TABLE OF CONTENTS
TOC o "1-3" h u HYPERLINK l _Toc26935 ABSTRACT PAGEREF _Toc26935 2.
HYPERLINK l _Toc25985 CHAPTER 1 PAGEREF _Toc25985 2
HYPERLINK l _Toc4867 INTRODUCTION PAGEREF _Toc4867 3.
HYPERLINK l _Toc11726 CHAPTER 2 PAGEREF _Toc11726 5
HYPERLINK l _Toc5164 DETECTION AND CLASSIFICATION OF HARD EXUDATES PAGEREF _Toc5164 5
HYPERLINK l _Toc6853 2.1. BLOCK DIAGRAM PAGEREF _Toc6853 5
HYPERLINK l _Toc15 2.2 DETECTION OF HARD EXUDATES PAGEREF _Toc15 6
HYPERLINK l _Toc12897 2.2.1 REGION OF INTERESTS PAGEREF _Toc12897 6
HYPERLINK l _Toc19589 2.2.2 GENERATION OF MOTION PATTERNS PAGEREF _Toc19589 6
HYPERLINK l _Toc9151 2.2.3 FEATURE SELECTION PAGEREF _Toc9151 6
HYPERLINK l _Toc16981 2.2.4 ABNORMALITY DETECTION PAGEREF _Toc16981 7
HYPERLINK l _Toc31281 2.3 CLASSIFICATION OF HARD EXUDATES PAGEREF _Toc31281 7
HYPERLINK l _Toc22126 CHAPTER 3 PAGEREF _Toc22126 9
HYPERLINK l _Toc14903 RESULTS PAGEREF _Toc14903 9
HYPERLINK l _Toc22925 CONCLUSIONS PAGEREF _Toc22925 11
HYPERLINK l _Toc16101 REFERENCES PAGEREF _Toc16101 12
Diabetic Retinopathy is the damage caused to the blood vessels in retina due to diabetes. The severe case of diabetic retinopathy leads to vision loss. It is important to diagnose diabetic retinopathy in earlier stage. In this work automatic methods for detection various lesions of diabetic retinopathy from color fundus images are explained. The retinal structures which include blood vessels, optic disc and fovea are also detected. The prominent lesions present in an abnormal color fundus image include the brighter lesion such as hard exudates and darker lesions such as microaneurysms and haemorrhages. The severity of the disease based on location of the hard exudates in the retina is also explained. Hard exudates are detected by a supervised learning technique on normal color fundus images. The global features of normal color fundus image are captured using a feature extraction technique. Based on this feature the images are classified to be normal or abnormal. The classification of abnormal image as moderate or severe is done by considering the rough rotational symmetry of the macula of a normal color fundus image. The presence of red lesions is detected based on its appearance on the color fundus image. Hard exudates are detected with an accuracy of 95% and classified with an accuracy of 96%.
The diabetic retinopathy is a disease occurring in persons suffering from diabetes. The disease leads to progressive damage of retina and eventually in vision loss. Thus the detection of diabetic retinopathy in its earlier stage is very significant. Any damage to the tiny blood vessels of the retina results in diabetic retinopathy. This also leads to leakage of blood and other fluids resulting in swelling of retinal tissue. Manual detection of diabetic retinopathy lesions by an ophthalmologist is difficult as it requires more time to analysis the color fundus images. Thus automated screening techniques for lesion detection have great significance in saving cost and time. The automatic disease detection system can highly reduce effort of ophthalmologist to limit the immediate attention to the severe cases. Diabetic retinopathy can be automatically detected by examining the different lesions present in the color fundus images. The different lesions that may be present in an affected retina are microaneurysms, haemorrhages, hard exudates etc. Figure. 1.1shows the color fundus image with different retinal structures and lesions labeled on it.
Figure1.1: Color fundus image with retinal structures and
Microaneurysms are the red dots seen in the layers of retina which represent out pouching of the retinal capillaries and are the first sign of diabetic retinopathy. Haemorrhages are present in severe case of the disease with bleeding into the deep layer of the retina. Hard exudates are minute, yellow, and well defined deposits of lipo–protein.The hard exudates are detected by a supervised learning technique on normal images. The severity of diabetic retinopathy depends on the proximity of hard exudates to the macula. If the hard exudates are present outside macula the disease is moderate and severe if it is present inside the macula. The methods presented are limited to the detection of hard exudates only. In this paper, both hard exudates and red lesions of diabetic retinopathy are detected and the severity of the disease is also analyzed.
CHAPTER 2DETECTION AND CLASSIFICATION OF HARD EXUDATESHard exudates appear as bright structures with well defined edges and variable shapes. The block diagram showing the different stages of hard exudates detection and classification of the disease is shown in Figure. 2.1. First a decision module validates the presence or absence of hard exudates in a color fundus image. Once the existence of hard exudates is confirmed next a second module assesses the macular region to measure the risk of the disease. Thus, a two-stage methodology for both detection and assessment of the disease is proposed. A supervised learning technique is used for detecting the hard exudates. The global characteristics of the normal color fundus image are analyzed and are used to discriminate it from abnormal images. The rotational symmetric feature of the macula of a normal image is used to access the severity of the disease.
2.1. BLOCK DIAGRAM
Figure 2.1: Block diagram for detection and classification of hard exudates
2.2 DETECTION OF HARD EXUDATESA circular region of interest with fovea as the center is cropped from the RGB image. Motion patterns are generated and studied to detect the hard exudates. Different steps in the detection of hard exudates are explained next.
2.2.1 REGION OF INTERESTS
Since the severity of disease is found by analyzing the location of hard exudates with respect to the macula, the images used for hard exudates detection usually focus around the macular region. Thus a circular region with macula as the center is cropped and optic disk is masked by a black rectangular mask. The green channel of resulting image forms the input for all subsequent processing.
2.2.2 GENERATION OF MOTION PATTERNS A smear pattern is generated by inducing motion in a single image. Sequences of rotated images are generated by inducing motion in given image. The rotated images are combined by using a function to combine the intensities at each pixel location to give a motion pattern . Here a motion pattern is generated for the green channel of the region of interest with optic disc masked.
The difference between the motion patterns of normal image and abnormal image with hard exudates can be clearly seen. The motion pattern of normal image does not have any white patches whereas the motion pattern of the abnormal image has white patches on it.
2.2.3 FEATURE SELECTIONTo effectively describe the motion pattern, a descriptor formed from the Radon space is used. The Radon transform of is the integral of a function f(x,y) along a line oriented at angle ‘a’ and distance ‘r’ from the origin. The image is projected to get a vector response for each angle. By combining the responses for different orientations the desired feature vector is obtained. The extent of hard exudates present in the image is extended in the motion pattern and is reflected in the projection based feature vector. Thus for an abnormal image the feature vector will have many peaks due intensity of hard exudates whereas for a normal image the feature vector will have comparatively uniform values which result in a compact normal subspace. The feature vector thus obtained is used to learn the subspace of normal images.
2.2.4 ABNORMALITY DETECTIONA classification the boundary is fixed around the normal subspace. A new image to be tested is transformed to this normal subspace. If it lies inside the normal subspace boundary, the image is classified to be normal, else abnormal. A PCA DD (principal component analysis data descriptor) is used for classification.
PCA DD: In a PCA classifier, a linear subspace is defined corresponding to the normal cases. The subspace is defined by the Eigen vectors corresponding to the covariance matrix of the training set. The feature vector for a new image is projected to this subspace and is reconstructed. Then based on a reconstruction error new case is classified to be normal.
Different steps in abnormality detection are:
Step 1 Some normal images are selected for training and radon transform of these images are found.
Step 2 The Eigen vector corresponding to the covariance matrix of images in training set is calculated.
Step 3 The average of all the Eigen vectors of normal images in the training set are taken and stored as dataset.
Step 4 The Eigen vector of the covariance matrix of the radon transform of each new image is calculated and projected to the dataset.
Step 5 The difference in the data is found and if it’s above a given threshold the image is said to be abnormal else normal.
2.3 CLASSIFICATION OF HARD EXUDATESThe circular ROI with macula as the center is the area of key interest as any hard exudates within this region indicates high risk of disease. The macula is a relatively darker structure compared to other regions in the fundus image. The macula also possesses a rough rotational symmetry. This symmetry information is used to find the risk of exhibiting the disease. If the degree of symmetry is above a particular threshold, it shows that the abnormality is not inside the macula and thus the image is of moderate severity else the image shows high risk of disease.
A symmetry measure is given by the distance between second norms of the histograms of pair of diametrically opposite patches. The macula is divided into eight patches.For each patch the histogram of 10 bins are computed. But for measuring the symmetry only the last five bins are used since intensity due to hard exudates is reflected in the higher bins of the histogram. The severity of the abnormal image is found by comparing the measure of symmetry this image to a threshold. Let Smin and Smax be the minimum and maximum symmetry values for normal images used in the training set for the detection of abnormal images. Now the severity of the given abnormal image is found by comparing the symmetry measure of this image against a threshold.
CHAPTER 3RESULTSAfter segmentation, each image is characterized by its corresponding segmented region.It is necessary to discriminate the extracted region as exudates or non exudates. The selected sets of features are standard deviation of the intensity, mean, intensity, size, edge strength and compactness.Each image was classified as normal or abnormal according to the presence or absence of exudates. A patient was classified as abnormal if the presence of exudates is found else it is classified as normal.
CONCLUSIONSAutomatic method for the detection of retinal structures such as hard exudates in color fundus images are explained. The severity of the disease based on the location of hard exudates with respect to the fovea was also analyzed.
REFERENCES1 Ream , Pradeepa.R , “Diabetic retinopathy: An Indian perspective, Indian Journal of Medical Research”, 125 (2007) 297-310.
2 Lena Kallin Westin “Receiver operating characteristic (ROC) analysis, Evaluating discriminance effects among decision support systems”, UMINF 01.18, ISSN-0348-0542.
3 U R Acharya, C M Lim, E Y K Ng, C Chee, T Tamura, “Computer-based detection of diabetes retinopathy stages using digital fundus images” , Proc. IMechE Vol. 223 Part H: J. Engineering in Medicine.
4 L. Giancardo, F. Meriaudeau, T. Karnowski, K. Tobin, E. Grisan, P. Favaro, A. Ruggeri, and E. Chaum,”Textureless macula swelling detection with multiple retinal fundus images,” IEEE Trans. Biomed. Eng., vol. 58, no. 3, pp. 795–799, Mar. 2011.
CO1:Learnt basic principles of medical image communication,image coding and pattern classification/recognition
CO2:classify the various medical image processing algorithms and fundamental methods for image enhancement
PS02:solve complex issues related to aquisition and analysis of a diverse range of biomedical signals and images using modern tools and processing algorithms