An automated tracking approach for extraction of retinal vasculature in fundus images.

Purpose: To present a novel automated method for tracking and detection of retinal blood vessels in fundus images.
Methods: For every pixel in retinal images, a feature vector was computed utilizing multiscale analysis based on Gabor filters. To classify the pixels based on their extracted features as vascular or non-vascular, various classifiers including Quadratic Gaussian (QG), K-Nearest Neighbors (KNN), and Neural Networks (NN) were investigated. The accuracy of classifiers was evaluated using Receiver Operating Characteristic (ROC) curve analysis in addition to sensitivity and specificity measurements. We opted for an NN model due to its superior performance in classification of retinal pixels as vascular and non-vascular.
Results: The proposed method achieved an overall accuracy of 96.9%, sensitivity of 96.8%, and specificity of 97.3% for identification of retinal blood vessels using a dataset of 40 images. The area under the ROC curve reached a value of 0.967.
Conclusion: Automated tracking and identification of retinal blood vessels based on Gabor filters and neural network classifiers seems highly successful. Through a comprehensive optimization process of operational parameters, our proposed scheme does not require any user intervention and has consistent performance for both normal and abnormal images.