A Meta-Learning Approach for Diabetic Retinopathy Severity Grading

Diabetes is a prevalent kind of chronic disease that results in different complications. One of the most severe diabetic problems ends up with blindness, termed medically as diabetic retinopathy (DR). The reason for blindness due to DR is the lack of proper treatment and monitoring before it progresses to the severe stage. As a result, computerized diagnosis assists physicians in detecting DR early, saving both money and time. Current research uses a Multipath Convolutional Neural Network technique to extract features before classifying lesions according to their severity. Conversely, this model has a high time complexity, which may affect the classifier's performance. To overcome these issues, a new technique is employed to improve DR detection performance. After preprocessing, feature extraction is done to extract 22 global and local features like microaneurysms, exudates, hemorrhages, contrast, entropy, spatial correlation information, and remaining features from the images. Then, meta -learner -based prediction uses weighted stacking -based ensemble classifiers (WSEC). Finally, a Meta Learn Enhanced Recurrent Neural Network (ML-ERNN) is built and deployed to improve the classification's performance. This study works with APTOS and Kaggle datasets. The criteria chosen for model evaluation are precision, recall, F1 -score, and accuracy. This model can be highly effective in forecasting retinal illnesses and helps reduce vision loss rate.