Convolutional Neural Networks-Based Deep Learning for Diabetic Retinopathy Detection

Abstract

Diabetic retinopathy (DR) is a major complication of diabetes that can cause permanent vision loss, affecting about 35% of people with type 2 diabetes worldwide. However, existing diagnostic models often struggle with class imbalance and limited generalizability across diverse real-world datasets. Early detection is crucial, yet manual screening is time-consuming and depends on expert assessment. This study develops an automated DR diagnostic system using deep learning to classify fundus images by severity. The model uses an EfficientNetB3 CNN pretrained on ImageNet, combined with CLAHE preprocessing to enhance image contrast. The preprocessing steps include resizing, CLAHE, normalization, and data augmentation (±20° rotation, horizontal flipping, and ZCA whitening). The dataset is the Gaussian-filtered APTOS 2019 set, consisting of 2,750 images across five DR levels (0–4). The model achieved 95% training accuracy and 75% validation accuracy, with overfitting observed after epoch 14. While training performance was high, evaluation metrics (Precision, Recall, F1-Score, and AUC) indicate the need for early stopping or regularization to improve generalization. Overall, CNN-based deep learning can effectively automate DR detection, though further optimization is required for better performance on unseen data. Clinically, this automated pipeline offers a reliable decision-support tool to prioritize high-risk patients for immediate ophthalmological review