The Classification Method Of DNA Microarray Data Based On Evolvable Hardware

Since all kinds of cancers have their own features, specified treatments are desired toachieve the maximize efficacy and minimize toxicity, which makes the cancerclassification as a key in the cancer treatment. Nowadays the clinical cancer diagnosis ismainly based on the morphological information. Some tumors may have significantdifference on clinical manifestations, although they share similar morphologicalinformation. Therefore, different treatments are required. In recent years, the developmentof DNA microarray technology provides a new approach for the cancer research. Thecomplex relationship among genes can be discovered through microarray expressionprofile, which provides basis for the research of pathogenesis of cancer and relatedtherapeutic drugs selection. DNA microarray data-based cancer classification has becomeone of the key fields in cancer research. However, the microarray dataset has somedisadvantages, including limited number of samples, high dimension, high noisy, highredundancy and unbalanced data distribution, which brings a great challenge for DNAmicroarray data-based cancer classification.A powerful classification performance is hard to be obtained using a conventionalpattern recognition scheme due to some special features of DNA microarray data. Due tothe fact that the current DNA microarray-based cancer classification has thedisadvantages of long learning time, high recognition period, and poor readability of thelearning results etc, an effective and accurate classification system which is based on theevolvable hardware (EHW) technology is proposed in this thesis. In contrast to thetraditional hardware where structures and functions are irreversible fixed once thehardware is manufactured, EHW refers to a type of hardware whose architecture andfunctions can change dynamically and autonomously by interaction with the environment.Based on the efficient and fast programmable logic devices, the EHW classificationsystem has the advantages of online adaptation, real-time processing ability and areadability of the learning results. For solving the above mentioned limitations in atraditional DNA microarray-based cancer classification system, this thesis discusses thefollowing contents.1) In order to solve the limitations of poor stability and low recognition rate within asingle EHW classifier, an EHW-based multi-classifier model is proposed for the classification of microarray data. Firstly, DNA microarray data is processed through afeature selection operation with a filter-based signal to noise ratio (SNR) feature selectionscheme. Then, various base classifiers are evolved by the virtual reconfigurablearchitecture (VRA)-based EHW through learning datasets several times. Finally, theoutputs of all base classifiers are integrated by a majority voting strategy. In order toreduce the complexity of the evolution of base classifier, an incremental evolutionstrategy is adopted in the evolutionary process. A pipeline technique is also employed inthis thesis to reduce the system evolution times.2) In this work, a selective ensemble learning method is introduced to reduce theconsumption of the system hardware resources, and further improve the systemrecognition rate. The original training set is randomly divided into a training set and avalidation set. In order to increase the differences among the base EHW classifiers, theoriginal data is partitioned several times. Different base EHW classifiers are evolvedthrough learning their corresponding partitioned training sets. Finally, the validation setsare used to test the evolved classifiers and the base classifiers with output less error areselected for the final system integration. The experimental results on different datasetsshow that the proposed method is able to build a stable and efficient classification systemfor DNA microarray data.