Detection of hard exudates in retinal images using soft computing techniques

Abstract

Diabetic Retinopathy (DR) is the most common retinal disease for a diabetic which causes blindness in the working age population. It occurs when the blood vessels inside the retina are damaged due to the blood pressure of vessels. Continuous screening system of diabetic retinopathy is one of the significant ways to know the symptoms, to find the best solutions. Early stage detection of diabetic retinopathy helps the medical professionals or ophthalmologist to avoid the causes of vision loss or blindness. Early signs of DR are the emergence of micro aneurysms, haemorrhages and hard exudates. Moreover, the automatic segmentation of DR lesions is attracted by several researchers due to the demand of regular screening of diabetic retinopathy and digitized information in ophthalmology. The proposed system uses top hat filtering technique to improve the quality of images, avoid the unnecessary distortions in the image and sharpens the edges of the hard exudates region. Then improved histogram equalization technique is used for enhancing the retinal OCT images. To remove the retinal blood vessels, the morphological closing operation is performed over the enhanced image which changes the structure or type of the objects utilizing a structuring component. After pre-processing step, the pre-processed output is segmented to isolate the hard exudates region. In segmentation, modified Region Based Active Contour method is applied that accurately segment the images with inadequately characterized boundaries. Moreover, three steps are performed to segment the hard exudates region from the pre-processed image: The steps involved in hard exudate region segmentation are (i) OD and hard exudates detection (ii) Segmentation of detected OD and hard exudates (iii) Removal of OD region from the segmented image. Next thresholding operation is performed by changing the gray image into binary image After that the segmented image is extracted by using Gray Level Cooccurrence Matrix (GLCM) which evaluates the pair of pixels with the values obtained from the retinal OCT images. Finally, a novel classification technique Random Forest is performed with the help of extracted feature values, which classifies the hard exudates region effectively. It is supervised classification approach which produces the forest with an amount of tress. In this thesis, the proposed Random Forest classification is compared with the various classification techniques (CNN, SVM, ANN, KNN, DNN and CCNN) to classify the hard exudates region. The proposed method acquired the images from Diabetic Retinopathy Database (DIARETDB0) database which is in Digital Imaging and Communications in Medicine (DICOM) format newline