Abstract:
Retinal fundus images are coloured images obtained through specially designed cameras through a dilated pupil of the patient. Analysis of these images is being used to detect retinal vascular abnormalities to provide insight into onset or severity of retinopathies specially hypertensive or diabetic retinopathy. One of the common yet significant change that occurs is the change in vascular shape; in that the vessel(s) becomes non-periodically twisted; more generally termed as an increase in tortuosity. This thesis presents a simple and reasonably accurate algorithm to classify a vessel as tortuous or not, through determining a set of features. A new set of features is proposed in this research for reliable detection of vascular changes. The proposed method uses One Class SVM (OC-SVM), commonly used for anomaly detection. The reason for using OC-SVM is that the ratio of tortuous vessels as compared to normal ones is very low and they mostly appear as anomaly when compared with normal vessels. A local dataset of 100 fundus images is used for evaluation. The dataset has manually extracted vessels, veins and arteries as ground truth and also contains annotation with respect to vessel tortuosity. The experiments are conducted by randomly dividing data into 60 per cent for training and 40 per cent for testing. The experiments are repeated 10 times and average results are reported. The results show that the proposed system provides an efficient non-invasive technique to detect tortuous vessels and an important step towards detecting IRMA.
