Automated Classification of Echo-Cardiography Images Using Texture Analysis Methods

This work studies the prospects and efficacy of identifying and classifying disease conditions from digital medical images by applying image processing and pattern recognition techniques. Specifically echo-cardiography images showing the left ventricle of the human heart are being used to identify three cardiac conditions viz. normal, dilated cardiomyopathy and hypertrophic cardiomyopathy. To differentiate between the classes, texture information extracted from the images are used to generate data models using three different approaches : by using a complex Gabor filter, by using Hu invariant moments and by using Law’s texture detection method. In all cases the system is first trained in a supervised manner using training samples for each class, and then the system is tested using similar but not identical testing samples. Each test sample is classified in an automated manner by computing differences with the trained samples of each class and identifying which class produces the least average difference. Each approach is studied by varying associated parameters to find out optimum performance levels. Recognition accuracies produced are found to be comparable with the best results reported in extant literature.