Optimizing Diabetic Retinopathy Detection with Advanced SMOTE-based Data Augmentation

Diabetic retinopathy (DR) is a prevalent condition associated with prolonged diabetes, which impacts the ocular functions of individuals and has the potential to result in irreversible visual impairment. The early detection of DR is most vital for effective and better treatment, as the symptoms often manifest in later stages. Ophthalmologists use fundus images of the retina as part of the manual diagnostic process for diabetic retinopathy (DR). Most of the existing models utilized publicly accessible datasets for their research. Publicly available datasets exhibit a notable disparity in the distribution of data. The presence of data imbalance can significantly hinder the performance of classifiers, particularly when the minority class holds significant importance. This challenge becomes even more arduous when the dataset consists of more than two target classes. The primary aim of the proposed model is to create an automated model for classifying exudates for diabetic retinopathy using the Random Forest (RF) algorithm using three oversampling methods, such as ROS (Random Over Sampling), SMOTE (Synthetic Minority Oversampling Technique) and ADASYN (Adaptive Synthetic Sampling Approach). The model was trained and evaluated using the publicly available DR datasets DIARETDB0 and DIARETDB1. The findings of this study demonstrate that SMOTERF has superior performance compared to other data-level techniques. Along with this, the study proposed a new and advanced SMOTE-based data augmentation technique for more accurate DR classification results.