Hybrid Approach—Diabetic Retinopathy Classification Through Activation Function Optimization

Another common disease resulting from diabetes is called diabetic retinopathy, which requires a well-defined and effective classification so that medical decisions could be made. In this research effort, our work embarks on an in-depth exploration of diabetic retinopathy classification using deep learning techniques. Our study starts by focusing extensively on data preprocessing that involves activities like standardization of image dimensions as well as employing data augmentation strategies to enhance the quality and variation of our dataset. This paper highlights studies about different deep learning models for diabetic retinopathy. EfficientNetB0 has 92.0% accuracy which demonstrates how good it is in detecting retinopathy due to diabetes. DenseNet performed strongly at 88.0%. Remarkably, this Hybrid Model surpasses all others with an impressive 96.0% accuracy through combining activation function optimization with architectural improvements. It shows how innovative methods can greatly improve the diagnosis of this eye condition referred to as diabetic retinopathy’s capabilities. To begin with, the crux of our research is on the examination of activation functions via detailed analysis on five prominent ones: Rectified Linear Unit (ReLU), Leaky ReLU, Hyperbolic Tangent (Tanh), Sigmoid, and Swish. Empirically, it is suggested that ReLU and Sigmoid are always better than other functions, hence culminating into an accuracy of over 96% for the ‘ReLU-Sigmoid’ combination. Furthermore, it was revealed by this study that there is a substantial reduction in error rates, thereby emphasizing on improving diagnostic accuracy in models. Three distinguished models are considered for differentiating architectural variability of neural networks: EfficientNetB0, DenseNet, and a new Hybrid Model, which melds EfficientNetB0 and DenseNet architectures together to function as one entity. With a performance peak of 96% accuracy rate coupled with remarkably low error rates averaging 4%, it is unrivaled among all other approaches. These findings have immense implications for diabetic retinopathy classification by emphasizing the importance of activation functions and architectural design choices. From this viewpoint, the paper serves as guiding light for medical image analysts through insights, methodologies, and inventions that can transform patient care and diagnostic precision in ophthalmic healthcare. The use of deep learning methods calls for a special graphics processing unit (GPU) designed to develop and evaluate classification model (Ishtiaq et al. in Multimed Tools Appl, 79:15209–15252, 2020).