Retinal Vessel Tortuosity Characterisation Using Boosted-Linear Discriminant Analysis

Retinal vessel tortuosity is an early indicator of different retinopathies. Although various automated methods in determining retinal vessel tortuosity have been proposed in the literature, there are needs for further study. This study extracted three different features namely distance metric, normalized hybrid metric and non-normalized hybrid metric from the thinned vessels. The weights of vessel data samples were dynamically updated using the Adaboost with linear discriminant analysis (LDA) and the feature correlation was used to facilitate the selection of the best feature combination at each of the boosting iteration rather than a single feature that minimizes the weighted error at each of the iterations. Adaboost with LDA method is then used for the classification of the retinal vessels as either tortuous or normal using a majority voting method. The proposed method achieves the accuracy rate of 100% for the training sample sizes of 70%, 80% and 90%.