Using Mathematical Modeling and the Linear Discriminant Analysis to Classify Retinal Vessels into Arteries / Veins
بواقنة, الاء محمد
An-Najah National University
Retinal vessels classification into artery / vein is an important issue; since it helps in early detection of serious retinal diseases, this can avoid blindness. Many features can be extracted from fundus images to distinguish between vessels, but some of these features are redundant, disturbing and time consuming, hence leading to inefficient classification results. In this work, methods for dimensionality reduction are discussed: Linear Discernment Analysis (LDA) and Principal Component Analysis (PCA). We also modified and implemented a two stage method in which PCA is used first to project data into a lower dimension, then LDA is implemented on the resulted projected data to obtain a one dimensional data. Through building a program using MATLAB, all these steps and other image processing are executed to classify vessels. Images from DRIVE database are implemented in some preprocessing steps and features are read from the vessels, then the dimension is reduced using dimensionality reduction technique. Finally, major voting for three classifiers (K-NN, SVM, and Naïve Bayes rule) are used to classify projected data. In the final stage of this work, results are measured and evaluated through sensitivity, specificity and accuracy metrics. Results of the different approaches showed that both LDA and the two stage method are able to reduce the feature space into one dimensional data space and the process time. Although both techniques enhance the classification but Two-stage method showed slightly better results. The proposed two stage method classified the vessels with average accuracy 86.4% on the 14 testing images.