DEEPDIABETIC: DEEP NEURAL NETWORK APPROACH FOR EARLY DIAGNOSIS OF DIABETIC EYE DISEASES

Abstract

Medical imaging diagnosis, image detection, and image classification are all successfully and significantly impacted by deep learning (DL). Diabetic eye disease will be the primary cause of vision loss worldwide, and diabetes is a serious public health problem. The DeepDiabetic framework is a multi-classification deep learning model that was presented in this study to diagnose and classify four distinct diabetic eye diseases: cataract, glaucoma, diabetic macular oedema, and diabetic retinopathy (DR). 1228 photos from six distinct datasets (DIARETDB0, DIARETDB1, Messidor, HEI-MED, Ocular, and Retina) were used to evaluate the suggested models. We evaluated the deep learning models' performance using two distinct geometric augmentation techniques, known as online augmented and offline augmented, in addition to the original dataset. EfficientNetB0, VGG16, ResNet152V2, ResNet152V2 + Gated Recurrent Unit (GRU), and ResNet152V2 + Bidirectional GRU (Bi-GRU) are the five architectures whose performances are examined in this study. These deep learning architectures are thoroughly analysed and evaluated utilising four classes of public fundus datasets (DR, DME, Glaucoma, and Cataract). To the best of our knowledge, the literature does not include any more deep learning models for selecting amongst these models for these particular disorders. The EfficientNetB0 model performs better than the other four suggested models, according on the experiment findings. Based on fundus pictures, the EfficientNetB0 model obtained accuracy, recall, precision, and AUC of 0.9876, 0.9876, and 0.9977, respectively. While the accuracy of the other research was only 88.33%, 89.54%, 97.23%, and 80.33%, respectively, our EfficientNetB0 model reaches 98.76%. Our EfficientNetB0 model delivers much greater accuracy, recall, precision, and AUC when compared to other research such as Fast-RCNN, RCNN-LSTM, and InceptionResNet. The results show that our suggested models—particularly the EfficientNetB0 model—are noticeably more accurate than the most advanced models.