2538 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 9, SEPTEMBER 2012 An Ensemble Classification-Based Approach Applied to Retinal Blood Vessel Segmentation Muhammad Moazam Fraz ∗ , Paolo Remagnino, Andreas Hoppe, Bunyarit Uyyanonvara, Alicja R. Rudnicka, Christopher G. Owen, and Sarah A. Barman Abstract—This paper presents a new supervised method for seg- mentation of blood vessels in retinal photographs. This method uses an ensemble system of bagged and boosted decision trees and uti- lizes a feature vector based on the orientation analysis of gradient vector field, morphological transformation, line strength measures, and Gabor filter responses. The feature vector encodes information to handle the healthy as well as the pathological retinal image. The method is evaluated on the publicly available DRIVE and STARE databases, frequently used for this purpose and also on a new pub- lic retinal vessel reference dataset CHASE_DB1 which is a subset of retinal images of multiethnic children from the Child Heart and Health Study in England (CHASE) dataset. The performance of the ensemble system is evaluated in detail and the incurred accuracy, speed, robustness, and simplicity make the algorithm a suitable tool for automated retinal image analysis. Index Terms—Ensemble classification, medical image analysis, retinal blood vessels, segmentation. I. INTRODUCTION F UNDUS imaging is being increasingly used to establish retinal normality, and to diagnose/monitor retinal abnor- mality. A number of retinal blood vessel features (e.g., arteriolar microaneurysms, nicking, narrowing) have been linked to sys- temic disease, and the morphological characteristics of retinal blood vessels themselves have been associated with cardiovas- cular and coronary disease in adult life [1] and with retinopathy of prematurity in infancy [2]. The morphology of retinal vessels (particularly arterioles) has also been linked to cardiovascular risk factors both in early and adult life [3]. The detection and analysis of the retinal vasculature is useful in the implementation of screening programs for diabetic retinopathy, the evaluation of retinopathy of prematurity, foveal avascular region detection, Manuscript received January 14, 2012; revised April 27, 2012; accepted June 16, 2012. Date of publication June 22, 2012; date of current version August 16, 2012. Asterisk indicates corresponding author. ∗ M. M. Fraz, P. Remagnino, A. Hoppe, and S. A. Barman are with the Digital Imaging Research Centre, Faculty of Science, Engineering and Com- puting, Kingston University London, Surrey KT1 2EE, U.K. (e-mail: moazam. fraz@kingston.ac.uk; P.Remagnino@kingston.ac.uk; a.hoppe@kingston.ac.uk; s.barman@kingston.ac.uk). B. Uyyanonvara is with the Department of Information Technology, Sirind- horn International Institute of Technology, Thammasat University, Bangkok 10200, Thailand (e-mail: bunyarit@siit.tu.ac.th). A. R. Rudnicka and C. G. Owen are with the Division of Population Health Sciences and Education, St. George’s, University of London, London SW17 0RE, U.K. (e-mail: arudnick@sgul.ac.uk; cowen@sgul.ac.uk). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TBME.2012.2205687 arteriolar narrowing detection, the determination of the rela- tionship between vessel tortuosity and hypertensive retinopathy, measurement of vessel diameter to diagnose cardiovascular dis- eases and hypertension, and computer-assisted laser surgery [4]. The retinal map generation and branch point detection have been used for temporal or multimodal image registration, retinal im- age mosaic synthesis, optic disc identification, and fovea local- ization and for biometric identification [4]. Retinal vessels are composed of arteriolars and venules, which appear as elongated branched features emanating from the optic disc within a retinal image. Retinal vessels often have strong light reflexes along their centerline, which is more apparent on arteriolars than venules, and in younger compared to older patients, especially those with hypertension. The vessel cross-sectional intensity profiles approximate to a Gaussian shape, or a mixture of Gaussians in the case where a central vessel reflex is present. The nonvessel region in the retina is not smooth due to the presence of the bright and dark lesions which includes hemorrhages, exudates, drusen, and the optic disc boundary. Most of the existing retinal segmentation methodologies are evaluated on the healthy retinal images free from the pathologies; therefore, their performance can be considerably degraded in the presence of lesions. This paper presents a new supervised method for segmenta- tion of blood vessels by using an ensemble classifier of boosted and bagged decision trees. The feature vector is based on gradi- ent orientation analysis (GOA), morphological transformation with linear structuring element; line strength measures and the Gabor filter response which encodes information to success- fully handle both normal and pathological retinas with bright and dark lesions simultaneously. The classifier based on the boot strapped and boosted decision trees is a classic ensemble classifier which has been widely used in many application areas of image analysis, but has not been applied within the frame- work of retinal vessel segmentation for automated retinal image analysis. The obtained performance metrics illustrate that this method outperforms most of the state-of-the-art methodologies of retinal vessel segmentation. The method is training set robust as it offers a better performance even when it is trained on the DRIVE database [5] and tested on the STARE database [6], thus making it suitable for images taken under different conditions without retraining. This attribute is particularly useful when implementing the screening programs over a large multiethnic population where there is a large variability in the background pigmentation level of the acquired retinal images. Moreover, the algorithm is computationally fast in training and classifi- cation and needs fewer samples for training. The classification accuracy of the ensemble can be estimated during the training 0018-9294/$31.00 © 2012 IEEE