Enhancing Diabetic Retinopathy Detection with Image Processing Techniques and Deep Learning EfficientNetB0: A Comparative Analysis

Diabetic retinopathy (DR) is a major global cause of blindness, underscoring the critical need for early and accurate detection. This study addresses this challenge by conducting an in-depth comparative analysis of various image processing techniques coupled with the state-of-the-art EfficientNetB0 model for DR identification. The researchers employed retinal images from the Asia-Pacific Teleophthalmology Society (APTOS) 2019 Blindness Detection competition on Kaggle, a valuable dataset for this purpose. The study considers five image processing methods: Median Subtraction, Ben Graham's Filter, Adaptive Histogram Equalization, Histogram Normalization, and Selective Gamma Correction. These techniques aim to enhance image quality and information content, which is crucial for training the model effectively. The study's primary performance metric is accuracy, recall, precision, and F1 score, which measures the model's ability to classify images correctly. The outcomes reveal that the choice of image pre-processing techniques significantly influences the model's performance. In summary, the study demonstrates that by combining advanced image pre-processing methods with a powerful deep learning model, Ben Graham's Filter can notably enhance diabetic retinopathy detection performance with 86% accuracy. Such advancements are instrumental in the early diagnosis and intervention of DR, thereby mitigating the risk of blindness in affected individuals.