Department of Electronics and Communication Engineering

Permanent URI for this collection

http://nits.ndl.gov.in/handle/123456789/91

Browse

Browsing Department of Electronics and Communication Engineering by Author "S., Muralitharan"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Classification of Diabetic Retinopathy through Segmented Microaneurysm Features Using Deep Learning Algorithm
    (Department of Electronic and Communication Engineering, NITS, 2024-06-06) S., Muralitharan
    Diabetes Retinopathy (DR) is a consequence of diabetes that can result in blindness. Elevated blood sugar levels cause retinal capillaries to shrink, resulting in the production of tiny, blob-like formations known as microaneurysms (MA). Detecting these little MAs manually is challenging and usually requires specific training from ophthalmologists. As a result, there is an urgent need for automated MA detection and DR categorization approaches to speed up diagnosis and treatment for a greater number of patients. This thesis is divided into two chapters, both of which focus on DR classification utilizing Machine Learning and Deep Learning methods. The MA detection procedure typically consists of three stages: preprocessing, augmentation, and classification. Preprocessing is very important for improving retinal characteristics and eliminating imaging artifacts, which are a difficult element of retinal fundus image analysis In this paper, we provide a new Improved Non Local Mean filter (INLMF) that successfully eliminates image artifacts. Furthermore, we present an improved U-Net architecture for MA segmentation, in which traditional U-Net blocks are substituted with customized Res-Net blocks. These Res-Net blocks are made up of three residual blocks connected by a dropout layer, which provides increased performance. Our goal is to detect DR early in the presence of MAs. To do this, we provide a unique deep CNN model that diagnoses DR severity based on the number of MAs present in fundus pictures, categorizing it as normal, mild, or moderate. Both strategies are carefully tested on publicly available datasets such as e-ophtha, with our segmentation strategy outperforming state-of-the-art measures.