Abstract
dc:descriptionThe study of chromosomes is one of the major areas of study for modern genetics because the chromosomes are the carriers of all the genetic material (DNA) of an organism that are transferred from generation to generation by means of reproduction. The assignment of each chromosome to each class from a chromosome image takes time and demands great experience to avoid mistakes that can lead to misdiagnosis. For this reason there have been developed algorithms for image processing and automated analysis of chromosomes. There are several methods and techniques for the cultivation of chromosomes each of which leads to a different type of image. For example, if the chromosomes are cultured according to the protocol G-Banding the resulting image is a gray level image. In this thesis we deal with M-FISH protocol which leads to a multichannel image (6 channels). In this technique the biological experiment has been constructed so that each of the 24 chromosome types (1-22, X, Y) would be reflected in a different color. The purpose of this thesis is the identification and classification of human chromosomes from multichannel M-FISH images. Initially, we developed a method based on the Watershed transform for the region segmentation (grouping pixels with similar characteristics) of chromosomes. The Watershed transform requires a measure of separability between similar areas and for this reason we chose to calculate the multichannel gradient. In this way we achieve a clear separation between areas with different color corresponding to a different chromosome class. The segmentation results are quite satisfactory compared to other methods reported in the literature on the same M-FISH basis images. After segmentation we perform region classification using a statistical classifier that employs the Bayes rule. This classifier is simple to develop and implement and provides satisfactory classification performance. Compared with existing approaches that use Pixel by Pixel classification the proposed region-based method showed better results. We also study the effectiveness image smoothing using Vector Median Filtering and its variants and provide comparative experimental results. One of the problems in the chromosome classification methodologies using multichannel M-FISH images is the fact that they demand a labeled training set to build the classifier. For example a Bayes classifier requires estimating the parameters such as mean and covariance for each of the 24 chromosomes classes. The existence of a methodology that does not require a labeled training set is therefore essential. Such an unsupervised methodology is presented in this thesis. First, we segment the M-FISH image using the Watershed transform to remove the background. Then we estimate which of the remaining pixels have been hybridized or not using the EM algorithm in each of the 5 channels of the image. Then we use a Gaussian Mixture Model to classify each pixel into one of the 24 classes of chromosomes. To build this model for the first time we exploit prior information about which chromosome class emits to each of the five channels. The adaptation of the parameters of Gaussian Mixture Model by using the Maximum a Posterior Expectation Maximization method (MAP EM) results in an increase in the rate of correct classification. It is noteworthy that the proposed unsupervised methodology achieves higher classification rates when compared to supervised classification methodologies. One of the problems for automatic chromosome segmentation methods is the problem of the occlusion of chromosomes. In particular, two important factors influencing the segmentation are the following Chromosomes that overlap with one another, Chromosomes which adjoin one another. We have developed a method that addresses both these problems successfully. Initially we apply a recursive Watershed transform to get an initial assessment of areas of chromosomes. Then for each area of the Watershed transform we determine high curvature points around the perimeter of the chromosomal region. From these points will begin a gradient path which crosses the region and separates the chromosome region where two chromosomes are tangent to each other. If two or more chromosomes overlap each other, then the path splits the chromosomes into pieces. Then we form the Region Adjacency Graph and categorize each area using a region Bayes classifier. If a pair of neighboring regions shares the same class then they are joined together. The method was tested on chromosome images and the success rate of the method was satisfactory. In addition we compared the method with other segmentation methodologies such as the pale paths and the results were much better, especially for the case of overlapping chromosomes.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2012
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Karvelis, Petros
- Καρβέλης, Πέτρος
Subjects
dc:subject × 8Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/28462