Detection of Hard Exudates from Diabetic Retinopathy Images using Fuzzy Logic R.H.N.G. Ranamuka and R.G.N. Meegama

Transcription

1 Detection of Hard Exudates from Diabetic Retinopathy Images using Fuzzy Logic R.H.N.G. Ranamuka and R.G.N. Meegama Abstract Diabetic retinopathy, that affects the blood vessels of the retina, is considered to be the most serious complication prevalent among diabetic patients. If detect successfully at an early stage, ophthalmologist would be able to treat the patients by advanced laser treatment to prevent total blindness. In this paper, we propose a technique based on morphological image processing and fuzzy logic to detect hard exudates from diabetic retinopathy retinal images. At the initial stage, the exudates are identified using mathematical morphology that includes elimination of the optic disc. Subsequently, hard exudates are extracted using an adaptive fuzzy logic algorithm that uses values in the RGB colour space of retinal image to form fuzzy sets and membership functions. The fuzzy output for all the pixels in every exudate is calculated for a given input set corresponding to red, green and blue channels of a pixel in an exudate. This fuzzy output is computed for hard exudates according to the proportion of the area of the hard exudates. By comparing the results with hand-drawn ground truths, we obtained sensitivity and specificity of detecting hard exudates as 75.43% and 99.99%, respectively. Keywords Diabetic retinopathy, hard exudates, retinal images I. INTRODUCTION There are three major complications of diabetes which lead to blindness. They are retinopathy, cataracts, and glaucoma among which diabetic retinopathy is considered as the most serious complication affecting the blood vessels in the retina. Diabetic retinopathy (DR) occurs when tiny vessels swell and leak fluid or abnormal new blood vessels grow hampering normal vision. The early stage of diabetic retinopathy is referred to as microaneurysm which appears as tiny, dark red spots or miniscule hemorrhages forming clusters with circular shape. The size varies from microns and it approximates to 1/12th diameter of an average optic disc. Hemorrhages, that appear inside deeper layers of the retina, form a round or flame shape. However, when they appear in large numbers, such a feature is considered as non-proliferative retinopathy. Cotton wool spots are yellowish white, fluffy lesions in the nerve fiber layer and are also called soft exudates. These spots are created as a result of swelling of the nerve fiber axons. While soft exudates are very common in diabetic retinopathy, hard exudates are typically bright, reflective and not common in diabetic retinopathy. They appear as white or cream colored lesions on the retina with different sizes. Each hard exudate consists of blood plasma and lipids leaked from blood vessels. The aim of this research is to develop a system to detect hard exudates in diabetic retinopathy using non-dilated diabetic retinopathy images. The exudates are identified using morphological methods and categorized into hard exudates and non-hard exudates using an adaptive fuzzy algorithm. Detection and treatment of diabetic retinopathy at an early stage help prevent total blindness. Therefore, early detection of diabetic retinopathy is very important because ophthalmologist would then be able to treat the patients by advanced laser treatment. R.H.N.G Ranamuka and R.G.N. Meegama are with the Department of Statistics and Computer Science, Faculty of Applied Science, University of Sri Jayewardenepura, Sri Lanka; ngranamuka@gmail.com, rgn@sjp.ac.lk 34 II. RECENT WORK Salvatelli et al. [1] have analyzed and tested several correction techniques where the aim was to establish a qualitative assessment of the adequacy of the different methods for preprocessing stages in a DR diagnosis system. They have obtained best results for the color model using RGB for image analysis and HSI for actual processing of images. To address non-uniform illuminations, they have used morphology and local enhancement techniques. The results have been obtained by homomorphic filtering for luminance correction, together with morphologic filtering for contrast enhancement. The processing stages were tested using Fuzzy C-Means and local Hurst (self correlation) coefficient for unsupervised segmentation of the abnormal blood vessels. Iqbal et.al [2] have presented a method to identify diabetic retinopathy using digital signal processing and image processing techniques. They have employed color space conversion, zero padding of image edges, median filtering and histogram equalization with overlap mean for the image preprocessing stage. A method to enhance exudates using fuzzy morphology is proposed in [3] where, a color fundus image is converted to grey scale image first and followed by a fuzzy morphological closing operation to enhance boundaries of exudates. At the final stage, the resulting image is added to the original image to obtain the enhanced image. These enhanced images provide good results for clinical examinations. Sopharak et al. [4] have employed morphological operators for exudates detection on diabetic retinopathy patients nondilated pupil and low-contrast images. They have used this technique due to its speed and requiring low computing power making it ideal to be used in a rural setup where access to power is difficult. They have followed certain steps for optic disc detection prior to identification of exudates. After the optic disc is eliminated, mathematical morphology is used to detect the exudates. Because the location of exudates is important with respect to the macula, they have assumed the macula as the darkest region in retinal image. Sixty images were tested in an AMD Athlon 1.25 GHz personal computer and it had taken approximately 3 minutes to process an entire image including detection of the optic disc. They have managed to detect exudates with sensitivity of 80% and specificity of 99.5%.

2 Akara et al. [5] have proposed a Fuzzy C-means (FCM) clustering method to detect exudates. Contrast enhancement is applied before four image-based features, namely, intensity, standard deviation on intensity, hue and a number of edge pixels, are provide as input parameters to a coarse segmentation routine using FCM clustering method. Before exudates are detected, the optic disc is identified using an entropy feature. They have applied a FCM clustering algorithm for the segmentation with subsequent morphological reconstruction to obtain better segmentation results. The difference image is thresholded and reconstructed to obtain the final result. Expert ophthalmologists handdrawn ground-truths were then compared with the output for validation. They have used sensitivity, specificity, positive predictive value (PPV), positive likelihood ratio (PLR) and accuracy to evaluate the overall performance of the algorithm and have reported sensitivity, specificity, PPV, PLR and accuracy to be 87.26%, 99.24%, 42.77%, and 99.11%, respectively. Kumari and Narayanan [6] have presented a method using feature extraction, template matching and enhanced MDD classifiers to detect exudates. It has been tested for fifty images where the location of the optic disc was identified accurately for 39 images. A machine learning approach to detect exudates from nondilated retinal images is presented in [7] where a series of experiments were conducted to select specific features such as pixel intensity after preprocessing, the standard deviation of the preprocessed intensities in a window around the pixel, the pixel hue, the number of edge pixels in a window around the pixel, the ratio between the size of the pixel s intensity cluster and the optic disc. A Bayes classifier is used to select features and subsequent classification. Walter et al. [8] have presented a method to detect exudates using variations in high grey levels. Contours of exudates were identified using morphological reconstruction techniques while morphological filtering and watershed transformations were used to identify the optic disc. The mean sensitivity and the mean predictive value were reported to be 92.8% and 92.4%, respectively. A method to detect exudates automatically from color fundus retinal image using color histogram thresholding is presented in [9]. The preprocessing of color fundus retinal images is done using color space conversion and fundus region detection. The RGB space is transformed to CIELab color space and then binarization and mathematical morphology are employed to detect fundus regions. Non linear diffusion segmentation is performed to determine the variations in exudates and lesion boundary criteria pixels. Because the optic disc and exudates exhibit similar brightness, the optic disc is detected and localized with the aid of region based segmentation and color histogram. The optic disc is identified using the maximum pixel value of the color histogram. The exudates are detected by thresholding the color histogram to classify pixels belonging to hard and soft exudates. Fig. 1. Main stages of the proposed algorithm Basha and Prasad [10] have proposed an approach using fuzzy logic for hard exudates detection from digital fundus images. Initially, morphological segmentation is used to detect abnormal regions in fundus images. Fuzzy sets for different values in the color space are used to form fuzzy rules. The output is calculated as the average of the values in the different color space. III. METHODOLOGY The algorithm proposed in this paper detects hard exudates in diabetic retinopathy using principles of mathematical morphology and fuzzy logic in a pipeline of routines as shown in Fig. 1. At the initial stage, the exudates are identified using mathematical morphology. This stage can be divided into three sub-stages, namely, pre-processing, optic disc elimination and exudates detection. Pre-processing involves color space conversion, noise removal, contrast enhancement and Gaussian filtering. The color fundus image is converted to HSI image and the I-band of the fundus image is used for further processing [4, 5]. In order to eliminate the optic disc, it is assumed that the component with the largest circular shape of a fundus image is the optic disc [7, 9]. The second stage involves classification of exudates as hard exudates using fuzzy logic. Fuzzy rules are formed using fuzzy sets derived from the RGB fundus image. These fuzzy rules are used to decide the presence of hard exudates in diabetic retinopathy images. 35

3 . Fig 2. Output images of optic disc elimination stage: (a) applying morphological closing operator, (b) thresholded image using Nilblack s method, (c) thresholded image using percentile method, (d) largest circular connected component, (e) inverted binary image with largest circular component and (f) optic disc is eliminated from the pre-processed image. A. Optic disc elimination At first, the closing operator with a flat disk shape structuring element is applied to the pre-processed image. The resultant image is binarized using a thresholding technique. The new binary image Ω consists of all the connected regions C i such that C, C C 0, i, j m, i j (1) k m k i j where m = 1,2,,k is the number of connected components. Other than the background, the component having the largest number of pixels with circular shape among C i contains the optic disc and therefore, extracting this component separates the optic disc from all other structures in the retinal image. If the largest connected component is R i, the compactness C of R i is measured by A( Ri ) C( Ri ) 4 (2) 2 P ( R ) i where A(R i ) is the number of pixels in i th region and P i is the number of the pixels around i th region. The compactness is measured by applying two thresholding techniques, namely, the P-tile method [11] and Nilblack s method [12,] separately. The optic disc can be considered as the largest connected component that provides high values of compactness among these two methods. In our algorithm, the optic disc is eliminated before detecting exudates as the optic disc and exudates contain similar color and intensity. Figure 2(a) shows the image after applying morphological closing operator with a flat disc to eliminate high contrast blood vessels. Then, the optic disc could be identified as the largest circular connected component in the fundus image. A weight of 1.3 is used in the Nilblack s method for Thresholding as in Fig. 2(b). The region containing the optic disc is brighter than other regions in the retinal images. It was discovered that the optic disc occupies approximately 2% of bright regions in fundus images while the rest being the background. This percentage is used to perform the percentile method to obtain the binary image as in Fig. 2(c). The largest connected component which provides a value of high compactness among these two methods is considered as the optic disc. In this case, as seen in Fig. 2(d), the binary image obtained after applying Nilblack s method provides a high compactness for the largest circular component. B. Detection of exudates The next step in our strategy is to identify exudates from the image from which the optic disc was eliminated. The morphological closing operator with a flat disc shape structuring element is applied to this image to remove blood vessels as both exudates and blood vessels exhibit high contrast. The standard deviation of the resultant image f 2 is calculated using the following equation where f 1 represents the image obtained by applying the closing operator. 1 2 f2 ( x) ( f1( i) f ( x)) (3) 1 N 1 i W ( x) where x is a set of all pixels in a sub-window W(x), N is the number of pixels in W(x) and, μ f1 (x)is the mean of f 1( i).the triangle method is used to obtain the thresholded image after enhancing the local contrast of image f 2.This enables us to extract every minute bright region together with 36

4 Fig. 3. Membership functions of linguistic variables (a) x r, (b) x g, (c) x b and (d) X out. Fig 4: Output images of exudates detection stage: (a) applying morphological closing operator, (b) standard deviation of the image, (c) thresholded image using triangle method, (d) removal of unwanted borders, (e) flood filling of holes, (f) marker image, (g) morphological reconstruction, (h) thresholded image and (i) superimposing on original image borders of large bright regions. As the detection of exudates are confounded by borders of both the optic disc and the image, we subtract the dilated optic disc region, which is detected previously, from the thresholded image after removing the image border. Then, morphological closing and dilation operator is applied to obtain closely distributed exudates. At the next step, flood filling is carried out on all holes in the resultant image in order to create a marker image for morphological reconstruction. During morphological reconstruction, the peaks in the marker image dilate until the contour of the marker image fits under the mask image. The difference image between the 37 resulting image of the previous step and the intensity band of the original image is taken for thresholding. The output of this thresholded image is super-imposed on the original RGB image to extract the exudates. The output images at each stage of these image processing routines are shown in Fig. 4. C. Classification of hard exudates using Fuzzy Logic The final stage in our proposed technique is to identify the exudates as hard exudates using fuzzy logic. We use values in the RGB color space of retinal image to form the fuzzy set and membership functions. It uses the red, green and blue

5 value of a pixel as three input values (x r, x g, x b ) for the fuzzy inference system giving a single output. In order to calculate the output of given x r, x g and x b for a specific rule, the fuzzy inference system provides the degree of membership to the output variable X out as shown in Fig. 3. A de-fuzzification function, based on the centroid method, is used to compute the final output for the identification of hard exudates. The method presented in this paper determines the fuzzy output for a given input set (x r, x g, x b ) corresponding to red, green and blue channels of a pixel in an exudate using fuzzy logic. These fuzzy outputs are evaluated for all the pixels in every exudate in the retinal image. A region is considered to be a hard exudate if the average fuzzy value is greater than This crisp logic is created according to the membership function of linguistic variables of where an exudate is at least a weak hard exudate if the output crisp value of an exudate becomes greater than Subsequently, the fuzzy output is computed for hard exudates according to the proportion of the area of the hard exudates. Given below are the fuzzy rules that we apply to detect hard exudates. In these fuzzy rules, & and denote the AND and OR operators, respectively. 1.if x r ϵ R1 x g ϵ G1 x b ϵ B4 then X out is nothardexudate 2.if x r ϵ R2 & x g ϵ G2 & x b ϵ B1 then X out is weakhardexudate 3.if x r ϵ R2 & x g G2 & x b B1 then X out is nothardexudate 4.if x r ϵ R3 & x g ϵ G3 & x b ϵ (B1 B2) then X out is weakhardexudate 5.if x r ϵ R3 & x g ϵ G3 & x b ϵ B3 then X out is nothardexudate 6.if x r ϵ R3 & x g G3 then X out is nothardexudate 7.if x r ϵ R4 & x g ϵ G3 & x b ϵ B1 then X out is mediumhardexudate 8.if x r ϵ R4 & x g ϵ G3 & x b ϵ B2 then X out is weakhardexudate 9.if x r ϵ R4 & x g G3 then X out is nothardexudate 10.if x r ϵ R5 & x g ϵ (G2 G3 G4) then X out is nothardexudate 11.if x r ϵ R5 & x g ϵ G5 & x b ϵ (B1 B2) then X out is HardExudate 12.if x r ϵ R5 & x g ϵ (G6 G7) then X out is nothardexudate 13.if x r ϵ R6 & x b ϵ B3 then X out is nothardexudate 14.if x r ϵ R6 & x g ϵ (G2 G3) then X out is nothardexudate 15.if x r ϵ R6 & x g ϵ G4 & x b ϵ (B1 B2) then X out is HardExudate 16.if x r ϵ R6 & x g ϵ G5 & x b ϵ (B1 B2) then X out is HardExudate 17.if x r ϵ R6 & x g ϵ G6 & x b ϵ (B1 B2) then X out is HardExudate 18.if x r ϵ R6 & x g ϵ G7 then X out is nothardexudate 19.if x r ϵ R6 & x b ϵ B3 then X out is nothardexudate 20.if x r ϵ R7 & x g ϵ G6 & x b ϵ (B1 B2) then X out is severehardexudate 21.if x r ϵ R7 & x g ϵ G5 & x b ϵ (B1 B2) then X out is nothardexudate 22.if x r ϵ R7 & x g ϵ (G2 G3 G4) is nothardexudate IV. RESULTS AND DISCUSSION For testing the proposed algorithm, we have chosen twenty images from the publicly available diabetic retinopathy dataset DIARETDB0 and DIARETDB1 [13]. These images were taken from Kuopio University hospital and captured with few 50 degree field-of-view digital fundus cameras. We have selected the images with a size of 1500x1152 pixels to test the proposed technique using Matlab version The identified exudates are classified as hard exudates using fuzzy rules that were mentioned previously. In this approach, we first classify each exudate as a hard exudates by assigning a crisp value for each hard exudates. By using this crisp value, we compute a value for all the exudates according to the proportion of its area. Figure 5(a) shows the fundus image which has soft exudates and a very tiny hard exudate. The algorithm identified the soft exudates and showed that hard exudates do not exist in this image. The fundus image with weak hard exudates is shown in Figure 5(b). The algorithm has detected that approximately 42% of diabetic 38 retinopathy hard exudates in this image. Strong hard exudates are depicted in the Fig. 5(c) with 89% hard exudates. We selected sensitivity and specificity, two widely used parameters used in research literature, to test the performance of the proposed technique. These measures are calculated using four parameters, namely, the true positive (TP) rate (the number of hard exudates pixels correctly detected), the false positive (FP) rate (the number of non-hard exudates pixels wrongly detected as hard exudates pixels), the false negative (FN) rate (the number of hard exudates pixels not detected) and the true negative (TN) rate (the number of non-hard exudates pixels correctly identified as non-hard exudates pixels) as follows: Sensitivity TP ( TP FN ) (4) Specificit y TN (5) TN FP As seen in Table 1, we managed to obtain an average sensitivity of 75.43% and an average specificity of 99.99%. In a majority of the images, the algorithm resulted in a specificity of 100%. V. CONCLUSION This research proposes a novel technique to identify exudates using morphological methods and categorize these exudates into hard and non-hard exudates using principle of fuzzy logic. The strength of this approach is the ability to determine whether each exudate is hard exudates or not, separately. We have used the intensity band of the HSI image at this stage. As fundus image generally contain a high amount of noise, different pre-processing techniques can be applied for noise suppression and enhancing features to equalize regions showing uneven contrast. Fig 5: Classification as a percentage of all exudates as hard exudates (a) 0%, TABLE (b) 42% 1 and (c) 89%. QUANTITATIVE PERFORMANCE OF THE PROPOSED METHOD FOR DETECTION OF HARD EXUDATES. Image Sensitivity % Specificity %

6 Image0 1 Image0 2 Image0 3 Image0 4 Image0 5 Image0 6 Image0 7 Image0 8 Image0 9 Image1 0 Image1 1 Image1 2 Image1 3 Image1 4 Image1 5 Image1 6 Image1 7 Image1 8 Image1 9 Image2 0 Averag e REFERENCES [1] A. Salvatelli, G. Bizai, G. Barbosa, B. Drozdowicz and C. Delrieux, A comparative analysis of pre-processing techniques in color retinal images, Journal of Physics, Conference series, vol. 90, [2] M.I. Iqbal, N.S Gubbal, A.M. Aibinu and A Khan, Automatic diagnosis of diabetic retinopathy using fundus images, Masters Thesis, Blekinge Institute of Technology, October, [3] A.B. Mansoor, Z. Khan, A. Khan and S.A. Khan, Enhancement of Exudates for the Diagnosis of Diabetic Retinopathy using Fuzzy Morphology, In Proc 12 th IEE International Multi-topic Conference (INMIC), [4] A. Sopharak, B. Uyyanonvara, S. Barman and T.H. Williamson, Automatic detection of diabetic retinopathy exudates from non-dilated retinal images using mathematical morphology methods, Computerized Medical Imaging and Graphic, vol. 32, no.8, pp , [5] A. Sopharak, B. Uyyanonvara and S. Barman, Automatic Exudate Detection from Non-dilated Diabetic Retinopathy Retinal Images Using Fuzzy C-means Clustering, Journal of Sensors, vol. 9, no. 3, pp , [6] V.V. Kumari and N.S. Narayanan, Diabetic Retinopathy Early Detection using Image Processing Techniques, International Journal on Computer Science and Engineering, vol. 02, no. 02, pp , [7] A. Sopharak, K.T. New, Y.A. Moe et al., Automatic Exudates Detection with a Naive Bayes Classifier, In Pro. of the 2008 International Conference on Embedded Systems and Intelligent Technology, pp , [8] T. Walter, J.-C. Klein, P. Massin and A. Erginay, A Contribution of Image Processing to the Diagnosis of Exudates in Color Fundus Images of the Human Retina, IEEE Transactions on Medical Imaging, vol. 21, no. 10, [9] S. Kavitha and K. Duraiswamy, Automatic Detection of Hard and Soft Exudates in Fundus Images using Color Histogram Thresholding, European Journal of Scientific Research, vol. 48, no. 3, pp , [10] S.S. Basha and K.S. Prasad, Automatic Detection of Hard Exudates in Diabetic Retinopathy using Morphological Segmentation and Fuzzy Logic, International Journal of Computer Science and Network Security, vol. 8, no. 12, [11] M. Taghizadeh and M.R. Mahzoun,. Bidirectional Image Thresholding Algorithm using Combined Edge Detection and P-Tile algorithms, The Journal of Mathematics and Computer Science, vol. 2, no. 2, pp , [12] R.N.B. Rais, M.S. Anif and I.A. Taj, Adaptive Thresholding Technique for Document Image Analysis, 8 th IEEE International Multi-topic Conference (INMIC), Lahore, Pakistan, December, [13] [www] 39