AUTOMATED CLASSIFICATION OF RETINAL DISEASES IN STARE DATABASE

Abstract

People of all age groups are affected by a number of retinal diseases. These diseases are identified by conducting different medical examinations primary of which are visual examinations. One of the key issues in visual diagnosis of diseases is the human error due to poor decision making, for that a number of research projects are conducted which use the visual data and directly or the symptoms and generate decisions. An upcoming inter-disciplinary technology named Computer Aided Medical Diagnostic System provides precise detection and prediction of disease. Automated image analysis methods are far more helpful for early identification and evaluation of disease as compared to cryptic and time taking manual techniques of digital medical imaging. This thesis aims to develop an automated method for identification of eye disorders that affect the human retina which if left unidentified may result in blindness due to delayed detection and analysis. Image data was acquired by publically available STARE Database having fundus images and by implementation of exclusion and inclusion criterion it was pre-processed on MATLAB. Initial pre-processing increased the significance of the data to be analyzed. From 186 images, 16 diseases and 22 features were deduced. A support vector machine classifier was used for automated identification and classification, resulting in an accuracy of 94% and specificity of 98%. In the chosen technique sensitivity, specificity and accuracy of the results was affected by the problem of one-sided data. For the reduction of dimensionality of data (redundancy reduction) principal component analysis was employed. 5, 10 and 22 Principal components were obtained to reduce the amount of variables. PCA was performed prior to training of the SVM, results for different data dimensionality were compared for completeness.