Comparative Analysis of Deep Learning Models for Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a global disease that leads to imaginative and foresighted impairment and is considered the largest cause of it. The progress from No DR to Proliferative DR is critical in portraying the progression of the disease, and hence, therefore classification of those stages is very crucial in its timely diagnosis and treatment. The paper compares the performance of four deep learning models: VGG16, VGG19, Inception V3, and InceptionResNetV2 for the classification of DR in its five stages. VGG16 and VGG19 had an excellent performance when it came to classifying the early levels but poor in the more complex tiers of high-level classification. Inception V3, with its advantage on a multi-scale function extraction, achieved good results in the advanced class and was classified under both Moderate and Severe DR but failed for the most challenging cases. The model, InceptionResNetV2, combined the Inception modules with residual connections and outperformed all alternative fashions by achieving more than 96% accuracy in all tiers while showing top accuracy in Severe DR at 98.05%. The results showed the importance of multi-scale function extraction and residual learning for effective DR classing, especially for better tier differentiation.