Pattern Recognition and Machine Learning Framework for Automated Analysis of Retinal Images

Abstract

The retina is one of the few locations in the human body that allows direct noninvasive visualization of its anatomical components. A comprehensive analysis of retinal microvasculature structures provides potential clinical biomarkers towards early diagnosis and prognosis of systemic and neurodegenerative diseases. The research focus of this thesis is to develop a series of novel pattern recognition and machine algorithms for automated analysis of retinal vasculature which includes - segmentation of vascular tree, classification of vessels into artery/vein and identification of vessel bifurcation and crossover points. Besides, several geometrical properties at crossover points are analysed to study and quantify the influence of various systemic diseases. Accurate segmentation of retinal vessels is challenging due to the varying nature of vessel calibre, the proximal presence of pathological lesions, strong central vessel reflex and relatively low contrast images. Most existing methods mainly rely on carefully designed hand-crafted features to model the local geometrical appearance of vasculature structures, which often lacks the discriminative capability in segmenting vessels from a noisy and cluttered background. To address this issue, a novel visual attention guided unsupervised feature learning (VA-UFL) approach is proposed to automatically learn the most discriminative features, without complex domain expertise. The VA-UFL approach inherits the combined knowledge of visual attention mechanism and multi-scale contextual information to selectively visualize the most relevant part of the structure in a given local patch. The experiment results show that the proposed approach is shown to be robust to segmentation of thin vessels, strong central vessel reflex, complex crossover structures and fares well on abnormal cases. Further, the discriminative features learned via visual attention mechanism is superior to handcrafted features, and it is easily adaptable to various kind of datasets, where generous training images are often scarce. Detection and classification of vessel junctions are extremely challenging due to spatially varying nature of vessel calibre, which often results in a very close appearance of false bifurcation or crossover points. Existing approaches model the orientation of vessels in a local neighbourhood, without explicitly considering the vessel shape information, which might aid in resolving ambiguities. To address this problem, a novel vessel keypoint descriptor (VKD) is proposed, which is derived from ithe projection of log-polar transformed binary patches. The VKD along with shape based features aids in accurate localization of junctions and classifying them into bifurcations/crossovers. Evaluation results on five challenging datasets show that the designed system is robust to changes in resolution and other variations across datasets. Several geometrical properties at crossover points are analysed to detect and quantify the morphological changes linked to hypertension, stroke and other systemic diseases. To this end, a complete system for detecting arteriovenous (AV) nicking is presented. The entire system is solely based on analysis of vessel morphology, which requires no knowledge of artery/vein class label for vessel segments. Both local orientation and width of vessels are estimated with the aid of VKD, to detect and quantify the vascular changes at crossover points for identifying AV nicking. The proposed solution indicate that the crossover geometrical properties can be considered as essential biomarkers in assessing the progression of various systemic diseases. Separation of arteries and veins is a fundamental prerequisite in the automatic detection of vessel-specific biomarkers associated with systemic and neurodegenerative diseases. In this thesis, a novel graph search metaheuristic approach is proposed for the automatic separation of arteries/veins (A/V) from color fundus images. The proposed method exploits local information to disentangle the complex vascular tree into multiple subtrees, and global information to label these vessel subtrees into arteries and veins. Based on the anatomical uniqueness at vessel crossing and branching points, the vascular tree is split into multiple subtrees containing arteries and veins. Further, the identified vessel subtrees are labelled with A/V based on a set of handcrafted features trained with random forest classifier. The experimental results demonstrate the superiority of the proposed approach in outperforming state-of-the-art methods for A/V separation.