Image enhancement combined with EfficientNet-B7 for grading classification of diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is a complication caused by poorly managed diabetes that affects the eyes. According to the World Health Organization (WHO), 422 million people worldwide have suffered from DR in the past ten years. Manual detection using retinal fundus images is time-consuming and requires experienced ophthalmologists. This study proposes a deep learning method using the pre-trained model EfficientNet-B7 to identify this disease automatically. Five levels of DR will be classified: no-DR, mild-DR, moderate-DR, severe-DR, and proliferative-DR. The model was trained using "APTOS 2019 blindness detection" dataset, and image augmentation was performed. Image segmentation techniques such as contrast limited adaptive histogram equalization (CLAHE) and real enhanced super resolution generative adversarial network (Real-ESRGAN) were applied during preprocessing to improve the model's accuracy significantly. The implementation of CLAHE resulted in the validation accuracy improvement from 76.6% to 83.4% compared to no segmentation, while the combination of Real-ESRGAN and CLAHE increased the accuracy to 93.7%. Future research can explore the combination of CLAHE with other image processing techniques apart from the Real-ESRGAN model.