Diabetic Retinopathy (DR) is the major cause of blindness caused by the damage to the blood vessels in the retina from diabetes. Exudates are fatty deposits on the retina which appear as yellowish regions in fundus image. To evaluate the new method, we examined 158 fundus images, including 84 abnormal images and 74 normal images.
Diabetic Retinopathy (DR) is the major cause of blindness caused by the damage to the blood vessels in the retina from diabetes. Exudates are fatty deposits on the retina which appear as yellowish regions in fundus image. To evaluate the new method, we examined 158 fundus images, including 84 abnormal images and 74 normal images.
Diabetic Retinopathy (DR) is the major cause of blindness caused by the damage to the blood vessels in the retina from diabetes. Exudates are fatty deposits on the retina which appear as yellowish regions in fundus image. To evaluate the new method, we examined 158 fundus images, including 84 abnormal images and 74 normal images.
Diabetic Retinopathy (DR) is the major cause of blindness caused by the damage to the blood vessels in the retina from diabetes. Exudates are fatty deposits on the retina which appear as yellowish regions in fundus image. To evaluate the new method, we examined 158 fundus images, including 84 abnormal images and 74 normal images.
Sharath Kumar P N , Rajesh Kumar R Health Informatics and Software Technology Group Centre for Development of Advanced Computing Thiruvananthapuram, India sharath_kumar_pn@yahoo.co.in Dr. Anuja Sathar , Dr. Sahasranamam V Department of Ophthalmology Regional Institute of Ophthalmology Thiruvananthapuram, India anujasathar@gmail.com
Abstract Diabetic Retinopathy (DR) is the major cause of blindness caused by the damage to the blood vessels in the retina from diabetes. It cannot be prevented but early detection through fundus imaging by an ophthalmologist can prevent further vision loss. Presence of microaneurysms, hemorrhages, cotton-wool spots and exudates are the symptoms of mild DR. Of these, the detection of exudates is one of the important factors in the early diagnosis of DR. Exudates are fatty deposits on the retina which appear as yellowish regions in fundus image. Fundus images show considerable variation in brightness which makes automatic detection of exudates difficult. In this study, we are proposing a new method for preprocessing and false positive elimination towards the reliable detection of exudates. The brightness of the fundus image was changed by the nonlinear curve with brightness values of the hue saturation value (HSV) space. To emphasize brighter yellow regions (exudates), gamma correction was performed on each red and green components of the image. Subsequently, the histograms of each red and green component were extended. After that, the exudates candidates were detected using histogram analysis. Finally, false positives were removed by using multi-channel histogram analysis. To evaluate the new method for the detection of exudates, we examined 158 fundus images, including 84 abnormal images with exudates and 74 normal images. The sensitivity and specificity for the detection of abnormal and normal cases were 88.45% and 95.5% respectively. Keywordsdiabetic retinopathy; fundus images; exudates; histogram analysis; preprocessing; false positive elimination. I. INTRODUCTION In India, there are approximately 31.7 million patients with diabetes and every year 1.7 million new patients with diabetes are recorded [1]. Approximately 5.6 million are thought to suffer from diabetic retinopathy (DR). This disease can be detected and prevented from developing into blindness if it is treated at an early stage. However, it has been recorded that people losing their vision following the onset of DR are increasing year by year. Retinal photographs obtained by the fundus camera are used to diagnose DR. Ophthalmologists examine the presence of hemorrhages, microaneurysms, and exudates in order to diagnose DR. By using fluorescein angiograms, ophthalmologist can detect changes in blood vessels. However, it is difficult to use fluorescein as a contrast medium for diagnosing all the medical examinees subjected to mass screening. In past, many studies have been reported on the use of fundus images in the detection of DR [27]. Hatanaka et al. proposed a CAD scheme for the detection of hemorrhages and exudates in ocular fundus images [2]. Niemeijer et al. proposed a method to describe and evaluate a machine learning-based automated system to detect exudates and cotton-wool spots in retinal photographs and differentiate them from drusen for early diagnosis of DR [3]. Sopharak et al. presented a method for detecting exudates using fuzzy C-means clustering [4]. Grisan et al. proposed a method for the segmentation of dark lesions in fundus images based on local thresholding and pixel density [5]. They reported a mean detection rate of 94% of the lesions present in an image based on a dataset of 60 annotated images. Yazid et al. proposed a method to sharpen the edge to simplify the segmentation process for cotton wool spots and exudates through ramp width reduction [6]. Xiaohui and Opas have presented a two way strategy method for the detection of lesions [7]. Based on different properties of bright lesions and dark lesions, bottom-up and top-down strategies are applied respectively to cope with the main difficulties in lesions detection such as inhomogeneous illumination. In bright lesion detection, a three-stage, bottom-up approach is applied. After local contrast enhancement preprocessing stage, two step Improved Fuzzy C-Means is applied in Luv color space to segment candidate bright lesion areas. Finally, a hierarchical SVM classification structure is applied to classify bright non- lesion areas, exudates and cotton wool spots. In the above methods, authors have also mentioned about two problems along with the detection of exudates. One is the wide variation in the brightness and image resolution other one is the high percentage of false positives (FPs). Fig. 1 (a) and (b) shows the typical fundus images without and with exudates respectively. In this study, we also address these by developing methods for fundus image normalization and FPs elimination.
Fig. 1. Typical fundus images. (a) Normal retina. (b) Retina having exudates. 2013 IEEE Recent Advances in Intelligent Computational Systems (RAICS) 978-1-4799-2178-2/13/$31.00 2013 IEEE 277 II. METHODS A. Overall Scheme Our overall detection scheme consists (1) image acquisition, (2) image normalizati of optic disc, (4) detection of exudates elimination of FPs using multi-channel hi Further details are described as follows.
Fig. 2. Flowchart for detecting exudates in fundus imag B. Image acquisition A dataset of 158 fundus images were fundus camera. 89 fundus images were obta size of 1,500 1,152 pixels and 24-bit color, were obtained with an array size of 2,144 24-bit color and 29 fundus images were obta size of 4,288 2,848 pixels with 24-bit images were annotated by an ophthalmolog 89 fundus images were taken from the D which was already annotated [8]. An example image is shown in Fig. 3 (a). Subsequently matrix was first standardized to 640 x 480 p detailed subsamples from the original imag processing efficiency. C. Image normalization Due to the flash light used in obtaining t there is an atypical change in the color of th To bring uniform change in the color of the f applied a scheme of brightness correction us value (HSV) space. First, the brightness va space were calculated. Second, the calculate was then smoothened using 9 x 9 median global correction was performed on the smoo value of HSV space and the finally ob correction value Bc(i, j) is given by the follow Bc(i, ]) = 1 - {I(i, ]) - 1 Image Acquisition Image Normalization Extraction of optic disc Detection of Exudates Elimination of false positives in optic disc Finally detected exudates s of five stages: ion, (3) extraction s candidates, (5) istogram analysis. ges. captured using a ined with an array , 40 fundus images 1,424 pixels with ained with an array color. 69 fundus gist and remaining Diaretdb1 database e of a color fundus y, the scale of the pixels by obtaining e data to improve the fundus images, he fundus images. fundus images, we sing hue saturation alues of the HSV d brightness value n filtering. Lastly, othened brightness btained brightness wing equation: 1] 2
Where V(i, j) is the smoo HSV space. Next, the red value R(i, j) value B(i, j) were changed b equally corrected from the ce fundus image (as shown in Fig. corrected by Bc(i, j) were then p The gamma value was experim Fig. 3 (c)). Finally, the histogra image were extended (as show color images were then conve selecting only the green comp which will have better contra fundus images were standardize
(a)
(c) Fig. 3. Color contrast enhancement. Brightness of fundus image was c brightness values of HSV space. (c adjusted in the dynamic range. (d) G green component. D. Detection of optic disc and The grayscale image obtain to detect optic disc and exuda exudates were then extracted where the threshold values corr disc were obtained. This is been Before the exudates candid understand the distribution of lesions viz. hemorrhages, exuda vessels and optic disc which oc study analysis, intensity profile hemorrhages being the darkest vessels which is not as dark exudates which appear more in color and finally the brightest intensity distribution of a fundu corresponding to each one of th hemorrhages and last peak bein 4 (a,d). This study knowledge while looking for the intensity r othened brightness value of the ), green value G(i, j), and blue by Bc(i, j). The brightness was enter to the skirt region of the 3 (b)). The color fundus images processed by gamma correction. mentally set to 0.5 (as shown in ams of each red, green and blue wn in Fig. 3 (c)). The corrected erted into grayscale images by ponent (as shown in Fig. 3 (d)) ast of the vascular details. The ed by using these processes.
(b)
(d) (a) Original color fundus image. (b) hanged by the nonlinear curve with c) Image after gamma correction and Grayscale image obtained by selecting exudates candidates ned in the previous step was used ates. Accordingly optic disc and d using the histogram analysis responding to exudates and optic n described as follows. dates are extracted, one should intensity values of the different ates and other objects viz. blood ccur in a fundus image. From our e of a fundus image starts with lesions, second being the blood k as hemorrhages, next is the number having yellowish bright being the optic disc. So, in the us image one can find four peaks he four objects with first peak for ng for optic disc as shown in Fig. is being utilized in our system range of exudates and optic disc. 278
(a) (b) (c) Fig. 4. Exudates detection approach. (a) Histogra (c) Portions of optic disc d (e) Optic disc candidate extraction from
( ( ( am from green channel. (b) Exudate candidates along w etected from green channel. (d) Histog red channel. (f) Localization of optic disc
(d)
(e)
(f) with optic disc candidate detection. gram from red channel. using binary filling operation. 279
(a)
(b) Fig. 5. Elimination of false positives by combining the exudates detection from green channel and optic disc extraction from red channel. (a) Binary image showing the finally detected exudates after eliminating false positives. (b) Detected Exudates outlined in the fundus image. The exudates candidates were extracted with the help of the threshold value chosen from the peaks in the histogram obtained from the green component (as shown in Fig. 4 (a)). As discussed earlier, we choose the third peak to extract exudates candidates. With the chosen value, simple binary thresholding is applied to the green component and the possible exudates candidates were detected. Detected exudates are outlined as shown is Fig. 4 (b). Similarly, optic disc localization was performed by thresholding the green component of the RGB color image. Here, we choose the fourth peak to extract optic disc and the threshold value corresponding to the fourth peak was obtained from the histogram of the green component as shown in Fig. 4 (a). Due to the blood vessels present inside the optic disc region, optic disc is not extracted in full but in small portions as shown in Fig. 4 (c). The method to overcome this is explained in the next stage. E. Elimination of false positives using multi-channel histogram analysis While extracting exudates using automatically obtained threshold value from histogram of the green component, few candidates in the optic disc region will also be detected as shown in Fig. 4 (b). In order to remove those falsely detected exudates candidates, we use multi-channel histogram analysis method. Firstly, we obtain a threshold value from the peaks corresponding to optic disc which is found from the histogram of the red component as shown in Fig. 4 (d). Using this threshold value, optic disc candidate is fully extracted as shown in Fig. 4 (e). This is done by thresholding the red component to the selected threshold value. The shape of the extracted optic disc region is then approximated to a circle followed by a binary filling operation. Here, approximation was done by considering the average size of the optic disc obtained from our image dataset. The fully extracted and outlined optic disc is as shown in Fig. 4 (f). The idea of using the red component is because optic disc being the brightest and dense object in the fundus images. Also, from the section II.D we saw that thresholding applied to the green component resulted in multi- object optic disc detection as shown in Fig. 4 (c) instead of single-object identification from the red component as shown in Fig. 4 (f). This is due to the blood vessels present in the optic disc. In nature, blood vessels are dark which cant be found from green component. But with the red component, optic disc can be found fully as the intensities of optic disc and blood vessels will be more saturated. While extracting we made sure that only optic disc candidates will be detected. This is done with the help of binary opening operation. Secondly, we make use of the exudates detected by thresholding the green component as obtained in section II.D. Finally, using the optic disc and exudates (including candidates in optic disc) pixel informations obtained from red and green component respectively, candidates in the optic disc which presented as false exudates are removed. This is done by binary filtering operation considering the fully extracted optic disc as a mask. Fig. 5 (a) shows the finally detected exudates from the fundus image and Fig. 5 (b) shows the outlining of finally detected exudates. III. RESULTS The contrast of exudates in the images was enhanced. The color of the processed image was standardized by the histogram normalization method. However, if the fundus image has some laser marks, the marks are changed brighter and more obscure than the original color pixel. A dataset of 158 fundus images having three different resolutions was used for the experiment. The images were obtained from two different sources and have sufficient variations in color, illumination and quality. These included 69 images acquired in a diagnostic examination at the Regional Institute of Ophthalmology (RIO), Thiruvananthapuram, India and 89 images taken from the Diaretdb1 database [8]. The images in the dataset were classified by an ophthalmologist as abnormal or normal image based on the 280 presence or absence of exudates respectively. Of the 69 images obtained from RIO, 22 were normal and 47 were abnormal. The Diaretdb1 database contained 37 normal and 52 abnormal images. Here, to evaluate our method of detecting exudates, we examined 158 fundus images. Exudates were detected in 74 out of 84 images with actual exudates and only 4 false positives were detected in the 74 images without exudates. By using our scheme, we succeeded in obtaining satisfactory results with a sensitivity of 88.45% (74/84) and specificity of 95.5% (70/74). Our method could remove as many FPs. However, our method could not detect approximately 10% of exudates. The main reason was that exudates that were not well focused became undetected. Table I summarises the results of our method for three data sets captured under three different resolutions. Table I also gives sensitivity (SN) and Specificity (SP) individually for the three data sets. Elimination of incorrectly detected exudates was done based on the method of finding the thresholds from the histograms of red and green channels. In Table II, the sensitivity and specificity results of our exudates detection algorithm are presented along with other published results from in the literature. Here, the comparison has been made with the results claimed by the respective authors and the image datasets used for their studies which are different from ours and from one another. Even though the result (Specificity) claimed by Sopharak et al [4] are higher than our proposed method; their method was validated for only 10 images.
TABLE I. ACCURACY DETAILS OF OUR SYSTEM
Data set Type of Image Diagnosed Images Correctly Identified Images
Accuracy (%) 1 (1500 X 1152) DR Normal 37 52 32 49 SN = 86.48 SP = 94.23 2 (2144 X 1424) DR Normal 27 13 24 12 SN = 88.88 SP = 92.3 3 (4288 X 2848) DR Normal 20 9 18 9 SN = 90 SP = 100
TABLE II. COMPARISON OF RESULTS FOR EXUDATES DETECTION Author No. of Images No. of Images with exudates SN (%) SP (%) Hatanaka et al. [2] 109 16 77 83 Niemeijer et al. [3] 300 105 95 86 Sopharak et al. [4] 10 DNP* 86.86 99.55 Our Method(Single channel) 158 84 88.45 91 Our Method(Mutli channel) 158 84 88.45 95.5 DNP* - data not provided in the paper IV. CONCLUSION In this study, a new method for automatic detection of exudates in the digitized fundus images using multi-channel histogram analysis is presented. This method can be applied to the computer-aided diagnosis (CAD) system for diagnosing eye diseases. The accuracy of the method was evaluated against the ground-truth annotated by the expert ophthalmologists. It was demonstrated that the algorithm can detect abnormalities related to the DR with high accuracy and reliability. The result of this work on exudates detection in the fundus images demonstrated that the efficiency and accuracy of the diagnosis of the DR can be considerably improved using computer aided detection methods. In the future, the integrated fundus image analysis scheme will be further improved and more clinical cases will be reported for evaluating its accuracy. The techniques employed in our study will help in improving diagnostic accuracy as well as in improving the workflow efficiency of the DR screening. Our next focus will be to develop a full fledged automated system to detect other symptoms of mild DR viz. microaneurysms, hemorrhages and cotton-wool spots which will help in assisting ophthalmologists. ACKNOWLEDGEMENT The Department of ophthalmology, Regional Institute of Ophthalmology (RIO), Thiruvananthapuram provided us with the required set of fundus images. The ophthalmologists and domain experts from RIO were consulted for gaining domain knowledge and for the verification and establishment of ground truth for the experiment. REFERENCES [1] Diabetic retinopathy: An Indian perspective, M. Rema & R. Pradeepa Madras Diabetes Research Foundation & Dr Mohans Diabetes Specialities Centre, Chennai, India, Indian J Med Res 125, March 2007, pp 297-310. [2] Y. Hatanaka, T. Nakagawa, Y. Hayash, A. Fujita, Y. Mizukusa, M. Kakogawa, K. Kawase, T. Hara, and H. Fujita, CAD scheme for detection of hemorrhages and exudates in ocular fundus images, in Proc. SPIE Medical Imaging 2007: Computer-aided Diagnosis, San Diego, 2007, vol. 6514, pp. 65142M-1-65142M-8. [3] Niemeijer M, van Ginneken B, Russell SR, Suttorp-Schulten MS, Abrmoff MD, Automated detection and differentiation of drusen, exudates, and cotton-wool spots in digital color fundus photographs for diabetic retinopathy diagnosis, Invest Ophthalmol Vis Sci. 2007 May;48(5):2260-7. [4] Sopharak A, Uyyanonvara B, Barman S, Automatic Exudate Detection from Non-dilated Diabetic Retinopathy Retinal Images Using Fuzzy C- means Clustering, Sensors. 2009; 9(3):2148-2161. [5] Grisan E, Ruggeri A, Segmentation of candidate dark lesions in fundus images based on local thresholding and pixel density, Conf Proc IEEE Eng Med Biol Soc. 2007;2007:6736-9. [6] Yazid, H. ; Arof, H. ; Mokhtar, N, Edge Sharpening for Diabetic Retinopathy Detection, IEEE Conference on Cybernetics and Intelligent Systems (CIS), pp. 41 44, June 2010. [7] Xiaohui Zhang ; Chutatape, O., Top-down and Bottom-up Strategies in Lesion Detection of Background Diabetic Retinopathy, IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), vol. 2, pp. 422-428, June 2005. [8] Standard Diabetic Retinopathy Database, School of Industrial Engineering and Management, Lappeenranta University of Technology, Finland. http://www2.it.lut.fi/project/imageret/diaretdb1/ 281