رسائل ماجيستير
English
2019
Automated Approach for Detection and Grading of Retinal Lesions Using Hybrid Transform
أسامه قاسم جمعه , حسين فاضل حمدان, سهير حسين طالب يعقوب
كلية الهندسة-جامعة بابل · العراق
Several pathological conditions can cause damage and hurt to the retina of the eye, such as diabetic retinopathy, and this disease can cause unalterable damage to the eye. It is one of the eye diseases and is a major cause of blindness. Initial signs are damaged blood vessels of the eye, which form lesions in the retina. The beginning of this disease starts with red lesions, for instance, microaneurysm and hemorrhages. After a period, these red spots become ulcerated lesions and seem yellowish in color; these are exudates. This is the most serious stage of the disease, called maculopathy, because of the near lesions from the fovea, which causes lack of vision.
So, early examination of the retina helps reduce the effect of the disease on the eye. In addition, diabetic patients are regularly subjected to take a digital image of the retina, and then they are analyzed by trained individuals, which consumes a large amount of time and incurs a high cost because the diabetic population is large. Fortunately, this field of study is very important and has a benefit in introducing systems for automated detection of the disease. The use of computers can aid the auto-detection and grading of these diseases, since computers can expedite the test and correct the diagnosis.
In this work, an improvement of computerized algorithms is proposed to use them in the medical screening system for detecting retinal lesions and grading risk to obtain the suitable treatment decision for the patient, which will assist clinicians to diagnose diabetic retinopathy at its early stage.
In this work, a new algorithm for detecting hard exudates in retinal images in an automated way is proposed. It uses a new algorithm named here as Triple Circular Segmentation and a Kalman Filter. Morphological operations are also applied to enhance and refine the Kalman filter results. To equalize the uneven illumination associated with fundus images, in addition to removing noise from the image, a pre-processing stage is used. By this algorithm, it is obtained a sensitivity of 88%, positive predictive value of 87%, accuracy of 99.5%, and specificity of 99.9% when taking 60 images as test images and 20 images to test the performance from different databases. These results are excellent in lesion detection operations as compared with other existing techniques.
Diabetic retinopathy grading is an important issue after detecting lesions of the retina to estimate their risk and to take a suitable decision for treatment. Here, the grading of diabetic retinopathy is examined by consistent medical approaches to build a computer model for grading in an automated way, which improves the efficiency of diabetic screening services. After the grading of diabetic retinopathy, Delta Neural Network Learning Rule is used to give suggestions to a doctor for suitable treatment for the patient. Here, sixteen different cases are trained, and it takes 20.519063 seconds with 152,962 iterations. The neural network diagnoses four treatment cases: urgent, moderate, mild, and normal. It is also found from the results that the neural network is a very fast algorithm to give these decisions. In addition, the program used for carrying out the process is MATLAB Program version 2016, and the computer is HP Core i5.