Retinal blood vessel image segmentation and classification of epileptic seizure EEG signals for computer-aided diagnosis

Abstract

In the last decade, the use of computer-aided diagnosis (CAD) is greatly promoted and has improved the diagnosis of diseases to great extent. It has potential to assist doctors in taking final decisions with CAD result as a “second opinion”. Continuous monitoring of Electroencephalogram (EEG) is very cumbersome and may not be possible in some situations, in those places CAD can be a better alternative. This thesis present a new methodology for CAD by finding repetitive local patterns in the biomedical signals. Accuracy of diagnosis depends on how precise abstraction of diseases is captured in these features, and how significant these features are from diagnosis point of view. Recently authors in [1] and [2] experimentally show that diagnosis of certain diseases like epilepsy and diabetes can be achieved by analysing local neighbourhood patterns. A very common example of local pattern is local binary pattern (LBP) which describes local textures, pattern recognition also includes analysis of high level features like Eigen values of hessian matrix like in Frangi vesselness filter. In this work two methodologies are proposed, the first methodology is intended to analyse local pattern in the neighbourhood of retinal image to findcorrelated structure (i.e. retinal vessels) which in turn segments the retinal blood vessels and is evaluated on two well-studied and different databases DRIVE and STARE. It must me noted that Detection or segmentation of retinal blood vessels greatly helps in identifying vessel abnormalities, which is characteristic of retinal vascular disorders including diabetic retinopathy (DR). The second methodology is developed for detection of local pattern corresponding to seizure that effectively perform classification of epileptic seizure and also is evaluated on two databases. There is a dire need of CAD for detection and classification of epileptic seizure as diagnosis of epilepsy based on the visual inspection of EEG signals can be cumbersome and may take a long time, especially for long-duration EEG signals .The first database is of epilepsy obtained from university of Bonn, it has recordings from patients during epileptic attack and in absence of it. It also has recording from normal persons. The second database is obtained from Sir Ganga Ram Hospital, New Delhi it has recording from patients suffering from epilepsy. The advantage of the both methodologies otherthan the very high accuracy, are very low timing complexity that makes these methods well suitable for near real time applications and for devices with limited resources. These advantages makes the proposed methodology well suitable for computer-aided diagnosis of epileptic seizures and segmentation of retinal blood vessels.