Development of an automated system to classify retinal vessels into arteries and veins

Abstract

There are some evidence of the association between the calibre of the retinal blood vessels and hypertension. Computer-assisted procedures have been proposed to measure the calibre of retinal blood vessels from high-resolution photopraphs. Most of them are in fact semi-automatic. Our objective in this paper is twofold, to develop a totally automated system to classify retinal vessels into arteries and veins and to compare the measurements of the arteriolar-to-venular diameter ratio (AVR) computed from the system with those computed from observers.

Our classification method consists of four steps. First, we obtain the vascular tree structure using a segmentation algorithm. Then, we extract the profiles. After that, we select the best feature vectors to distinguish between veins and arteries. Finally, we use a clustering algorithm to classify each detected vessel as an artery or a vein.

Our results show that compared with an observer-based method, our method achieves high sensitivity and specificity in the automated detection of retinal arteries and veins. In addition the system is robust enough independently of the radii finally chosen, which makes it more trustworthy in its clinical application. We conclude that the system represents an automatic method of detecting arteries and veins to measure the calibre of retinal microcirculation across digital pictures of the eye fundus.

Introduction

The retina is the only place where the vascular system is visible using simple instrumentation. There are several studies providing some evidence of the association between the calibre of the retinal blood vessels, i.e. arteries and veins, and diabetes [1], as well as cardiovascular diseases such as hypertension, strokes, and coronary heart disease [2], [3], [4], [5], [6], [7]. Narrower retinal arteriolar calibre has been associated with hypertension and may even precede clinical hypertension [3], [4], [5], [8], [9]. In contrast, wider venular calibre has been shown to be associated with an increased risk of stroke and cardiovascular events [7], [10], [11], [12]. A measurement that has been used to quantify these variations in the calibre of the vessels is the arteriolar-to-venular diameter ratio (AVR) [3].

To achieve accurate and, more importantly, reproducible results, computer-assisted procedures have been proposed to measure the calibre of retinal blood vessels from high-resolution photopraphs [3], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27]. Most of these procedures are in fact semi-automatic [14], [15], [16], [17], [18], [20], [21], [25], [26], in the sense that they involve a manual function where a skilled observer must first determine the vessel type, artery or vein. In fact, our group recently described a semi-automatic computerised system for evaluation of the AVR that has shown very good reproducibility [28]. Using this method we observed a regression of early hypertension-related alterations in retinal microcirculation after 6 months of antihypertensive treatment [17].

In light of this, we have tried to improve the specificity of the linear method we used by applying a model based on elastic curves called ‘snakes’ [29], with which the specificity is improved, supporting the same sensitivity in the detection of retinal vessels, but which continues to have the limitations of a semi-automatic method that needs the intervention of an observer, thus limiting its application to clinical practice.

An automatic method was already proposed in 2003 by Grisan and Ruggeri [13]. Two were the main differences with respect to the method we proposed here. First, in Grisan and Ruggeri's method, images were first pre-processed in order to compensate both, intra and inter-variability. Second, analyses were restricted to main vessels, excluding small arterioles and venoules [13]. In fact, whereas for main vessels (i.e. those belonging to the main vascular arcades) results were very good (only 6.7% of error), results corresponding to secondary vessels were not so satisfactory (20.7% of error).

Recently, Niemeijer et al. [27] also propose an automatic procedure in classifying retinal vessels into arteries and veins. The procedure provides very good results (area under the ROC of 0.88). However, it is applied only to large vessels, because they tend to be easier to classify and they do not require the use of vessel connectivity Information [27].

Our main objective in this paper is to develop a totally automated system to classify all the retinal vessels (large and smaller) into arteries and veins. However, as stated above, our final aim is to develop a system that can be easily applied to clinical practice. Therefore, our secondary objective is to compare the measurements of the AVR computed with the results obtained from the system with those computed with the results from observers. In fact, very recently, Tramontan et al. [30] proposed a web tool for the AVR estimation, very similar to that described here.

This paper is organised as follows. After the short introduction given in this section, we describe in detail our classification methodology in Section 2. In Section 3 we provide the material and the materials in order to validate our proposed system. In Section 4, we provide the results of the validation and also use these results to measure the AVR. Finally, we finish with Section 5, where we provide some concluding remarks and Section 6 where we give our conclusions.