Enhanced retinal image captioning system for identifying and differentiating severity levels of diabetic retinopathy

Abstract

The objective of this research was to formulate a clinical decision support framework leveraging AI towards utilizing retinal fundus images for the identification and categorization of the four distinct stages of diabetic retinopathy, namely proliferative, severe, moderate, and mild. The devised system architecture integrated Long Short-Term Networks (LSTM), Generative Adversarial Networks (GAN), and pre-trained convolutional neural network (CNN) models. Following an exhaustive performance analysis, the most optimal image captioning model was identified and recommended to ophthalmologists for the purpose of identifying and categorizing diabetic retinopathy. Notably, the results revealed that employing ResNet50 with LSTM, in conjunction with enhanced retinal images, yielded superior accuracy of 0.975. The proposed methodology holds transformative potential for the realm of diabetic retinopathy diagnosis and classification, facilitating early detection and intervention to mitigate vision loss in individuals affected by diabetes.