Abstract
Diabetic retinopathy (DR) is a common consequence of diabetes mellitus resulting in vision-impairing lesions on the retina. Treatment of DR in its early stages can extensively minimize the chance of blindness. Diverse machine learning approaches were developed for DR detection; however, the classical models may create certain limitations including overfitting issues, data requirements, and vanishing gradient problems. To mitigate these shortcomings, this research proposed a Wolf social leader algorithm-enabled Bi-directional long short-term memory (WS-BiLSTM) for DR detection. The integration of a weighted shape-based texture pattern enhances the capability of the model to extract pertinent texture and shape features. Additionally, the ResNet 101 model obtains the informative regions from the fundus images which leads to attaining better performance. The statistical features extracted from the input fundus images enhance the robustness of the framework. The hyperparameters of the WS-BiLSTM model are optimized using the suggested Wolf social leader algorithm, which imitates the social dynamics of American jackals and the hunting characteristics of gray wolves. In addition, the model improves the performance effectively with high detection performance and achieved accuracy, sensitivity, and specificity of 96.32%, 97.21%, and 95.42% compared to other convolutional methods.