Quantification of blood vessel calibre in retinal images of multi-ethnic school children using a model based approach
Introduction
Examination of retinal blood vessel offers an opportunity to view and image the circulatory system directly (both arterioles and venules). Abnormalities of retinal vessels have been prospectively associated with vascular outcomes in adult life, including coronary heart disease (CHD), stroke and cardiovascular mortality [1]. In particular, narrowing of retinal arterioles has been related to CHD [2]. Changes in retinal vessel calibre in later life have also been associated with established risk factors for cardiovascular disease; narrow arterioles being linked to high levels of blood pressure and markers of body size [3]. These latter associations have also been observed in childhood [4], and retinal vessel tortuosity has also been associated with a number of established risk factors for cardiovascular outcomes in the first decade of life [5]. The unusual variations in diameter along a vein which is termed as venous beading is one of the most powerful predictors of proliferate diabetic retinopathy [6]. Hence, retinal vessel morphology may be an important early marker of circulatory status, and accurate assessment of vessel width, as well as variations in width along the length of a vessel, may be an important physio-marker of vascular health. To date, computerized assessment of vessel calibre from retinal images has been heavily reliant on operator involvement, which is open to measurement error and time consuming, limiting its use in large population based studies. However, reliable blood vessel segmentation and vessel diameter measurement are two critical and challenging technical tasks to automate using image processing. The morphological characteristics of retinal images of children and premature infants are very different from those of the adult retina. Choroidal vessels are more visible alongside the retinal vessels in retinal images taken from premature infants. A bright central reflex is more apparent in arterioles than venules and is typically found in the retinal images of younger patents. In this paper, we report an automated methodology for measuring retinal vessel diameter in child retinal images. The proposed methodology aims to quantify the blood vessel calibre in noisy and pathological retinal images of school children with uneven illumination and contains complex vessel profiles. We have utilized a new public retinal image dataset named as CHASE_DB1 [7], for evaluation of the methodology. This dataset includes retinal images of 9- and 10-year-old children of different ethnic origin, along with the ground truths of blood vessel segmentation as well as for vessel width measurement. The database includes images with stark differences in background levels of retinal pigmentation (being more pigmented in South Asians compared to white Europeans). The proposed system combines methods that were previously described by our group [8], [9], [10], [11], also proposes new techniques that automatically localizes and segments the retinal vasculature and estimates width along the length of a vessel.
The organization of the paper is as follows. In Section 2, we give an overview of retinal vessel width measurement techniques. Section 3 presents the details about the recently introduced retinal image database CHASE_DB1, which is used to evaluate the methodology. The methodology and implementation details are presented in Section 4. In Section 5, the methodology is quantitatively and qualitatively evaluated on vessel segments containing various types of vessel profiles, as well as using different monochromatic representations of colour space. Finally, the conclusions are given in Section 6.
Section snippets
Overview of retinal vessel width measurement techniques
The detection, segmentation of retinal blood vessels and measurement of vessel calibre are two important and non-trivial tasks in any automated system. A detailed review of methods to detect vessels can be found in [12], [13]. These methods locate vessels, but do not directly determine vessel width, and additional methods are needed for accurate measurement of the vessel width. Many different techniques have been used to estimate vessel width, all of which are predicated upon the idea of
Materials
The CHASE_DB1 [7], a recently introduced retinal vessel reference dataset [8] acquired from multi-ethnic school children, is used to assess the diameter measurement results. The CHASE_DB1 database is comprised of the retinal images obtained as part of the Child Heart and Health Study in England (CHASE), a cardiovascular health survey in 200 primary schools in London, Birmingham, and Leicester [22]. The part of the investigation involving ocular imaging was carried out in 46 schools and
Method summary
An initial segmentation process was needed to extract the vessel segment profiles. An ensemble classification based approach for vessel segmentation has been proposed by our group [8], [25] and we have used the vessel probability map image obtained by this method as the initial vascular tree. A binary segmentation image is generated by thresholding the probability map image. A scale space skeletonization approach is then applied to the binary vessel segmentation to find the vessel centrelines.
Results and discussion
The methodology is evaluated on CHASE_DB1; a retinal image database that contains images of multi-ethnic children aged 10–11 years. The images have a lot of inter-image and intra-image variation. They are seen to have non-uniform background levels of pigmentation (being more pigmented in South Asians compared to white Europeans), poor contrast of blood vessels as compared with the background, and vessels, particularly arterioles, often have bright central vessel reflexes. The results show the
Conclusion
We have introduced a robust and accurate methodology for measuring the calibre of vessel segments in retinal images of multi-ethnic children. The algorithm uses the vessel probability map image resulting from ensemble classification to find vessel centrelines resulting in a one pixel wide binary vascular tree image. The vessel branch points and crossovers are detected and removed from the vessel centreline image, therefore resulting in forming an image containing vessel segments. Later, the
Acknowledgements
The authors thank the participating pupils, and their parents; the Birmingham Optical Group (Birmingham, UK) who provided a Nidek NM-200D handheld fundus camera for the duration of the study. The retinal imaging component of CHASE was supported by a grant from the BUPA Foundation (755/G25).
References (42)
- et al.
Retinal microvascular abnormalities and 10-year cardiovascular mortality: a population-based case–control study
Ophthalmology
(2003) - et al.
An approach to localize the retinal blood vessels using bit planes and centerline detection
Comput Methods Programs Biomed
(2012) - et al.
Blood vessel segmentation methodologies in retinal images – A survey
Computer methods and programs in biomedicine
(2012) - et al.
Algorithms for digital image processing in diabetic retinopathy
Comput Med Imaging Graph
(2009) - et al.
Automated grading of venous beading
Comput Biomed Res
(1995) - et al.
Characterization of changes in blood vessel width and tortuosity in retinopathy of prematurity using image analysis
Med Image Anal
(2002) - et al.
Automatic model-based tracing algorithm for vessel segmentation and diameter estimation
Comput Methods Programs Biomed
(2010) - et al.
Automatic detection and characterisation of retinal vessel tree bifurcations and crossovers in eye fundus images
Comput Methods Programs Biomed
(2011) - et al.
Quantitative measurement of changes in retinal vessel diameter in ocular fundus images
- et al.
A model based method for retinal blood vessel detection
Comput Biol Med
(2004)