Improving Early Detection of Diabetic Retinopathy: A Hybrid Deep Learning Model Focused on Lesion Identification

Diabetic Retinopathy has been found to be the leading cause of sight impairment in most parts of the globe, particularly in diabetic patients. An early detection of DR in retinal images could considerably reduce the chances of blindness. The present paper develops a hybrid deep learning approach using DenseNet, InceptionV3, and ResNet architecture combined with Random Forest Classifier, XGBClassifier, Decision Tree Classifier, and LightGBM Classifier. Preprocessing techniques consisting of green channel extraction, top-hat/bottom-hat transformations and data augmentation are used. The comparison of the model performance metrics such as precision, recall, F1 score, and AUC are done. It can be seen that the hybrid model has a strong difference in classifying different stages of DR. Comparison of a number of models is also made that comes up with the best model. Even more remains for the near future: improve sensitivity of the model and extend this approach to larger datasets.