Indian Journal of Science and Technology

Abstract

Background: Blood vessel tortuosity analysis in fundus image is used to develop a Clinical Decision Support System (CDSS) to diagnose hypertensive retinopathy, retinopathy of prematurity, cardiovascular problem and stroke. Methods: A Novel approach using machine learning algorithms has been proposed in this paper to determine global tortuosity in a clinical perspective. After preprocessing and blood vessel extraction, eight dimensional feature vector is formed by evaluating tortuosity. By applying Correlation based feature selection procedure top four features are selected for classification purpose. We have collected images from database and hospital, and images are graded by senior ophthalmologist under two labels namely normal and tortuous image. Findings: Highest sensitivity is obtained in the case of SVM classifier. By keeping clinical classification as ground truth highest sensitivity of 96.6% is achieved in case of SVM with radial basis function as kernel. Feature selection process improves the overall sensitivity by 4% and also reduces the computational complexity. In the proposed method decision making is done based on the four features followed by classification, which outperform the other methods in the literature. By using the novel feature-classifier combination, highest sensitivity is obtained. Improvement: In the proposed method decision making is done based on the four features, which outperform the other method proposed in the literature. By using the novel feature-classifier combination, highest sensitivity is obtained.
Keywords: Bayesian Classifier, Correlation Feature Selection, K-Nearest Neighbor, Random Forest, Support Vector Machine, Tortuosity