Browsing by Author "Mutyala, Sukumar Vamsi"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Classification of Diabetic Retinopathy and Macular Edema through Segmented Retinal Vessels using Graph Neural Network
    (National Institute of Technology, Silchar, 2026) Mutyala, Sukumar Vamsi
    Every year, countless people go blind from diabetic retinopathy and macular edema and the tragic reality is that most of them didn’t have to. These conditions can largely be stopped in their tracks, but only if they’re caught early enough, before the damage becomes permanent. That early window depends entirely on a doctor’s ability to see what’s actually going on inside the eye. This is why having accurate, detailed maps of the retina’s blood vessels matters so much. The problem is that current deep learning models tend to fall short in exactly the ways that matter most. They produce broken, incomplete vessel maps missing the tiny capillaries and dropping connections between vessels which are precisely the features a doctor needs to identify the earliest warning signs. To solve this problem, this paper introduces a three stage hybrid framework. It takes three steps to operate the pipeline. First, a dual-encoder graph convolutional network (DE-DCGCN-EE) is used to detect the boundaries of the vessels and model the relationships among the colour channels. Second, a novel VessGAT-SAM2 combination of SAM2 and a Graph Attention Network repairs broken vessel segments, ensures vessel continuity in thin capillaries and generates topologically accurate masks. Third, the multi modal classifier combines the fundus image, the refined vessel mask and the SLIC superpixel maps to predict diabetic retinopathy grade (0-4) and the presence of macular edema. The whole pipeline is trained end-to-end with a joint loss function that reflects the correlation between both conditions. For that purpose, experiments were conducted on four datasets which are publicly available: DRIVE, PRIME-FP20, Messidor-2 and IDRiD. The outcomes were impressive. The overall results were superior to the traditional ones like U-Net, standalone SAM2 and GCN based ones in terms of both accurate segmentation of the vessels and classification of the disease. Most crucially, it remained robust in a variety of datasets, and continued to reveal the fine vessel structures that are important to real clinical significance. This is an exciting step towards the development of automated screening systems, which may help to prevent vision loss due to diabetic eye disease.