Classification of cancer microarray data using block processing and transforms for accurate prediction

Abstract

The objective of this thesis is to develop an efficient Microarray newlineData Classification (MDC) system for cancer prediction. The prediction is newlinedone based on dominant features extraction from microarray data using signal newlineprocessing and machine learning algorithms. The dominant features are newlineextracted using transforms such as Empirical Wavelet Transform (EWT), newlineDiscrete Wavelet Transform (DWT) and Stationary Wavelet Transform newline(SWT) and the classification is performed using K-Nearest Neighbour (KNN) newlineand Support Vector Machine (SVM). newlineThe proposed cancer prediction from gene sequences of MDC newlinesystem consists of (i) feature extraction by DWT, SWT and EWT, (ii) feature newlineselection by t-test and (iii) classification by KNN and SVM. Initially, newlinemicroarray data is applied with above wavelet transforms. EWT provides newlinemore sparse representation than DWT and SWT. In feature selection, Block- newlineBy-Block (BBB) procedure with predefined block sizes in powers of 2, starts newlinefrom 128 to 2048 is applied. BBB avoids the processing of whole microarray newlinedata. BBB also prevents the information loss, while selecting whole newlinemicroarray data. The selected feature set from MDC has high discrimination newlinebetween classes for MDC through KNN and SVM. newlineThe MDC system predicts cancer from five microarray datasets newlinesuch as colon, breast, leukemia, Central Nervous System (CNS) and ovarian. newlineThe proposed methods such as DWT-MDC, SWT-MDC and EWT-MDC are newlineapplied to above Gene Microarray data sets. The optimal block size for colon newlinecancer is 128, 256 for CNS and leukemia dataset and 512 for breast and newlineovarian dataset. In this proposed system for colon cancer newline