An improved Tasmanian Devil Optimization algorithm based EfficientNet in convolutional neural network for diabetic retinopathy classification

The accurate segmentation and identification of retinal blood vessels play a crucial role in detecting diabetic retinopathy (DR), an eye condition associated with diabetes that can lead to vision loss due to abnormalities in the blood vessels within the fundus. This paper introduces novel methods for DR detection. Initially, a database of input images is subjected to pre-processing, followed by applying a modified level set technique to segment the blood vessels. Once the specific region is segmented, texture and color features are extracted. Subsequently, a categorization of different stages of DR, including Severe Non-proliferative diabetic retinopathy (S-NPDR), Moderate NPDR (Mo-NPDR), and Mild NPDR (M-NPDR), is achieved using the proposed Improved Tasmanian Devil Optimization (ITDO) algorithm based on the EfficientNet in convolutional neural network (EN-CNN). The implementation of this work is carried out using Python 3.6 software for DR classification. The results demonstrate exceptional performance, surpassing previous methods with an accuracy of 99.09%, sensitivity of 98.78%, specificity of 98.89%, and a computational time of 09.67 s. The proposed technique exhibits superior classification performance compared to existing methods.