Segmentation of fundus images for diabetic retinopathy evaluation

Diabetes, one of the most common diseases, leads to various complications. One of these complications, diabetic retinopathy, is the leading cause of permanent blindness in adults. Regular screening for this disease not only reduces the risk of vision loss for patients but also lowers healthcare costs through early diagnosis. To establish an efficient public health screening system, it is essential to reduce the workload on doctors. Therefore, automating diabetic retinopathy checks has become a popular topic in recent years. A system that enables the detection of diabetic retinopathy must learn to recognize signs of the disease, whether they are explicit or hidden. In this thesis, we focused on the direct segmentation of lesions required for classifying the disease. For this purpose, we used the IDRiD dataset, which includes four important lesion types: microaneurysms, hemorrhages, exudates, and soft exudates. In this study, we addressed the challenges arising from the limited number of images and lesions in the dataset. To achieve this, the following dataset-focused methods were developed: -Pre-processing methods to enhance the visibility of retinal lesions. -Data augmentation method by adding lesions to healthy images. -A three-stage transfer learning method. With the help of these methods, various parameters and combinations of other techniques were tested, resulting in achieving the second-highest performance value in the literature. The positive effects of the listed methods were tested and presented. Keywords: segmentation, diabetic retinopathy, data augmentation, deep learning, artificial intelligence, image processing, denoising