Exploration of Computational Intelligence Techniques for Static, Time-series and Imagery Biosignal Processing

Abstract

This thesis presents the application of computational intelligence techniques to signal processing of static, time-series and imagery bio-signal data. In the process two very important diseases were diagnosed. These diseases are breast cancer and cardiac arrhythmia. Out of the different detection techniques for breast cancer the ones that were used in this research are the Fine Needle Aspiration (FNA) and the mammography. The fine needle aspiration (FNA) procedure consists of excising a small sample of suspected lesion from the breast using a fine needle. The sample is studied under a microscope for the cell sizes and shapes. From the collective characteristics of these features the pathologist decides whether the cell is malignant(cancerous) or benign(noncancerous). Data for the FNA technique was obtained from the Diagnostic Wisconsin Breast Cancer (DWBC) database, an example of static data. The database contains many malignant and benign sample feature value and their results. To assist physicians in diagnosis, a computational intelligence detection approach was devised. In this method experiments were performed using the computational intelligence network of Cartesian Genetic Programming evolved Artificial Neural Network (CGPANN). Feature values of the samples were normalized and a part of them used to train a CGPANN. The trained network was then tested with rest of the samples. Experiments conducted with the FNA dataset resulted in more than 99% accuracy. The second diagnostic method, the mammography, is also used widely for breast cancer screening. It consists of taking a high resolution x-ray image of the breast that is suspected of cancer. The two main abnormalities in a breast can be masses and microcalcifications. In order to assist radiologists in diagnosis a method was developed, that can classify a mass or microcalcification appearing in a mammogram to be either benign or malignant. The data for this work was obtained from the Digital Database for Screening Mammography (DDSM), an example for imagery data. The method consists of calculating the Haralick’s statistical parameters of the suspected lesion. A CGPANN network was trained with large number of these parameters, extracted from mammograms found in the database. The trained network classified both masses and microcalcifications with accuracy=90.58%, sensitivity=85.32% and specificity=95.84%. In the case of cardiac arrhythmia the ECG signals were obtained from MIT-BIH database, an example of time-series data. For automatic detection of Cardiac Arrhythmia an algorithm was developed. This algorithm applies digital signal processing and logical operations to the time domain Electro-Cardiogram (ECG) signal and hence detects the fiducial points of an ECG complex. From the fiducial points, the lengths and slopes of a number of segments; and amplitudes of peaks are determined. These parameter values are applied to CGPANN to classify the beats. To make the system capable of classifying unknown ECG it was trained with the parameters extracted from ECG signals available at MIT-BIH database. All these parameters bear important information about the different arrhythmia. Three different experimental setups were designed, each setup improving the performance of the previous one. In the third setup, with the inclusion of digital logic unit, seven arrhythmia types were detected, with four types having accuracy value of 94% and above. In all experiments, the CGPANN was first trained with parameters extracted from a part of sample ECG, together with their arrhythmia types; and then tested with another part of the data. This algorithm can be implemented in real time on beat to beat basis. A future enhancement to this system is to implement the algorithm in programmable hardware and subsequently used in systems like Implantable Cardioverter Defibrilators (ICD) that need correct detection of life threatening beats to apply an electrical impulse to the heart at the right moment.