Optimizing Deep Learning for Diabetic Retinopathy Diagnosis

The detection of diabetic retinopathy traditionally requires the expertise of medical professionals, making manual detection both time- and labor-intensive. To address these challenges, numerous studies in recent years have proposed automatic detection methods for diabetic retinopathy. This research focuses on applying deep learning and image processing techniques to overcome the issue of performance degradation in classification models caused by imbalanced diabetic retinopathy datasets. It presents an efficient deep learning model aimed at assisting clinicians and medical teams in diagnosing diabetic retinopathy more effectively. In this study, image processing techniques, including image enhancement, brightness correction, and contrast adjustment, are employed as preprocessing steps for fundus images of diabetic retinopathy. A fusion technique combining color space conversion, contrast limited adaptive histogram equalization, multi-scale retinex with color restoration, and Gamma correction is applied to highlight retinal pathological features. Deep learning models such as ResNet50-V2, DenseNet121, Inception-V3, Xception, MobileNet-V2, and InceptionResNet-V2 were trained on the preprocessed datasets. For the APTOS-2019 dataset, DenseNet121 achieved the highest accuracy at 99% for detecting diabetic retinopathy. On the Messidor-2 dataset, InceptionResNet-V2 demonstrated the best performance, with an accuracy of 96%. The overall aim of this research is to develop a computer-aided diagnosis system for classifying diabetic retinopathy.