Diabetic Retinopathy Detection using the Genetic Algorithm and a Channel Attention Module on Hybrid VGG16 and EfficientNetB0

Abstract

Diabetic Retinopathy (DR), a result of diabetes, requires early detection to reduce the impact of the disease on vision. This study introduces a new system whose architecture is based on a combination of VGG16 architecture with EfficientNetB0 as well as an added body structure, which is the Channel Attention Module (CAM), to strengthen the channel maps and thus achieve improved classification accuracy. For further efficiency and consistency, the system employs a genetic algorithm for image normalization. The system shows great potential for improving clinical decision making and patient examination results when used in the diagnosis of DR. The evaluation results confirm the reliability of the system and the feasibility of using it in daily practice to address the acute challenge of early detection of DR. The model is well trained with a test dataset of 2900 images and demonstrates high accuracy of 95%. This high accuracy clearly shows the high reliability of the proposed hybrid model which is also confirmed by the precision and recall values. The achieved precision is 0.96 for class 0 and 0.94 for class 1, and the achieved recall is 0.94 for class 0 and 0.97 for class 1.