Design and Development of a Novel Framework for Severity Prediction of Diabetic Retinopathy

Abstract

Diabetes Mellitus (DM) is the root cause of the eyesight impairment known as newlinediabetic retinopathy (DR). In humans, it damages the retina, leading to vision issues newlineand eventual blindness. Statistics show that DR affects 80% of people with diabetes newlinewho have had the disease for 15 to 20 years. Thus, it has become a grave threat to newlinepeople s health and life. Therefore, it is essential to stop DR from developing into newlinesevere stages and causing blindness, so it is necessary for early detection and newlinediagnosis. This work focuses on classifying the severity of DR into the five newlinecategories such as normal, mild, moderate, severe, and proliferative DR (PDR). More newlineefficient and reliable design and development of medical image diagnosis is newlineenvisaged. The existence of arbitrary shapes, colours, and location makes illness newlineboundary segmentation a difficult process. newlineThe current systems accuracy in categorizing the various DR severity levels is newlineconstrained. In this work, a Stack Ensemble model is designed to enhance the newlineperformance of existing pre-trained models such as EfficientNetB0, MobileNet, and newlineXception to rate the DR severity. The performance of the proposed model is assessed newlineon the retina datasets from the Asia Pacific Tele-Ophthalmology Society (APTOS) newlineand KAGGLE. With an accuracy of 96% and 97% on the KAGGLE and APTOS newlinedatasets, respectively, the proposed model is surpassed state-of-the-art pre-trained newlinemodels. newlineIn stack ensemble, the model s accuracy is better than other models, but training takes newlinea long time. To achieve greater accuracy in categorizing diabetic severity prediction, newlinea model with minimal computational complexity and strong generalisation ability is newlineproposed. This work proposes a hybrid deep-learning approach for automatic severity newlinedetection in DR. To overcome the restriction of simple skip connections in the newlinesegmentation of DR pictures, the suggested Hybrid Residual U-Net (HRUNET) is newlineutilised for the contraction and extended routes from the down-sample to the upsample. newline