Detection of neovascularization in retinal images using multivariate m-Mediods based classifier

Abstract

Diabetic retinopathy is a progressive eye disease and one of the leading causes of blindness all over the world. New blood vessels (neovascularization) start growing at advance stage of diabetic retinopathy known as proliferative diabetic retinopathy. Early and accurate detection of proliferative diabetic retinopathy is very important and crucial for protection of patient's vision. Automated systems for detection of proliferative diabetic retinopathy should identify between normal and abnormal vessels present in digital retinal image. In this paper, we proposed a new method for detection of abnormal blood vessels and grading of proliferative diabetic retinopathy using multivariate m-Mediods based classifier. The system extracts the vascular pattern and optic disc using a multilayered thresholding technique and Hough transform respectively. It grades the fundus image in different categories of proliferative diabetic retinopathy using classification and optic disc coordinates. The proposed method is evaluated using publicly available retinal image databases and results show that the proposed system detects and grades proliferative diabetic retinopathy with high accuracy.

Introduction

Diabetic retinopathy (DR) is a progressive eye disease that is caused by the increase of insulin in the blood and can cause blindness if not detected timely [1]. DR is also caused by the microvascular complication of diabetes and it is one of the main sources of vision impairment. A number of studies have shown that DR is one of the major causes of blindness in industrialized countries [2]. One out of five patients with newly discovered type II diabetes has DR at the time of diagnosis, where DR almost never occurs in first five years after diagnosis of type I diabetes [1]. The common symptoms of diabetic retinopathy are blurred vision, floaters and flashes, and sudden loss of vision [3].

Healthy retina contains blood vessels, optic disc (OD), macula and fovea as main components [2] whereas an affected retina may also contain different signs of DR. DR is broadly divided into two stages i.e. non proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). NPDR also known as background DR contains the early signs of presence of DR such as microaneurysms and dot haemorrhages caused by the breaks in tiny small vessels called capillaries [4]. As the disease progresses, more signs of DR appear such as hard and soft exudates which are caused by the leakage of fats and proteins on the surface of retina known as severe NPDR. PDR is an advanced stage of DR and it is divided into two stages; i.e. neovascularization on optic disc (NVD) and neovascularization elsewhere (NVE) [4]. In PDR, retina sends signals for nourishment of oxygen deprived areas. As a result of this, new blood vessels start growing in different regions of retina to supply blood which is a good thing but these new vessels are weak and their walls are thin and fragile. These infant vessels may easily start leaking blood on surface of retina and cause severe vision loss, even blindness [5]. Fig. 1 shows digital images of healthy retina and retina affected with PDR.

Digital retinal images are used in computer aided diagnostic (CAD) systems for screening of DR and its different stages. Different signs of NPDR and PDR appear with different properties on the surface of retina and it is the goal of CAD systems to identify these signs for timely and accurate treatment of DR. A number of fundus image databases are publicly available for the purpose of evaluation and testing of proposed systems for the diagnosis of retinal diseases. Fig. 2 shows digital images of healthy retina and retina affected with PDR.

A number of computer aided diagnostic systems have been proposed in literature for early detection of DR and lesions related to NPDR [6], [7], [8], [9], [10]. All these systems covered early signs of DR such as microaneurysms and exudates. Although it is important to detect DR at its early stage, however early detection of PDR is also very important to save patient's vision. Only a few studies have been carried out on PDR. Goatman et al. [11] proposed a method for automatically detecting new vessels on the optic disc by detecting blood vessels and using support vector machine to categorize a vessel segment as normal or abnormal. They presented a good set of features to detect neovascularization but limited their scope with NVD only. This limitation makes it relatively easy to detect abnormal vessels from a specific region of interest instead of looking for abnormal vessels from whole image. A multiscale amplitude-modulation–frequency-modulation (AM–FM) based method for discriminating between normal and pathological retinal images containing neovascularization was presented in [12]. They used only 120 regions from just 15 images with different DR signs and classified those signs into different categories. A study on 27 fluorescein angiogram images was done by Jelinek et al. [13] to examine vascular pattern characteristics to detect PDR. Nayak et al. presented a simple artificial neural network based method for detection of PDR with help of just blood vessel area and perimeter [14]. A small database of 36 images was used and they achieved 90.91% accuracy. Although [12] and [14] have presented the methods for neovascularization, they focused a little bit on accurate extraction of blood vessels which is the core for detection of abnormal vessels.

Computer aided diagnostic system for PDR should be able to detect blood vessels accurately. A number of methods have been proposed for blood vessel detection and segmentation [15], [16], [17], [18], [19], but detection of neovascularization is still a difficult problem. Blood vessel extraction algorithms normally contain two parts, first is enhancement of blood vessels and second is the segmentation and classification of vessel pixels. Chaudhuri et al. [18] proposed a matched filter based method for blood vessel detection and it has been widely used for extraction purposes but it has been unable to find small blood vessels. Later, a threshold probing based technique was presented in [19] to improve the accuracy of matched filters. They analyzed the region based features of vessel structure. Mendonca et al. [20] used a first order derivative of Gaussian filter and a modified top hat operator for blood vessel enhancement and segmentation. Another probing algorithm using multithresholds was presented in [21].

The existing methods which we have discussed here with respect to neovascularization consist of different limitations such as (i) very little focus is given to accurate blood vessel segmentation, (ii) only few of them have extracted reliable features for detection of abnormal vessels, (iii) NVE and NVD are not detected collectively using a common method, and (iv) the classification stage is not given any proper focus and it is done by using existing SVM or neural network. In contrast to all these limitations, we propose a complete new system for PDR detection and grading which is very rare in literature. Main contributions of proposed system are representation of abnormal vessels with detailed feature set and the use of new dynamic modeling technique for detection of abnormal vessels using multivariate m-Mediods based classifier. We propose a new set of features to differentiate between normal and abnormal blood vessels and combine it with OD detection to grade the fundus image in different categories of PDR.

The rest of the paper is organized as follows: Section 2 describes the proposed system and its all phases in detail. The evaluation of proposed system using different retinal image databases and performance parameters is done in Section 3 followed by conclusion in last section.