Department of Instrumentation Technology, RVCE, Bengaluru, India

Transcription

1 A Survey on Microaneurysm Detection for Early Diagnosis of Diabetic Retinopathy Srilatha L.Rao 1, Deepashree Devaraj 2, Dr.S.C. Prasanna Kumar 3 1, 2, 3 Department of Instrumentation Technology, RVCE, Bengaluru, India Abstract: Diabetic Retinopathy is a major medical problem that causes damage to the eye. A need arises to detect it at an early stage. Currently, the trained eye care specialists are not able to screen the exponential increase in the number of Diabetic Retinopathy patients. An automated Diabetic Retinopathy screening system will enable the detection of lesions accurately, thus helping the ophthalmologists. Few Retinal Lesions include Micro aneurysm, Soft Exudates, Hard Exudates, and Retinal Haemorrhages. Microaneurysms are the earliest clinical signs of Diabetic Retinopathy. They are reddish in color and appear as small red spots on the retinal fundus images. Early detection of microaneurysm can help in the early treatment of Diabetic Retinopathy. In this paper, we mainly review and analyse the various methodologies in the detection of microaneurysm from the diabetic retinopathy images. Index Terms: Diabetic Retinopathy, Fundus Image, Image processing, Microaneurysm. I. INTRODUCTION Diabetes mellitus is one of the main causes of irreversible blindness worldwide which may cause the blood vessels to leak blood onto the retina eventually leading to loss of vision. According to WHO estimates, the total number of people affected with diabetes is projected to rise from 171 million in 2000 to 366 million in Diabetic related eye disease like Diabetic Retinopathy (DR) has been recognized as the major cause of Blindness.The patients do not notice the damage caused to the vision until it becomes severe. The damage caused by DR, if treated in early stages, can be prevented. An automated tool would be beneficial in the early detection of DR, assisting the ophthalmologists in treating the disease efficiently. Fig 1: Normal Retinal Fundus Image [6] The presence of micro aneurysm (MA), exudates, haemorrhages may indicate the degree of DR.MA appear as red dots. They are the focal dilations of the retinal capillaries. Exudates appear as yellowish color. They are leakages of proteins or lipids from blood vessel. Haemorrhages appear as round small red dots or blots. They occur when blood leaks from the retinal vessels. MA continues to get worse with the advancement of the disease. They are referred as non-proliferative stage, and if not treated immediately, they may transform into proliferative stage, causing retinal detachment. There are four stages used for grading DR, grade 0 for no DR, grade 1 for mild DR, grade 2 for moderate DR and grade 3 for severe DR. Each grade is classified by the number and appearance of MAs and haemorrhages. DR with MA has 6.2% possibility to develop into Proliferative Diabetic Retinopathy within a year. 148

2 Fig 2: Examples of Spot Lesions: MA, Haemorrhage, And Hard Exudate [5] In this paper, we mainly compare the methods for the automatic detection of microaneurysm based on parameters like sensitivity and specificity. II. METHODOLOGIES USED IN DETECTION OF MICROANEURYSM Automatic detection of micro aneurysm helps in diagnosing Diabetic Retinopathy at an early stage. The ophthalmologists have to examine a large number of images in order to perform mass screening. The cost of manual examination is prohibiting the implementation of screening on a large scale. A possible solution could be the development of an automated screening system for retinal images for the early detection of DR. Such a system should be able to distinguish between retinal images containing microaneurysm (MA) and normal retinal images, which will significantly reduce the workload for the ophthalmologists as they have to examine only those images diagnosed by the system as potential anomalies containing affected retina. The Existing methodologies work in three stages:-image pre-processing, Feature Extraction and Classification. First stage requires image processing for the removal of noise and contrast enhancement. It is mainly performed in green plane because MAs have higher contrast with the background. The Second stage involves extraction and selection of the essential feature vectors to detect microaneurysm. In the third stage, classification is done to categorize as microaneurysm candidate or non-micro aneurysm candidate. The general process for detection of microaneurysm is shown in Fig.3. Acquisition of Retinal Fundus Images Image Pre-processing Image Segmentation Feature Extraction Microaneurysm Classification Fig 3: Block Diagram for the Detection of MA In the following subsections, we describe the various approaches to detect microaneurysm. A. Morphological Approach Morphological processing is the most common method for the detection of lesions like micro aneurysm. Akara Sopharak et al.[1]proposed a morphology based method for the detection of MA. The digital retinal images were obtained from KOWA-7 non-mydriatic camera with the Field of view of 45. The Images were preprocessed in green plane since the red lesions have high contrast with its back ground in this color plane. It was then median filtered to attenuate the noise followed by Contrast Limited Adaptive Histogram Equalization (CLAHE).Shade Correction Algorithm was applied to the image to remove the slow background variation due 149

3 to non-uniform illumination. Exudates and Vessels are detected, prior to the MA Detection. Coarse Segmentation was done to identify the MA candidates. The extended-minima transform was applied on the shade corrected image with the threshold value of 0.05 and the detected exudates and the vessels were removed from the above image. Feature Extraction helps in extracting the essential features that distinguish MA pixels from the non-ma pixels.18 such features like pixels intensity value of shade corrected image, pixel s hue, Perimeter,Area Circularity, Eccentricity were obtained. Then, Fine Segmentation using naïve Bayes Classifier was applied.it was performed using 45 Non-Dilated retinal images using MATLAB program and a weka data mining software running for feature discretization and naïve Bayesian classification. Detected MA were compared with the ground truth images for verification and the sensitivity, specificity, precision and accuracy were found to be 85.68, 99.99, and 99.99% respectively. Usman Akram et al.[2] proposed a three-stage system for the early detection of MAs using filter banks. The three stages include candidate region extraction, feature vector formation and classification. The candidate region extraction works in three phases: In phase 1, contrast of dark regions is improved by using mathematical morphological operations, contrast normalization and filter banks. Phase 2 performs enhancement and segmentation of the blood vessels. The last phase of candidate region extraction eliminates all blood vessel pixels from candidate pixels in order to reduce the false positive regions. Gabor Filters were used for the enhancement of the lesions due to their fine frequency tuning and orientation selectiveness.ma can be distinguishable from the other lesions based on color, shape and size. Feature vectors are formed for each candidate region. They are divided into four subsets like shape based features, gray level features, colorfeatures and statistical features.hybrid classifier consisting of Gaussian Mixture Model, Support Vector Machine and an extension of multimodal m-mediod based modelling approach were used for classification of MA and non-ma. The Retinal Fundus Images were mainly acquired from the databases like DIARETDB0 and DIARETDB1 (219 Images).The sensitivity, specificity and accuracy were %, 99.69%, 99.40%. B. Fractal Analysis Rukmini et al.[3] proposed a method which comprises of two stages.the first stage comprises of image preprocessing and fractal Analysis. Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied on the image as it makes the region of interest more clear and visible. The Image is divided into non-overlapping regions of equal size. Local Histogram Equalization was done at every disjoint region. Bilinear interpolation was applied to eliminate the boundaries between the region.it was performed in the green plane since the contrast is higher in this color plane. The Blood Vessels were removed using Thresholding technique. During Thresholding, unwanted pixels may appear as noise. Morphological opening is used to remove the objects that are fewer than 35 pixels. The Boundary can be removed by subtracting a mask image. Fractal Analysis helps in distinguishing normal and abnormal retinal images. Fractal dimension was calculated using the Box-counting method. Box counting Method is based on the principle of the number of square boxes required to cover the fractal. If the fractal dimension of a given image was below the threshold value, it was classified as a normal retinal image else abnormal retinal image. Since, MA appears as round dots, Canny Edge Detection was applied for distinguishing MA from the other lesions. This may leave unfilled holes. Therefore, Morphological Reconstruction was performed.the edge detected image was subtracted from the morphologically reconstructed image to obtain the candidate MA. Various features were extracted like area, perimeter, diameter, circularity, aspect ratio. Receiver operating Characteristics curve was obtained, which displays the relationship between the sensitivity and specificity. Publicly available diabetic retinopathy database DIARETDB1 was chosen which consists of 89 color fundus images. The Sensitivity and the specificity was 89.5% and 82.1% respectively. C. HSV Method and Eccentricity Technique Preeyaporn et al.[4] proposed a method using the combination of HSV method, area identification and eccentricity technique. HSV color model is the method which mainly considers Hue, Saturation, value. It is most similar to Human color perception. Each color bar consists of 10 random MAs captured area. The samples of MAs colors in each color bar are analysed. The selected color was chosen as the base color for the retinal image diagnosis. The HSV values of the base color were added in the algorithm. The approximate HSV values were H average which was 0.019, S average which was and V average which was For the detection of MAs, the suspected images were applied with the algorithm to detect the target color. If the dot in the image lies 150

4 within the range of the target color, the position of the MA was detected. The target areas are separated from the background by applying threshold technique. It was converted into a binary image. The pixel areas of the MAs will be in the range pixels, in white colour in the binary image. Pixels with the value less than 200 will be noise. The shape of the pixels can be divided based on the eccentricity ranging from 0-1.The eccentricity of MAs range from If the eccentricity is less than 0.3,it is considered as a noise and if it is greater than 0.89,it is considered as a vein. The Accuracy was 93%. D. Local Rotational Cross-section Profile Analysis Istvanet al. [7] proposed detection of microaneurysm though local rotational cross-section profile analysis. The inverted green channel of the fundus image was acquired, since MAs, Haemorrhages and vasculature appear as bright structures. It was convoluted with a Gaussian mask with a variance of 1.0.MAs are local intensity maximum structures with a Gaussian like intensity distribution. Every MA region contains at least one regional maximum. A Local Maximum Region (LMR) is a connected component of a pixel with a constant intensity value and the neighbouring pixel has a lower value. Breadth -first search algorithm was applied for grayscale morphological reconstruction. Pixels with the same intensity is stored in a queue.the surrounding of the single maximum pixel in a MA was examined and the intensity values along the discrete line segments of different orientations was recorded. Thus, set of cross-sectional profiles was obtained. Peak detection was performed on the obtained cross- sectional profiles. Peak was detected and several statistical measures were calculated. For classification, naïve Bayes (NB) classifier was used. It was tested on 60 retinal images. This method has achieved a higher sensitivity at lower false positive rates i.e., 1/8 and ¼ FPs/image. Table 1 show the Evaluation and the comparison Results for MA Detection. Table 1: Evaluation and Comparison Results for MA Detection Authors Images Sensitivity Specificity Accuracy Akara 45 Images 85.68% 99.99% 99.99% Sopharak et al [1] Usman 219 Images 98.64% 99.69% Akram et from al [2] DIARETDB0, DIARETDB1 Rukmini et al [3] 89 Images from 89.5% 82.1% Not specified Preeyaporn et al [4] DIARETDB1 Not specified Not Specified Not specified 93% III. CONCLUSION Automatic Detection of MAs is a challenging task.the presence of microaneurysm is the first clinical sign of DR. The ophthalmologists will have to examine a large number of fundus images in order to perform mass screening. An Automatic detection of MAs will help the doctors in detecting DR at an early stage. Various Image Processing Techniques have been used for Enhancement, Segmentation, and Feature Extraction for detecting MA accurately. In this paper, some existing methods based on Image processing are discussed and their specificity, sensitivity and accuracy have been compared. REFERENCES [1] Akara Sopharak, Bunyarit Uyyanonvara and Sarah Barman, Simple hybrid method for fine microaneurysm detection from non-dilated diabetic retinopathy retinal imagesakara, Elsevier, Computerized Medical Imaging and Graphics pp , [2] M. UsmanAkram, ShehzadKhalid, ShoabA.Khan, Identification and classification of microaneurysms for early detection of diabetic retinopathy, Elsevier, Pattern Recognition, pp , [3] Rukhmini Roy, Srinivasan Aruchamy, Partha Bhattacharjee, Detection of Retinal Microaneurysms using Fractal Analysis and Feature Extraction Technique, IEEE, International conference on Communication and Signal Processing, April 3-5,

5 [4] Preeyaporn Yunuch, Noppadol Maneerat, Ronakorn Panj aphongse M.D, Ruttikorn Varakulsiripunth, Automatic Microaneurysms Detection Through Retinal Color Image Analysis,IEEE [5] Xin Zhang and Guoliang, Retinal Spot Lesion Detection Using Adaptive Multiscale Morphological Processing, Fan School of Electrical and Computer Engineering Oklahoma State University, Stillwater, OK [6] Oliver Faust & Rajendra Acharya U. & E. Y. K. Ng & Kwan-Hoong Ng & Jasjit S. Suri, Algorithms for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images: A Review, Springer,2010. [7] Istvan Lazar, Andras Hajdu, Retinal Microaneurysm Detection through Local Rotating Cross-Section Profile Analysis, IEEE Transactions on Medical Imaging Vol. 32, No. 2, February AUTHOR BIOGRAPHY Srilatha L. Rao: Ms. Srilatha.L.Rao is currently pursuing her M.Tech in Biomedical Signal Processing and Instrumentation (Final year) in Department of Instrumentation Technology at RVCE, Bengaluru, Karnataka,India. She has obtained her B.E Degree in Electrical and Electronics from Bangalore Institute of Technology, Bengaluru, Karnataka, India. Deepashree Devaraj: Mrs. Deepashree Devarajis currently an Assistant Professor at RVCE, Bengaluru, Karnataka, India. She has received her B.E Degree from Manipal Universityand M.Tech Degree from RVCE, Bengaluru, Karnataka,India.Her areasof interest includemedical Image Processing, Biomedical Instrumentation, and Signal Processing.She has guided many students at P.G level.she has to her credit many papers in international journals and national and international conference. Dr. S.C. Prasanna Kumar: Dr.S.C. Prasanna Kumar is currently the Professor and Head of the Department of Instrumentation Technology, RVCE,Bengaluru, Karnataka,India. He has received M.Tech and Ph.D. Degree. His areas of interestincludesignal Processing (Acoustic Signal Processing) Biomedical Instrumentation, Signal Processing. He has guided many students at Ph.D. and P.G.level.He has to his credit many papers in international journals and national and international conference. 152