Fundus Image Generation and Classification of Diabetic Retinopathy Based on Convolutional Neural Network

Abstract

To detect fundus diseases, for instance, diabetic retinopathy (DR) at an early stage, thereby providing timely intervention and treatment, a new diabetic retinopathy grading method based on a convolutional neural network is proposed. First, data cleaning and enhancement are conducted to improve the image quality and reduce unnecessary interference. Second, a new conditional generative adversarial network with a self-attention mechanism named SACGAN is proposed to augment the number of diabetic retinopathy fundus images, thereby addressing the problems of insufficient and imbalanced data samples. Next, an improved convolutional neural network named DRMC Net, which combines ResNeXt-50 with the channel attention mechanism and multi-branch convolutional residual module, is proposed to classify diabetic retinopathy. Finally, gradient-weighted class activation mapping (Grad-CAM) is utilized to prove the proposed model’s interpretability. The outcomes of the experiment illustrates that the proposed method has high accuracy, specificity, and sensitivity, with specific results of 92.3%, 92.5%, and 92.5%, respectively.