Diabetic Retinopathy Detection and Classification Using Deep Learning

Diabetic Retinopathy (DR) remains a significant cause of vision impairment worldwide, requiring accurate and timely severity assessment to prevent irreversible damage. Al- though automated deep learning systems have improved retinal image analysis, reliable multi-stage classification remains challenging due to variability in image quality and class distribution across severity levels. Conventional convolutional neural net- work architectures have demonstrated promising results however, achieving consistent performance across all DR stages while maintaining computational efficiency continues to be an active re- search problem. To address these challenges, this study proposes a transfer learning-based framework utilizing EfficientNet-B3 for five-class DR severity classification. EfficientNet-B3 employs com- pound scaling to balance network depth, width, and resolution, enabling effective feature extraction with optimized parameter utilization. The model is fine-tuned using ImageNet pretrained weights and evaluated on the APTOS 2019 dataset. Experimental results demonstrate strong validation performance and high ordinal agreement across severity categories. The trained model is further integrated into a Streamlit-based application to support real-time clinical screening. The findings indicate that the proposed approach provides a computationally efficient and practically deployable solution for automated DR severity assessment.