Cloud-Based Diabetic Retinopathy Severity Recognition System Using Ensemble Deep Convolutional Neural Network Classifier Model

One of the key reasons for visual impairments is due to the ignorance of diabetic retinopathy disease. This research study focuses on the early recognition of diabetic retinopathy disease from the fundus images and identifies its severity stages to make successful treatments against blindness risk. Some traditional approaches explored the decision tree, kernel-based support vector machine, and Naïve Bayes classifier models to extract the features from fundus images. Most of the researchers applied the modern approach of convolutional neural network model through transfer learning mechanism to extract relevant features from the fundus images. It helps in the diagnosis of diabetic retinopathy that may delay the prediction process and create inconsistency among the doctors. So, a deep learning-based approach is proposed in this research study to provide stage-wise prediction of diabetic retinopathy disease with a multi-task learning mechanism. As a result, the proposed deep convolutional neural network classifier with an ensemble model outperforms the existing classifier with EfficientNet-B4, EfficientNet-B5, SE-ResNeXt50 (380 × 380), and SE-ResNeXt50 (512 × 512) networking methods in the context of prediction correctness, sensitivity, specificity, macro F1, and quadratic weighted kappa (QWK) score metrics. Exploiting hyperparameter optimizations on the deep learning classifier model and multi-task regression learning approaches make significant improvements over the performance evaluation metrics. Finally, the proposed approaches make the effective recognition of diabetic retinopathy disease stages based on the human fundus image.