Segmentation, Classification and Morphometric Analysis of Retinal Vasculature

Abstract

The abnormality in shape and size of retinal blood vessels is considered as an important diagnostic indicator and predictor of many ophthalmic, microvascular, and systemic diseases. Automated assessment and morphometric analysis of retinal vasculature is a useful physio-marker and potential predictor of cardiovascular status in later life and enables large population-based mass screening for early diagnosis. This work aims to develop an image analysis pipeline for early detection and quantification of Hypertensive Retinopathy (HR), based on morphometric analysis of retinal vasculature. The progression of HR has a very different course in retinal vessels i.e. arterioles and venules, with some biomarkers associated with only one type of vessels. Robust vessel segmentation and classification of retinal vasculature into arterioles and venules is the first step in developing an automated image-based HR diagnostic system. This is followed by the measuring vessel width, calculating the arterio-venous width ratio (AVR) and vessel tortuosity index for HR grading. In this context, this work presents an optimized trainable B-COSFIRE filter for retinal vessel segmentation. Moreover, a fully convolutional neural network-based method is presented for simultaneous retinal vessel segmentation and classification into arteriole and venule. The method is further extended by proposing a segment and pixel level multi-loss function for vessel segmentation and AV classification. The region of interest around the retinal Optic Disc (OD) is identified by accurate localization of OD. The width of segmented arteriole and venule is computed in the specific region of interest around the OD, which is used to compute the arterio-venous width ratio (AVR). Furthermore, morphometric analysis of segmented retinal vasculature is performed to compute 14 tortuosity metrics and vessel tortuosity severity levels. The AVR and vessel tortuosity index are an important indicator to quantify HR severity. iii A new retinal image morphology (RVM) dataset of 704 retinal images is created with the pixellevel annotations available for vessel segmentation, arteriole-venule classification, and imagelevel labels for vessel tortuosity index and HR grade. This dataset will be made public to ignite further research in this domain. The segmentation and classification are evaluated on publicly available datasets of DRIVE, STARE, CHASE-DB1, and on the proposed AV dataset. The AVR and tortuosity indexes are evaluated on the proposed AV dataset. The results revealed an enhanced trainable B-COSFIRE filter for retinal vessel segmentation gives improved performance over existing methods. Moreover, the proposed semantic segmentation multi-loss function-based deep learning method outperforms the previous AV-classification methods in terms of sensitivity, specificity, and accuracy on the DRIVE, AVRDB and proposed AV datasets. The results show that the proposed pipeline can potentially be used to develop a robust computer-aided system for diagnostic retinal image analysis.