Spaces: An Efficient Algorithm Based on Immersion
Simulations,” IEEE Transaction on Pattern Analysis and Machine Intelligence, 13, No.6, pp. 583–598, June
1991.
[19] Otsu N., “A threshold selection method from gray-level histograms,” IEEE Trans on Syst., Man, Cybern., vol. 9,
no. 1, pp. 62–66,Jan. 1979.
[20] Duda. R. O., P. E. Hart, and D. G. Stork, “Pattern Classification,” San Diego: Harcourt Brace Jovanovich,
Second ed., November 2000.
[21] http://dip4fish.blogspot.com/2011/11/mfish-dataset-available.html
[22] Hyo Hun Choi, “Automatic Segmentation and Classification of Multiplex-Fluorescence In-Situ Hybridization
Chromosome Images”, Phd Dissertation, The University of Texas at Austin, 2006.
[23] Choi H., K. R. Castleman, and A. C. Bovik, “Joint segmentation and classification of M-FISH chromosome
images,” in Proc. 26th IEEE Annu. Int. Conf (EMBS), San Francisco, CA, 2004, vol. 1, pp. 1636–1639.
[24] Schwartzkopf W. C., A. C. Bovik, and B. L. Evans, “Maximum-likelihood techniques for joint segmentation-
classification of multispectral chromosome images,” IEEE Trans. Med. Imag., vol. 24, pp. 1593-1610, Dec.2005.

M-FISH KARYOTYPING - A NEW APPROACH BASED ON WATERSHED TRANSFORM

More Related Content

What's hot (19)

Similar to M-FISH KARYOTYPING - A NEW APPROACH BASED ON WATERSHED TRANSFORM (20)

More from IJCSEIT Journal (20)

Recently uploaded (20)

M-FISH KARYOTYPING - A NEW APPROACH BASED ON WATERSHED TRANSFORM