Enhanced Robustness in Early Multi-Class Detection of Diabetic Retinopathy Across Datasets Using EfficientNet and CLAHE

Abstract

Diabetic retinopathy (DR) is a severe complication of diabetes mellitus (DM) that can lead to vision loss. DR has a high prevalence among DM patients and is one of the leading causes of preventable blindness. Early detection of DR is crucial to prevent adverse effects on vision in later stages. This study utilizes EfficientNet and CLAHE methods to detect DR levels based on fundus images. Experiments were conducted on three datasets: DDR, IDRiD, and Messidor. These datasets contain five stages of DR: normal, mild, moderate, severe, and proliferative diabetic retinopathy (PDR). Our proposed model achieved a validation accuracy of 85.97% and a testing accuracy of 86.16%, outperforming other models such as Inception-ResNet-v2 (82.18%), Inception-V3 (78.79%), and ResNet-50 (74.32%). The validation and testing accuracy values above 85% indicate that the model can accurately predict labels. Interestingly, when a mixed dataset was used, the testing accuracy decreased. This decline in accuracy may be due to increased data variability, inconsistent preprocessing, and differences in image quality. Nonetheless, this model significantly contributes to preventing severe complications and vision loss due to DR.