Predicting Sevearity of Type 2 Diabetic Retinopathy Using Retinal Fundus Images with new DeepLearning Architectures from the Inception and Resnet Families

Due to its high morbidity and medical expenses, DR mellitus (T2DM) is a serious global health concern. Its burden can be lessened with effective early prognosis, but the available treatments are not enough. Predictive modelling of Retina Images (fundus) can enhance early diagnosis and care quality. Advanced convolutional neural networks (Inception-V4, ResNet V1, and ResNet V2) are used in this study to assess the severity of Diabetic Retinopathy caused by T2DM. Using de-identified fundus data from 9,948 individuals (1,904 with a T2DM diagnosed), we used transfer learning to train these models. We have performed simple data augmentation to rectify the class imbalance. The accuracy and AUC of the ensemble model were 94.21% and 98.53%, respectively. The findings imply that the combination of ResNet and Inception-V4 enhances classification power, providing a viable approach for early T2DM identification from fundus data that may find use in clinical settings. Future research still needs to focus on improving sensitivity and integrating real-time.