Enhanced Classification of Diabetic Retinopathy Using Deep Learning Models

Early detection is essential for effective treatment of diabetic retinopathy, a leading cause of blindness and vision impairment worldwide. Algorithms for deep learning-based classification have shown great promise in automating DR diagnosis. This study demonstrates the effectiveness of deep learning ensemble methods in medical imaging and validates their potential for real-world clinical applications. This study compares majority voting, weighted averaging, and cascaded fusion procedures to improve generalization and reduce misclassification rates by leveraging the different capabilities of different models. The model outperformed the standalone classifiers when trained and evaluated on a carefully chosen set of retinal fundus images. For automated DR classification, this ensemble approach is a comprehensible solution. This study illustrates the effectiveness of various deep learning-based ensemble techniques, such as majority voting, weighted averaging, and cascaded ensembles, using the FGADR dataset. The cascaded ensemble performed the best among them, achieving a noteworthy accuracy of 92.23%, which was reinforced by high precision (88%), FI-score(90%) and recall (91%) metrics. The outcome shows that for a wide range of datasets, ensemble approaches improve accuracy and yield more precise predictions.