Automated Localization of Optic Disc in Colour Fundus Images

Transcription

1 World Applied Sciences Journal 8 (11): , 013 ISSN IDOSI Publications, 013 DOI: /idosi.wasj Automated Localization of Optic Disc in Colour Fundus Images Hidayat Ullah, Zahoor Jan, Rashid Jalal Qureshi and Bilal Shams 1 Department of Computer Science, Islamia College Peshawar (Chartered University), Peshawar, P.O. Box: 510, Khyber Pakhtunkhwa, Pakistan Shaheed Zulfiqar Ali Bhutto Institute of Science and Technology, Dubai, U.A.E. 3 Information Technology, Kohat University of Science & Technology, 6000, Kohat, Pakistan, P.O. Box: Abstract: The fundamental step for developing screening systems of diabetic retinopay is optic disc detection. The knowledge of optic disc location can be used to detect oer anatomical features of e retina such as macula and fovea. Mostly OD need to be eliminated for e identification of exudates, e primary signs of diabetic retinopay. In is work, a novel, fast and robust technique is presented for e automatic localization of OD in retinal fundus images. The proposed approach proceeded rough two stages: (1) conversation of RGB to HSI colour space followed by smooing and local contrast enhancement and () OD region s center detection using resholding and maematical morphology. Using clinician reference (ground tru), e proposed meod was validated and compared wi oer OD detection meods presented in e recent past. Proposed meod achieved comparatively higher speed and accuracy on retinal images of variable quality (normal and abnormal) selected from four different datasets. The superior performance of e technique suggested at it could be effectively used as a pre-processing step for computer aided mass screening of retinal diseases. Key words: Diabetic Retinopay (DR) Exudates Fovea Macula and Optic Disc INTRODUCTION Diabetic retinopay, eye diseases are e primary reason of vision failure and its cruelty can lead to blindness. Fundus image and its properties are shown e Figure 1. The brightest area of retinal image is optic disc and blood vessels initiate from optic disc [1]. An optic disc can also be found as e major area. In some cases like e presence of large exudates in an image, ere may be areas in e image which are bigger an e optic disc area. Therefore optic disc masking is most commonly used as a pre processing for e detection of exudates. An optic disc information be able to study cruelty of diseases like glaucoma etc. A new, fast and robust technique is presented for e automatic localization of e center of e OD in retinal fundus images. The proposed approach proceeds rough two stages. First, in preprocessing, each Red Green Blue Fig. 1: Fundus images of retina and its core anatomical characteristics (RGB) colour of retinal images is transformed into Hue Saturation and Intensity (HSI) space followed by smooing by means of median filter and local contrast enhancement. In e secondly stage e OD center is detected by using Otsu s resholding and maematical morphology operations like dilation and filling. Correspondence Auor: Dr. Zahoor Jan, Department of Computer Science, Islamia College Peshawar (Chartered University), Peshawar, P.O. Box: 510, Khyber Pakhtunkhwa, Pakistan. Tel: +9-(0) , Fax: +9-(0)

2 World Appl. Sci. J., 8 (11): , 013 The paper is structured as follows: Section, consist processed intensity image. For image segmentation have e review of latest proposed meods for optic disc Otsu s algorim is applied. The optic disc is en detection is presented, section 3 consist of e proposed localized by largest connected component of circular technique for optic disc localization, section 4 gives shape. experimental results and section 5 conclude e proposed Akara et al. [10] and [11], e preprocessing step are work. same as in [8] and [9]. The OD is detected using e Various studies have been conducted regarding entropy of e preprocessed intensity image. Otsu s Optic Disc localization Colour retinal fundus images for algorim is applied for segmentation. Then binary various datasets. Below is a brief list of such literature: dilation was applied so at all e neighboring pixels are Hussain et al. [] present a technique based on also included. However e entropy filtering requires more top-hat operator wi a radius of 5 pixels structuring time and memory. element was applied to e green channel image, en V. Vijaya et al. [1] extract e OD by using smooing image by median filter using window size of a propagation rough radii meod. In general x+rcos, (3 3). Finally applying hough transform to detect main y+rsin were used to represent a circle in discrete space. circular element. The lies from 0 to 360. First radiuses for all 360 degree S. Sekhar et al. [3] detect optic disk and fovea. By were computed and en e mean radius is found. using maematical morphological operator followed by Next midpoint of line joining circum points is found. hough transform an optic disk is detected. The fovea is The mean center is e midpoints of x and y co-ordinates. localized using its connection wi an optic disk. The circle of best fit is drawn wi mean centre as e S. Sekhar et al. [4] localized OD by using centre and mean radius as e radius. The optic disc morphological operations wi Hough transform. A center is found and propagation rough radii meod is circular area of concern is found by dividing e employed and e entire optic disc region is blackened observable area in e image using maematical and removed. However by using propagation rough morphological operation. Then optic disc is detected by radii meod e chance of detecting exudates in e using hough transform. proximity of e Optic Disc increases. Hussain et al. [5] and [6] use a morphological top-hat Huiqi et al. [13] defined candidate regions in e operation to e image. First ey converted RGB image retinal image by clustering e brightest pixels. A different into HSI colour en to intensity band Contrast Limited scaling factor of PCA meod is applied to candidate Adaptive Histogram Equalization (CLAHE) was applied. regions to find e minimum distance among e original An area resholding meod was applied to obtained input image and its projection onto e disk space. An binary image. To remove vessels and large peaks a optic disc center is located at e point wi e minimum closing operator followed by opening was applied. distance in e candidate regions among all e scaling Then eroded image is subtracted from dilated image. factors. Nevereless, detection is time consuming. Finally centre of e OD is detected using hough Aliaa et al. [14] presented matched filter to match e transform to e image. vessels direction at e OD area. By using D Gaussian S. Kavia et al. [7] transformed original RGB image matched filter e retinal vessels were segmented. to HSV colour Next median filtering is performed for Thus, a vessels direction map of e segmented retinal smooing. Finally CLAHE is used to enhance local vessels is obtained using segmentation algorim. contrast. The optic disc is eliminated by using e Hough The optic disc center is represented by inning and transform which is predominant to e detection of filtering vessels. The difference among e proposed circular features. The OD is masked by using a disc of matched filter resized into four different sizes and e intensity value similarly to e average intensity of an vessels directions at e surrounding area of each image. An area enclosed by e hough circle is set to zero candidates of e OD center is considered. to eliminate e OD. Nevereless e technique based on Akara et al. [15] first, transformed to HIS colour Hough transform does not give e best result as e space from each original RGB image and en smooing Hough space is responsive to e resolution. image by median filter and CLAHE was applied to improve Akara et al. [8] and [9], transformed each RGB image contrast. Next a closing operator was applied and e into e HSI colour and en noise was removed and resultant image was binarized. The reconstruction by image was enhanced by using CLAHE to e I band. dilation operation was applied en. The resultant image The OD is detected using e entropy of e pre- was binarized using otsu s algorim. The OD is en 1580

3 World Appl. Sci. J., 8 (11): , 013 localized by e largest connected component of circular shape. A binary dilation operation was applied to make sure at all e neighboring pixels are also included in candidate region. Nevereless e meods based on morphological reconstruction require more time and memory. MATERIALS AND METHODS Pre-Processing Phase Input Fundus Image RGB to HSI Apply Median filter of 3x3 on I band OD Detection Phase Convert back to RGB Binarization: Apply ing Otsu s on Green Band Area filtering size 150 pixel Apply Dilation The Figure depicted our proposed algorim which consists of two steps. First in preprocessing step we convert RGB image to HSI colour space as e intensity band is separately used for enhancement. Next remove noise and enhanced e contrast of intensity band. In OD detection step, we converted e enhanced HSI colour space to green channel of RGB as e OD is more visible in green band. Then Otsu s binarization and morphological operation is applied to detect e OD. This = () t () t () t () t () t Apply CLAHE on I band t 1 L 1 () t = p() i and () t = p() i 1 i= 0 i= t Apply morphological Filling Optic Disc detection Fig. : Block diagram for e proposed meod where, is simple and fast detection of OD in term of time and and memory. The various steps involve in e proposed approach are explain below: 1 (t) = The background pixel s variance (less en reshold) Pre-Processing: In pre-processing first e original (t) = The foreground pixel s variance (above retinal image s RGB space step is transformed into HSI reshold) (hue, saturation and intensity) colour space because we make use of e intensity component in e later Weights i are probabilities of two classes separated stages. by a i variances and reshold t of ese classes. To reduce noise we applied a (3x3) median filter, As OD is round, erefore its selection desires to be however medina filter result in blurring e image and made exact to e biggest one among e regions whose erefore, in order to reduce e amount of blurriness, we shapes are round. Circularity (C) of e shape of an area used only (3x3) filter instead of higher order filter. is defined by e value of compactness. C as defined in To improve image contrast, a CLAHE is applied to e e equation (). intensity colour space [19]. CLAHE divides e images into related regions and on each region histogram Area C = Perimeter equalization is applied. This evens out e allocation of () grey values into image and us creates hidden features where e area is number of pixels in e region and e more detectable. total no of pixels around e border of each region is e perimeter. Optic Disc Detection: In OD detection e resultant HSI To remove false detected area, we applied area components are transformed back to RGB to get e pre filtering. Area filtering is used for identification of e processed image from e green channel. target area on e basis of its properties. It was tested and To binarize an image, e Otsu's binarization analyzed at area smaller an 150 pixels are not OD area. algorim [0] is applied on green channel to split e To include boundary of an OD, a morphological compound regions from e smoo regions. The Otsu's dilation is applied. A disc-shaped structuring element (s) binarization algorim oroughly searches for e of radius 11 is used in dilation operator. An image I reshold at minimizes e intra class variance, distinct dilation wi structuring element s is specified by I s. as a weighted sum of variances of e two classes (e.g. foreground and background): (1) where I is e image after otsu s binarization and s is e structure element. The structuring element s is placed wi its origin at (x, y) and e new pixel value is determined using e rule: 1581

4 World Appl. Sci. J., 8 (11): , if s hits I g (x,y) = 0 oerwise Next filling is performed. Filling used to seek for regions enclosed by closed contour. Hence optic disc is detected in e retinal image. The optic disc s center is e severity of e binary image obtained from resholding. RESULTS AND DISCUSSION In is work, 40 retinal images are used, obtained from Al-Shifa Trust Eye Hospital Kohat, Pakistan and available online databases i.e. DIARETDB0 [16], DIARETDB1 [17], DRIVE [18]. From Al-Shifa Trust Eye Hospital 50 images are selected (10 are normal and 40 are abnormal). DIARETDB0, 80 images (0 normal and 60 contain signs of diabetic retinopay), from DIARETDB1, 70 images (5 are normal and 65 shows abnormality), from DRIVE, 40 images. The image size varied from 565 x 584 to 1500 x 115 while e FOV coverage varied between 35 and 50. All images were JPEG compressed. Table 1: Optic disc detection results for proposed algorim Optic Disc detector algorim Test databases Test Set Correctly detected False detected Accuracy DIARETDB % DIARETDB % DRIVE % Al-Shifa % Total % The proposed algorim is tested for detection of an optic disc s center on e databases: DIARETDB0, DIARETDB1, DRIVE and Al-Shifa Trust Eye Hospital Kohat, Pakistan. The proposed technique achieved a success rate of 98% as optic disc was detected correctly in 34 out of 40 images selected from e DIARETDB0, DIARETDB1, DRIVE and Al-Shifa Trust Eye Hospital Kohat dataset. This evaluation considers two ings, namely correct detection i.e. e number of images where optic disc area was correctly detected and secondly false detection i.e. Table : Comparison between e meods [7] and [8] and proposed meod Optic Disc detector algorims Test set databases Test Set Meod [7] OD detection Meod [8] OD detection Proposed OD detection Meod DIARETDB % 9% 96% DIARETDB % 94% 97% DIVE 40 97% 96% 100% Al-Shifa 50 89% 88% 98% Total 40 93% 9% 98% Fig. 3: (a) Input RGB image (b) Intensity image (c) Image after median filter (3x3) on I band (d) Image after applying CLAHE on I band (e) Preprocessed OD (f) OD detection using otsu s binarization (g) after area filtering (h) after dilation and (i) after morphological filling. (j) OD center (k) Abnormal input RGB image, (l) OD.Center. (m) OD and oer area detected, (n) after area filtering 158

5 World Appl. Sci. J., 8 (11): , 013 (m) (n) Fig. 4: (a) to (l) Optic Disc s center detection in retina images from DRIVE, DIARETDB0 and DIARETDB1 datasets while (m) and (n) shows retinal images wi wrongly detected optic disc s center in Al-Shifa Trust Eye images e number of images where non optic disc area was computational complexity by avoiding iterative detected as optic disc. In Table 1 e efficiency of e procedures and minimizing e parameters usually used. proposed technique is shown by comparing it wi its The proposed technique achieved maximum performance counterparts and Table shows comparison of meods and it was close to e manually detected optic disc s [7] and [8] wi e proposed technique on e selected center by a retinal expert. Therefore e result obtained by databases. applying proposed meod of optic disc center was more The percentage detection rate of proposed meod is perfect in similarity to oer algorims such as [7, 8]. relatively superior an oers. The results of each step of The proposed technique is just one option, ere is still e proposed meod are shown in Figure 3. opportunity for more improvement rough more research Figure 4 shows Optic disc center in original RGB keeping in view eir performance and computational normal and abnormal image in Al-Shafa Eye Trust needs. Hospital retina images and available databases like DIARETDB0, DIARETDB1 and DRIVE dataset. ACKNOWLEDGMENTS CONCLUSION We would like to ank Dr. Irfan Aslam, Medical Superintendant at Al-Shifa Trust Eye Hospital Kohat, A simple and off e shelf image processing Khyber Pakhtunkhwa, Pakistan for support, techniques for e localization of an optic disc s center is coordination and supplying e retinal images used in presented in e paper. The idea is to reduce e paper. 1583

6 World Appl. Sci. J., 8 (11): , 013 REFERENCES 11. Akara Sopharak, Bunyarit Uyyanonvara and Sarah Barman, 009. Automatic Exudate Detection for 1. Kumari, V.V. and N. Suriyanarayanan, 008. Blood diabetic retinopay screeing. Research Article, vessel extraction using wiener filter and Science Asia, 35: morphological operation. International Journal of 1. Vijaya Kumari V. and N. Suriya Narayanan, 010. Computer Science and Emerging Technol., 1(4): Diabetic Retinopay- Early Detection Using Image. Hussain F. Jaafar, Asoke K. Nandi and Waleed Al Processing Techniques. International Journal on Nuaimy, 010. Automated Detection of Exudates Computer Science and Engineering, 0(): In Retinal Images using a Split-and-Merge Algorim. 13. Huiqi Li J. and Opus Chutatape, 001. Automatic In Proceedings of European Signal location of Optic Disc in Retinal Images. In e Processing Conference, pp: Proceedings of IEEE- ICIP, pp: Sekhar, S., W. Al-Nuaimy and A.K. Nandi, Aliaa Abdel-Haleim Abdel-Razik Youssif, Atef Zak- Automated Localisation of Optic Disc and Fovea in i Ghalwash and Amr Ahmed Sabry Abdel-Rahman Retinal Fundus images. In e Proceeding of 6 Ghoneim, 008. Optic Disc Detection from Normalized European Signal Processing Conference, Digital Fundus Images by Means of a Vessels 4. Sekhar, S., W. Al-Nuaimy and A.K. Nandi, 008. Direction Matched Filter. In e Proceedings of e Automated Localisation of Retinal Optic Disc using IEEE Transactions On Medical EUSIPCO, Hough Transform. In e Proceeding of 5 IEEE Switzerland, 7(1): International Symposium on Biomedical Imaging: 15. Akara Sopharaka, Bunyarit Uyyanonvaraa, Sarah From Nano to Macro, pp: Barmanb, Thomas H.Williamsonc, 008. Automatic 5. Hussain F. Jaafar, Asoke K. Nandi and Waleed Detection of Diabetic Retinopay Exudates from Al- Nuaimy, 011. Automated Detection and Grading Non-Dilated Retinal Images using Maematical of Hard Exudates from Retinal Fundus Images. In e Morphology Meods. Computerized Medical proceedings of 19 European Signal Processing Imaging and Graphics, 3: Conference, pp: Kauppi, T., V. Kalesnykiene, J.K. Kamarainen, 6. Hussain F. Jaafar, Asoke K. Nandi and Waleed L. Lensu, I. Sorri, H. Uusitalo, H. Kalviainen and Al- Nuaimy, 011. Detection of Exudates from Digital J. Pietila, 006. Diaretdb0: Evaluation database and Fundus Images Using a Region Based Segmentation meodology for diabetic retinopay Algorims. Technique. In e proceedings of 19 European Technical report, Lappeenranta University of Signal Processing Conference, pp: Technology, Lappeenranta. 7. Kavia, S., K. Duraiswamy and A.R.S. Sri Supreea, 17. Kauppi, T., V. Kalesnykiene, J.K. Kamarainen, 01. Detection of Exudates and Macula in Fundus L. Lensu, I. Sorri, A. Raninen, R. Voutilainen, Images to Estimate Severity of Diabetic Retinopay. H. Uusitalo, H. Kalviainen and J. Pietila, 007. International Journal of Communications and DIARETDB1 diabetic retinopay database and Engineering, 7(7): 4-9. evaluation protocol. Technical report, Faculty of 8. Akara Sopharak, Khine Thet Nwe, Yin Aye Moe, Medicine, University of Kuopio. Matew N. Dailey and Bunyarit Uyyanonvara, Staal, J., M.D. Abramoff, Meindert Niemeijer, Max A. Automatic Exudate Detection wi a Naive Bayes Viergever and Bram van Ginneken, 004. Ridge Based Classifier. In e proceedings of International Vessel Segmentation in Colour Images of e Retina. Conference on Embedded Systems and Intelligent In e Proceedings of e IEEE Transaction on Technology, pp: Medical Imaging, 3: Akara Sopharak, Bunyarit Uyyanonvara, Sarah 19. Laszlo Kovacs, Rashid Jalal Qureshi, Brigitta Nagy, Barman, Sakchai Vongkittirux and Andnattapol Balazs Harangi andras Hajdu, 010. Graph Based wongkamchang, 010. Fine Exudate Detection using Detection of Optic Disc and Fovea in Retinal Images. Morphological Reconstruction Enhancement, In e proceedings of Soft Computing Applications International Journal of Applied Biomedical (SOFA), 4 International Workshop, pp: Engineering, 10(1): Otsu, A reshold selection meod from 10. Akara Sopharak, Bunyarit Uyyanonvara and graylevel Histogram. In e Proceedings of e IEEE Sarah Barman, 009. Automatic Exudate Detection Transactions on Systems, Man and Cybernetics, from Non-dilated Diabetic Retinopay Retinal 9: Images Using Fuzzy C-means Clustering. Sensors, pp: