Classification of Diabetic Retinopathy Using Efficientnet-B7 with Hyperparameter Optimization

Diabetic Retinopathy (DR) is a microvascular complication of Diabetes Mellitus that results from retinal blood vessel damage and can lead to permanent blindness. Early detection is essential to prevent disease progression, yet DR is frequently identified only in advanced stages. This study focuses on classifying DR severity levels from retinal fundus images using the EfficientNet-B7 Convolutional Neural Network (CNN) architecture combined with Hyperparameter Optimization (HPO). Experiments were conducted by varying data splits, dense layer sizes, and learning rates on a balanced and augmented Kaggle DR dataset. The best training performance was obtained with a 90%-10% data split, 256 dense units, and a 0.01 learning rate, achieving 95.48% accuracy. The best testing performance used a 90%-10% data split, 32 dense units, and a 0.001 learning rate, reaching 95.81% accuracy. These findings indicate that EfficientNet-B7 with optimized hyperparameters can significantly improve DR classification accuracy and offers a promising automated approach for early DR detection.