Automatic detection of diabetic retinopathy including neovascularization based on morphological operations

("Diabetic retinopathy is a widely spread eye disease caused by diabetes complication. Screening to detect retinopathy disease can lead to successful treatments in preventing blindness especially at early stages. An automated decision support system for the purpose of detecting and classifying retinal abnormalities is carried out mainly using an image processing methods as presented in this thesis. The retinal images are automatically analyzed in term of pixel-based and region-based diagnostics accuracies after compared with ophthalmologist's hand-trut~ An adjusted morphology-based, thresholding and mathematical-based pixel segmentation methods are developed to segment the bright lesions from background and to distinguish from other retinal feature especially optic disc. Gradient classifier has been used to distinguish hard exudates and cotton wool spot from bright lesions segmentation result. The preliminary pixel-based hard exudates and pixel-base cotton wool spot analysis are used to support the fact that development of a reliable retinal abnormalities identification system is feasible. There are small dark lesions and large dark lesions detection methods presented in this thesis. Image enhancement, image restoration, morphology operator, thresholding and compactness properties techniques have been used in development of automatic dark lesions detection system. Lastly, neovascularization lesion identification is still a new study in automatic detection of diabetic retinopathy. Detection of neovascularization is important since it signifies the disease has reaches a vision-threatening phase. Therefore, image normalization, morphology-based operator, Gaussian filtering and thresholding techniques are used in developing of neovascularization detection. Moreover, a function matrix box-based on predefined criteria of neovascularization has been used in order to classify the neovascularization from natural blood vessel. The developed method is tested on a set of 303 images from different database sources to classify the images as abnormal or normal images. The proposed method managed to achieved 90.29% score for abnormal images and 100% score for normal images. Result after testing shows 85.39% sensitivity and 94.59% specificity for bright lesions, 70.68% sensitivity and 99.22% specificity for dark lesions and 63.9% sensitivity and 89.4% specificity for neovascularization. Further detail of lesions shows 61.24~ sensitivity and 98.43% specificity for hard exudates, 48.62% sensitivity and 98.75% specificity for cotton wool spots, 27.41% sensitivity and 99.94% specificity for microaneurysms and 67.93% sensitivity and 99.70% specificity for haemorrhages