Automated Detection of Optic Disc from Digital Retinal Fundus Images for Screening Systems of Diabetic Retinopathy

GAO Weiwei MA XiaofengZUO Jing

1 College of Mechanical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China 2 Department of Ophthalmology,Jiangsu Province Hospital of TCM,Nanjing 210029,China

Abstract:Optic disc (OD) detection is a main step while developing automated screening systems for diabetic retinopathy. We present a method to automatically locate and extract the OD in digital retinal fundus images. Based on the property that main blood vessels gather in OD,the method starts with Otsu thresholding segmentation to obtain candidate regions of OD. Consequently,the main blood vessels which are segmented in H channel of color fundus images in Hue saturation value (HSV) space. Finally,a weighted vessels’ direction matched filter is proposed to roughly match the direction of the main blood vessels to get the OD center which is used to pick the true OD out from the candidate regions of OD. The proposed method was evaluated on a dataset containing 100 fundus images of both normal and diseased retinas and the accuracy reaches 98%. Furthermore,the average time cost in processing an image is 1.3 s. Results suggest that the approach is reliable,and can efficiently detect OD from fundus images.

Key words:fundus image;spectral characteristics analysis;optic disc(OD);blood vessels;Otsu thresholding;match filter;segmentation

Introduction

The optic nerve is one of the most important organs in the human retina.The central retinal artery and central retinal vein emanate through the optic nerve,supplying the retina with blood.Optic disc(OD) is the entrance for blood vessel of retina and optic nerve into eye ball.In diagnosis of eye illnesses,accurate location and extraction of OD are basic conditions for further diagnosis and detection due to reasons as listed:

(1) Relative position between OD and macula obeys a certain model[1].In addition,macula is the most sensitive area to light and affects eyesight significantly.

(2) OD is an important indicator for some eye diseases[2],such as glaucoma，blood and fovea.

(3) Diameter of OD [i.e.papillary diameter (PD)] can be considered as a reference for measuring distance and size of other fundus features[3].

(4) In diagnosis of hard exudates (a kind of early lesion for diabetic retinopathy),OD is a kind of false positive which should be deleted first[4].

OD is the area with the highest brightness in a normal fundus image because almost all light is reflected.Furthermore,shape of OD is similar to a circle (an ellipse in most cases and seldom a standard circle).Although OD has significant features,it is very difficult to locate and extract with effects of these two factors.

(1) Images with abnormal light condition.

(2) Effect of lesions in fundus.

In order to realize accurate extraction,shape and position of OD have to be considered.Researchers proposed 4 kinds of extraction methods.

(1) Assuming OD is a set of pixels with the highest brightness[5-6].Because the method only employs brightness information,its results are affected by bright changes of fundus images.The feature shows a large variance that makes simple detection methods brittle,particularly in the presence of retinal disease such as hard exudates or cotton wool spots.

(2) Matching the template[7-8].A template based on shape and brightness is designed to detect OD.For images with OD lesions,the method can not provide accurate results because lesions may cause shape and bright changes of OD.

(3) Employing a property that blood vessels finally gather in OD area[9].Firstly,all blood vessels should be extracted.Secondly,fuzzy vessel section is constructed after deleting all intersections.Finally,convergent diagram can be obtained by voting method and the point with the highest convergent strength is defined as the center of OD.This method may define an intersection of blood vessels as the OD center and the calculating process is too complex.

(4) Matching filter method[10].The centre of OD is defined as the pixel with the minimum cumulative scores obtained by 4 direction filters with different properties constructed by blood vessel direction in OD area.A defect of this method is that calculation cost is relatively high.

These 4 methods of locating OD were compared and research results showed that the first 3 methods were not robust and only the matched filter method was feasible[11].In the process of automatic diagnosis of diabetic retinopathy,extracting OD should be the first step because the color and bright similarity of OD and lesions is high.Therefore,high request for efficiency and reliability of OD extraction is virtually important.But accuracy for boundary extraction is not a factor of high request because OD is only a false positive to lesions of diabetic retinopathy,such as hard exudates and cotton wool spots.So proposing a method with both reliability and efficiency is necessary for automatic diagnosis of OD.

In section 1,the characteristics of the image database under study is presented.Section 2 describes the detection method.We test the performance of the proposed method and compare the method with other methods in section 3.In section 4,conclusions of this paper are made.

1 Materials

In this study,a total of 100 images with variable color,brightness and quality were used.The images were provided by the Department of Ophthalmology,Jiangsu province hospital of TCM and were obtained from a non-mydriatic retinal camera with a 45° field of view.There were the presence of retinal disease such as hard exudates or cotton wool spots,as shown in Fig.1.The image resolution was 3504×2336 at 24 bit RGB.In order to save calculating cost,the size of all images was transferred to 300×300 pixels in this study.And Configuration of the computer employed in this paper was Intel(R) Core(TM) Duo E7500 CPU and 6.00 GB RAM.

Fig.1 Difficulties in location and extraction of OD for (a) Lesions with high brightness and large areas in fundus and (b) OD with pathological changes

2 Methods

Based on the property that main blood vessels gather in OD,Otsu thresholding segment was employed to obtain candidate OD area.And then the OD area was extracted based on the intersection of main blood vessel.Advantages of this method are as follows.

(1) High pertinency and efficiency.Only main blood vessels with high contrast need to be extracted.Existed OD extraction methods are based on accurate extraction of blood vessels which is very difficult to realize.The method proposed is no need to consider capillaries,decreasing difficulties of realization and increasing efficiency.

(2) High reliability.Great robust can be obtained due to employing the property that main blood vessels gather in OD area which is shown in Fig.2.

The proposed method can be divided in two steps,which are explained in detail in the next subsections.The overall scheme of the method developed in this work is depicted in Fig.3.

Fig.2 Schematic drawing of the retinal vasculature orientations

Fig.3 Flowchart of location and extraction of OD

2.1 Spectral characteristics analysis of main physiological structure

Spectral characteristics of different anatomical structures are different,because retinal pigments have different absorption properties and different wavelengths of light in the fundus of the penetration performance are also different.There are three kinds of pigment including lutein,hemoglobin and melanin in fundus.These three pigments of absorption properties with different monochromatic light are different.

In order to obtain spectral characteristics of OD,retinal artery and vein,monochromatic photography was used to analyze the spectral characteristics of them.Remove the stimulating filter of the fundus camera lighting light.All the fundi were taken one photo under the white light firstly,and then focus shooting while inserting the peak wavelength of 417-648 nm in turn.The contrastCis calculated as

(1)

whered1andd2were optical density of physiological structure measured and retina neighbouring.Figure 4 shows the result of the monochromatic photography.It indicates that the visibility of OD is higher with wavelength of 478 nm or greater than other wavelengths of light,and there are two peaks at 530 nm and 628 nm,especially visibility reaches the highest with red light of 628 nm.The visibility of main vessel is higher in 478-589 nm,especially visibility reaches the best with the wavelength of 570 nm.

On this basis,appropriate RGB channels can be chosen for different detection targets according to the spectral characteristics analysis results.

Fig.4 Contrast of OD and blood vessel under different monochromatic lights

2.2 Candidate area of OD

OD reaches the highest visibility in 628 nm red light according to the result of spectral characteristics analysis above.In this wavelength,edge of OD is clear,visibility of blood vessels from OD is very poor with the almost disappearance of nerve fiber and OD presents a uniform reflected light spot.Figure 5 shows R,G and B channel of a color fundus image.In R channel,OD is the clearest but blood vessels are vague.Therefore,R channel (fR) is proper to OD segmentation.OD shows the highest brightness in R channel,but some lesions,for example,cotton wool spots in Fig.1 also show the highest brightness.Therefore,only according to the luminance information of OD to segment it was obviously unreliable,but it is reliable to get the candidate regions of OD which are areas characterized by high intensity and the certain size.Consequently,R channel of the original fundus image was coarsely segmented by improving Otsu thresholding to obtain the candidate regions of OD which was shown in Fig.6.

Fig.5 Channels of the original fundus image:(a) red channel;(b) green channel;(c) blue channel

Fig.6 Candidate regions of OD

2.3 Detection of main blood vessels and determination of its direction

In order to determinate the direction of main blood vessels,we must segment the main blood vessels firstly.H channel of color fundus images in HSV space subtract the H channel operated by a closing,and then the main blood vessels were segmented as shown in Fig.7.

In fundus image,the blood vessels show connected curve with different thicknesses,and each piece can approximate by a straight line.Consequently,the direction of main blood vessels can be detected by means of straight line detection.When the straight line is within the blood vessels,the number of the pixels with the value of 1 is the most,or otherwise.So the direction of blood vessels in the pixel can be taken the direction of the most number of the pixels with the value of 1.The total number of direction is 12 by means of every 15 degrees in one direction in this study.This method is simple,effective with a low computational complexity.

Fig.7 Main blood vessel

2.4 OD segmentation based on direction of main blood vessels

In order to detect object,the method of matched filter need to design a template,and then use the template to find out the object.When the template gets a response values in a certain area higher than the threshold,it is that to get the object in this area.Thus,in order to detect the OD,a vessel direction matched filter (Fig.8) is proposed to roughly match the direction of the vessels.Furthermore,the weighted approach was added to the matched filter to give the more higher weight to the direction of blood vessels in the area of OD.The weigh matrix was shown in Fig.9.Calculate the response value of weighted matched filter in the image of directions,and the pixel having the highest response value is selected as the OD center shown in Fig.10(a).Consequently,the OD [Fig.10(b)] can be picked out by thethe OD center from the candidate OD area shown in Fig.6.

Fig.8 Matched filter

Fig.9 Weight matrix

Fig.10 Results of OD:(a) location;(b) extraction

3 Results and Discussion

3.1 Experimental results

About 100 images with 300×300 pixels were tested by both methods proposed by the author and Hoover.In all tests,the window size for extracting blood vessels is 5×5 pixels and size of template is 179×93 pixels.Test results are shown in Table 1.Especially,the formula of accuracy is the number of correct judgement divided by the total number.The detection results of these fundus images processed by the proposed method by Hoover[9]are also shown in Table 1.The proposed method shows a bit lower accuracy than that in Ref.[9],but the efficiency improvement is very significant.In large amount of diabetic retinopathy,the proposed method can provide better performance.As one part of a mass screening programme of fundus pathology,the segmentation method of OD proposed in this paper is more better.Additionally,the experimental results shown in Fig.11 indicate that the proposed method overcomes the influence of inhoogeneous contrast and lesions to the segmentation of OD.

Table 1 Segmention result of OD

Fig.11 Results of OD in fundus appearing lesions with high brightness and large areas:(a) candidate regions;(b) main blood vessel;(c) location;(d) extraction

3.2 Discussion of experimental results

Although there are so many methods for OD extraction,a novel method is proposed with both reliability and efficiency to optimize the total process of diabetic retinopathy diagnosis.OD extraction,as the preliminary work for further diagnosis,is meaningless without efficiency,even with 100% accuracy.According to Retinal Photography Screening for Diabetic Eye Disease proposed by British Diabetic Association[12]in 1997,the method proposed provides much better performance.In addition,in their health policy model,Javittetal.[13]suggested that a sensitivity of 60% or greater maximized cost-effectiveness in screening for diabetic retinopathy.Increasing screening sensitivity from 60% to 100% provided little additional benefit due to the frequency of screening and the likelihood that retinopathy cases missed at one visit will be detected at the next.That means efficiency of automated detection is more important when detection accuracy is adequate.For DR patients,the retinal lesions will be influenced by treatment,control and progression of diabetes.Therefore,only higher detection frequency is guaranteed,related lesions of fundus can be founded in time.Higher detection frequency is guaranteed by efficient automatic detection algorithm.The efficiency in Table 1 of the proposed method is obviously superior to that in Ref.[9].So,the proposed method in this paper is useful to efficient EXs automatic detection algorithm.

4 Conclusions

The paper presented a simple but novel method for OD segmentation using weighted matched filter according to the brightness of OD and the positional relationship between it and the main blood.The proposed approach achieved better results compared to the result reported in the literature.Furthermore,the proposed approach is simple,easy to implement and robust.The results are encouraging and it is more suitable for the first step of hard exudates segmentation which constitutes automated screening of diabetic retinopathy.So further steps would be the detection of hard exudates and the evaluation of localization and distribution of the detected exudates in order to detect macular edema.