Multi-level Hybrid Ensemble with Attention-Based Meta-Learner for Diabetic Retinopathy Prediction

Most common complication of Diabetic mellitus is Diabetic Retinopathy: It causes the lesion to occur upon the retina and affects vision if not diagnosed early it triggers blindness. Diabetic retinopathy should be treated by an early diagnosis to avoid irreversible loss of vision. In addition, the manual diagnosis by ophthalmologists is less efficient and can easily miss the smallest detail that, in some cases, may not be visible to naked human eyes compared to the computer-aided systems. This implies proposing an existing supervised learning strategy for detection of DR from retinal fundus images to a hybrid combination of both deep learning InceptionV3 and ResNet and a machine learning model, namely Random Forest and Support Vector Machine. The model architecture incorporates advanced neural networks fused with classifiers which is further tuned and added up with an attention mechanism ensuring robust and one of the most accurate classification model of DR and non-DR cases. The dataset comprises of 30,000 fundus images which is preprocessed and augmented to improve model performance, hence addressing class imbalance. Additionally, a front-end app with Grad-CAM analysis is developed to classify DR and Non-DR images and visualize where the model focuses during classification.