A convolutional neural network for the screening and staging of diabetic retinopathy

• DOI: 10.1371/journal.pone.0233514
• Publication Date: 2020-06-22T17:38:22Z
• Authors: Mohamed Shaban; Zeliha Ogur; Ali Mahmoud; Andrew Switala; Ahmed Shalaby; Hadil Abu Khalifeh; Mohammed Ghazal; Luay Fraiwan; Guruprasad Giridh...; Harpal Sandhu; Ayman S. El-Baz
• Keywords: Fundus (uterus); Fundus Photography; Cotton wool spots

Diabetic retinopathy (DR) is a serious retinal disease and considered as leading cause of blindness in the world. Ophthalmologists use optical coherence tomography (OCT) fundus photography for purpose assessing thickness, structure, addition to detecting edema, hemorrhage, scars. Deep learning models are mainly used analyze OCT or images, extract unique features each stage DR therefore classify images disease. Throughout this paper, deep Convolutional Neural Network (CNN) with 18 convolutional layers 3 fully connected proposed automatically distinguish between controls (i.e. no DR), moderate combination mild Non Proliferative (NPDR)) severe group NPDR, (PDR)) validation accuracy 88%-89%, sensitivity 87%-89%, specificity 94%-95%, Quadratic Weighted Kappa Score 0.91-0.92 when both 5-fold, 10-fold cross methods were respectively. A prior pre-processing was deployed where image resizing class-specific data augmentation used. The approach considerably accurate objectively diagnosing grading diabetic retinopathy, which obviates need retina specialist expands access care. This technology enables early diagnosis objective tracking progression may help optimize medical therapy minimize vision loss.