Comparative Analysis of Deep Learning Models Using Adaptive Epoch Training for Diabetic Retinopathy Classification

Diabetic retinopathy (DR) is a sound consequence of diabetes, leading to significant loss of vision worldwide. Early detection and classification are vital for effective management and treatment. This study evaluates the several deep learning model performances, including Sequential CNNs, InceptionV3, ResNet50, and VGG19, using an adaptive epoch training approach for the classification of DR severity. The training approach consists of an initial phase of 80 epochs followed by up to 20 additional epochs with early stopping to prevent overfitting. The proposed methodology addresses challenges such as inter-observer variability and the need for scalable solutions in healthcare. Our analysis reveals that pre-trained models show consistent performance improvements with extended training, while the custom CNN demonstrates significant gains. Experimental results depict the effectual output of our approach in achieving high accuracy and generalization, providing a promising tool which automates DR examination and testing.