EfficientNetB0-Based Automated Diabetic Retinopathy Classification in Fundus Images

Abstract

The most common cause of blindness among diabetics is diabetic retinopathy. This disease must be detected in its early stages as delaying treatment can result in permanent blindness. In today's modern world with the development of technologies, automated techniques have improved the accuracy and efficiency of detecting and classifying this disease. This study suggests a deep learning model utilizing EfficientNetB0 that focuses on designing a neural network structure that requires fewer parameters and computations compared to traditional architectures. Further to help in increasing the accuracy, the image is being processed with methods like CLAHE which is an image processing algorithm while keeping the noise of the images as low as possible. Also, augmentation is being done on the images and SMOTE is applied to handle class imbalance. We collected the APTOS dataset consisting of retinal fundus images. The accuracy that is being obtained is 98.78% indicating the model's effectiveness in detecting and classifying between the various stages of diabetic retinopathy.