Certain investigations on retinal vessel extraction and artery vein classification using machine learning and deep learning approaches

Abstract

Commonly seen retinal diseases like Diabetic retinopathy, Glaucoma, Hypertension and even some cardiac related diseases are diagnosed through the anatomical changes in vascular pattern and artery/vein ratio. The extraction of blood vessel from the retinal images and the vessel classification into artery/vein are required for the diagnosis based on vascular pattern analysis. The objective of this research work is the development of algorithms for the retinal vessel extraction and its classification into artery /vein vessels. The first module of the work comprises the supervised algorithms with BAT optimization algorithm based feature selection of the manually specified features for vessel extraction and artery/vein classification. The features which have more information about edges are preferable for vessel extraction. Hence features like Green channel intensity, Gradient, Gaussian filter, Phase congruence and Divergence are considered as feature vector. Then Bat optimization algorithm is applied to select the significant features from the feature vector. The vessel extraction by the Bat algorithm selected features is observed to be better than by all the initially specified features and it also reduces the computational burden by reducing the feature dimensionality. Features which have the information about brightness and reflectance of the vessels are required to discriminate artery with vein, hence intensity, profile and patch features are considered to extract this information. Then Bat optimization algorithm is applied to select the predominant features for artery/vein classification and the achieved artery/vein classification performance is better, compared with that of all feature based classification. In the second module of the work, a fully convolved neural network (FCNN) is proposed for vessel extraction and artery/vein classification. The proposed FCNN is the encoder - decoder architecture with five stages each newline