الفهرس 
CHAPTER 1Only 14 pages are availabe for public view
 |  | from | 181 |  |  |
| | | from | 181 | | |
Abstract
The organs of the human body may be infected by dierent diseases or lesions simultaneously. On the other hand, the infected organ may aect other organs in the human body. One of these diseases is diabetes. Diabetes is a chronic disease characterized by blood glucose level elevation. This elevation leads over time to severe damage of the human blood vessels (BV), eyes, and nerves. Diabetic retinopathy (DR) is primarily one of the common complications of diabetes. DR is a dangerous and progressive disease that may cause blindness suddenly without any indications. Therefore, it is necessary to continuously screen and audit the disease progress from early to severe grades. The DR grades are categorized into (mild, moderate, and severe) non-proliferative DR (NPDR) and proliferative DR (PDR). These grades are formulated by appearing multiple DR lesions simultaneously on the color retinal fundus image. There are many DR lesions, such as neovascularization (NV), vascular dilation or tortuosity, hemorrhages (HM), microaneurysms (MA), and exudates (EX). Multi-label classication (MLC) is deemed an eective and dynamic research topic in the medical image analysis eld. For ophthalmologists, MLC benets can be utilized to detect early DR signs and its dierent grades to initiate appropriate treatment and avoid DR complications. For developers, it is very important to lessen misclassication and increase the diagnosis accuracy. The developers and ophthalmologists face some problems and challenges, such as DR detection is accomplished by involving a well-trained physician. The availability of retina experts and well-trained graders is a signicant limitation. Even when they are available, there could be a time delay in graders submitting their DR grading and advice due to their busy schedule. It leads to delayed interpretation and loss of follow-up, miscommunication, and delay in DR severity management. The manual retina’s structural changes and blood vessels (BV) abnormalities detection may be inconsistent and time-consuming. It depends on the physician’s experience. Moreover, hand-crafted feature tools are sensitive to the contrast of fundus images. Fundus images have insucient contrast, noise, and artifacts, which lead to inaccurate detection of some DR signs. The features similarity between the eye anatomies and DR lesions is another challenge of the fundus images. Although the previous problems of fundus imaging modality, it is inexpensive, widely used in developing countries, and timely documents the retinal abnormalities. This thesis proposed a comprehensive multi-label (ML) computer-aided diagnostic (CAD) system based on hand-crafted segmentation and feature extraction methods. The proposed ML-CAD system exploits the MLC of DR grades using colored fundus photography. We utilized MLC based on problem transformation to enable future expansion. MLC idea depends on label correlations, which can result in unprecedented labels from the existing labels. Therefore, the proposed system detects and analyzes various retina pathological changes accompanying DR development. We extracted some signi- cant features to dierentiate healthy from DR cases as well as dierentiate various DR grades. First, we preprocessed the retinal images to eliminate noise and enhance the image quality using histogram equalization for brightness preservation based on dynamic stretching technique (HEBPDS). Second, the images were segmented to extract four pathology variations: BV, EX, MA, and HM. It was essential to integrate two dierent lters in order to produce an accurate BV network. Next, six signicant features were extracted using a gray level co-occurrence matrix (GLCM) with 12 descriptors, the four extracting pathological areas, and BV bifurcation points (BP) counts. Finally, the feature vector was supplied to the multi-label support vector machine (MLSVM) classier based on classier chain (CC) to distinguish normal and dierent DR grades. Moreover, we proposed a second ML-CAD system based on both handcrafted and deep learning (DL) techniques. It is an extension and promotion to the rst proposed one with some dierences. First, a binary classication phase was added. It depends on hand-crafted feature extraction. This phase is used to distinguish healthy from DR cases. Second, some post-processing steps were utilized in the framework to prepare images for the segmentation phase. Third, the fundus images were segmented by utilizing a customized, universal DL U-Net model. In fact, shallow classication models’ performance depends on the quality of the features fed into them. On the other hand, classication is mainly based on the accuracy of the segmentation phase. Therefore, we customized the U-Net model by establishing it deeper and customized its hyperparameters to provide precise results. Fourth, we increased the features that we extract to classify the DR grades accurately. We extracted 11 descriptors of the grey level run length matrix (GLRLM) on four directions, which are 0o, 45o, 90o, and 135o for each image. Finally, we evaluated the performance using six dierent performance metrics on nine datasets (four of them are ML) and compared it with many current conducted systems and methods. We proposed a third ML-CAD system mainly based on DL. DL achieved great success in medical image analysis. Deep convolution neural network (CNN) architectures are widely used mainly in MLC. The DR lesions are several and have many diculties to segment features. They can not be distinguished by utilizing conventional and hand-crafted methods. Therefore, the practical solution is to utilize an eective CNN model. Therefore, we proposed a novel hybrid, DL technique,” E-DenseNet.” We integrated the EyeNet and DenseNet models based on transfer learning. We customized the traditional EyeNet, embedded the dense blocks, and optimized the resulting hybrid E-DensNet model’s hyperparameters. The proposed system based on the E-DenseNet model can accurately diagnose healthy and dierent DR grades from various small and large ML color fundus images. The model was trained and tested on four dierent benchmark datasets that were published from the year 2006 to the year 2019. The rst proposed system was trained and tested on four datasets; two of them are ML. All datasets have ground truth (GT) to give the ability for evaluation. They include GTs of BV, EX, MA, HM, and DR grades. The proposed system was evaluated by six performance metrics. The rst proposed system achieved an average accuracy (ACC) of 89:2%, sensitivity (SEN) of 85:1%, specicity (SPE) of 85:2%, positive predictive value (PPV) of 92:8%, area under the curve (AUC) of 85:2%, and Disc similarity coecient (DSC) of 88:7% The second system was trained and tested on nine datasets with GTs; four of them are ML and evaluated on six performance matrices. It achieved 95:1%, 91:9%, 86:1%, 86:8%, 84:7%, 86:2% for ACC, AUC, SEN, SPE, PPV, and DSC, respectively. The third system was trained and tested on four ML datasets with grading GTs. It was evaluated by ve performance metrics. It achieved an average ACC equals 91:2%, SEN equals 96%, SPE equals 69%, DSC equals 92:45%, and the quadratic Kappa score (QKS) equals 0:883. The three systems’ experiments show encouraging results compared with other systems in segmenting the DR lesions and diagnosing the disease grades.