Date of Award
Doctor of Philosophy (PhD)
Computational Analysis and Modeling
DNA microarray is an efficient biotechnology tool for scientists to measure the expression levels of large numbers of genes, simultaneously. To obtain the gene expression, microarray image analysis needs to be conducted. Microarray image segmentation is a fundamental step in the microarray analysis process. Segmentation gives the intensities of each probe spot in the array image, and those intensities are used to calculate the gene expression in subsequent analysis procedures. Therefore, more accurate and efficient microarray image segmentation methods are being pursued all the time.
In this dissertation, we are making efforts to obtain more accurate image segmentation results. We improve the Segment Based Contours (SBC) method by implementing a higher order of finite difference schemes in the partial differential equation used in our mathematical model. Therefore, we achieved two improved methods: the 4th order method and the 8th order method. The 4th order method could be applied to segment both the cDNA microarray images and the Affymetrix GeneChips, while the 8 th order method could be applied to segment only the cDNA microarray images, due to the limitation of the current image resolution.
The mathematical derivation shows that both our 4th order method and 8th order method are better approximating the C-V model [Chan & Vese, 2001] than the SBC method, which means they will offer more accurate segmentation results than the SBC method. Besides mathematical proof, we do the practical experiments to double check the conclusion drawn from the mathematical derivation. Both the 4th order method and the 8th order method are used to segment microarray images, and the output segmentation results—the intensities of each probe cell in the microarray image—are being compared to the results from the SBC method and two other mainstream microarray image segmentation methods, the Globaly Optimal Geodesic Active Contours (GOGAC) method and the GeneChip Operating System (GCOS) software, for more valid evaluation.
To give the ground true values of intensities as the standard for different segmentation methods comparison, a microarray image simulator is introduced to generate the simulated images used in our experiments. The simulated microarray images have all the characteristics that real microarray images have, and the true intensity values of each probe spot in the image are provided by this simulator. Intensity values segmented by those segmentation methods are compared to the true intensity values. Therefore, we could evaluate that one segmentation method is more accurate than the other methods if its intensity values are closer to the true values.
We conduct several analysis procedures in the segmentation results comparison part to convince our analysis results. Intensity analysis, paired t-test and Unweighted Pair Group Method with Arithmetic Mean (UPGMA) hierarchy cluster experiments are applied to analyze intensity values of those methods. The segmentation output analysis results show that our 4th order method and the 8th order method could offer more accurate segmentation than the SBC method, the GCOS method and the GOGAC method on some kinds of the microarray images. There are accuracy improvements achieved with the 8 th order method over the 4th order method on the cDNA microarray image. On the Bovine type Affymetrix GeneChip image, there is no significant difference between the 4th order method and the 8th order method.
Li, Yang, "" (2014). Dissertation. 251.