A Lightweight Resource Constrained Automated Deep Learning Model for Diabetic Retinopathy Detection

Abstract

Diabetic Retinopathy (DR) is the main reason of blindness in diabetics requiring quick, precise diagnosis to prevent and manage vision loss. This dissertation contribution three important work to tackle of this critical healthcare challenge. newlineThis research introduces a strong mobile application based on deep learning, which provides an innovative and effective tool for DR detection in remote and resource-limited regions. newlineUsing the VGG model as the backbone of CNNs, the application can get a high accuracy (96%) newlineand well work offline. Therefore, this application is highly accessible, which is important in regions where health infrastructures are not developed enough to meet the current requirements. newlineThis mobile application represents a democratization effort such that patients in the countryside could be diagnosed and intervened promptly. newlineSecondly, this research introduces an innovative CNN architecture called DiabNet to greatly improve the performance of DR detection from fundus images in terms of both accuracy and efficiency as well as robustness. By exploiting the state-of-the-art apparatuses such as attention newlinemechanisms, skip connections, and batch normalization, the proposed DiabNet yields newlinethe state-of-the-art performance with the accuracy rate of 98.72%, which significantly outperforms the existed methods. This helps to bring the DR detection closer to early-diagnosis and prevention, which in turn can greatly mitigate the burden of diabetic-induced blindness in the society. Besides, by deploying the proposed DiabNet as a mobile application, the proposed approach extremely facilitates DR screening in a compact, affordable and accessible way. By newlinepassing their fundus images through a cell phone s camera, at anywhere and at any time, users are able to track their retinal health condition easily and privately. newlineFinally, as a hybrid model DyFuseNet successfully integrates MobileNetV2 and DenseNet-121 s significant strengths on understanding and presenting rich details of retinal images with impressive precision. By