EfficientNet-B0 Convolutional Network for Early Detection and Severity Classification of Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness worldwide. Early detection is essential to reduce the risk of vision loss; however, screening programs are often constrained by a shortage of qualified specialists. This study implemented a lightweight EfficientNet-B0 architecture for the automated classification of five diabetic retinopathy severity levels using a dataset of 3,662 color fundus images. The proposed methodology incorporated transfer learning, an 80:20 stratified data split, and a WeightedRandomSampler to address severe class imbalance (9.4:1 ratio between the majority and minority classes). The model was trained using the Adam optimizer with a learning rate of 1×10^-4 and an early stopping mechanism. Experimental results showed an overall accuracy of 84% and a macro-average AUC of 0.94, demonstrating strong discriminative capability. The highest performance was achieved in the No_DR class (F1-score = 0.98) and the Moderate class (F1-score = 0.79), whereas the Severe and Proliferative_DR classes remained more challenging because of their limited number of training samples.