Classification of Diabetic Retinopathy Using ShuffleNet V2 and Real-ESRGAN with CLAHE Image Enhancement

Abstract

Diabetic retinopathy (DR) is a microvascular complication of diabetes that can lead to blindness if not detected and treated early. Manual DR grading from fundus images is time-consuming and highly dependent on expert availability, motivating the need for automated and efficient decision-support systems. This study proposes a lightweight DR severity classification model using ShuffleNet V2 combined with a preprocessing pipeline consisting of Contrast Limited Adaptive Histogram Equalization (CLAHE) and Real-ESRGAN-based super-resolution. Unlike prior works that mainly employ these enhancement techniques with deeper or computationally expensive networks, this study explicitly investigates their synergistic integration with ShuffleNet V2 to improve lesion visibility while preserving computational efficiency for resource-constrained environments. Experiments conducted on the APTOS 2019 dataset demonstrate that the proposed combination significantly improves classification performance, achieving a best accuracy of 90.70%, with balanced precision, recall, and F1-score when optimized using Adam. Comparative analysis with SGD optimizer further reveals a trade-off between accuracy and inference speed. The results confirm that combining CLAHE and Real-ESRGAN with ShuffleNet V2 offers an effective and efficient solution for automated diabetic retinopathy grading, highlighting its suitability for large-scale screening and low-resource clinical deployment