Automated Five-Class Diabetic Retinopathy Detection Using Transfer Learning and Web-Based Clinical Deployment

Abstract

Diabetic Retinopathy (DR) is a serious eye condition caused by diabetes and is a major contributor to vision impairment across the world. Detecting the disease at an early stage is essential to prevent permanent vision loss, but large-scale screening is often difficult due to the limited availability of specialists and the high cost involved. This study proposes a lightweight system for classifying DR into five severity levels using transfer learning and Convolutional Neural Networks (CNNs). The solution is integrated into a web-based application developed using the Django framework, allowing easy access and practical usage. Retinal fundus images are processed through a standardized pipeline where images are resized to 224×224 pixels and normalized to a suitable range for model input. The trained model produces probability scores for each class, and the results are presented through a visual confidence chart to improve interpretability. The system is designed to run efficiently on standard CPU-based devices, making it suitable for deployment in resource-limited environments. Experimental results indicate that the model performs consistently across all five classes, making it useful for preliminary screening. Overall, the proposed approach provides an accessible and efficient solution to support early detection of diabetic retinopathy.

Introduction

The text discusses Diabetic Retinopathy (DR), a serious complication of diabetes that can cause blindness if not detected early. Since early stages are often asymptomatic, regular screening is important, but traditional manual diagnosis is time-consuming and inconsistent.

To address this, the study proposes an automated DR detection system using deep learning. The system employs a lightweight CNN with transfer learning and a Django-based web interface to classify retinal images into five stages of DR (from No DR to Proliferative DR), providing more detailed diagnosis than binary models.

The system follows a three-tier architecture (frontend, backend, and model layer). Images are preprocessed (resized, cropped, normalized) and passed to a MobileNet-based CNN that generates predictions. It also includes confidence visualization using probability charts to improve interpretability for clinicians.

Implementation uses tools like TensorFlow, Django, and SQLite, and is designed to run on standard hardware without requiring GPUs. Results show high accuracy and balanced performance across all five classes, with fast inference time and low resource usage.

Advantages include efficiency, accessibility, interpretability, and suitability for low-resource settings. However, limitations include dependence on image quality, limited dataset diversity, and lack of detailed lesion-level analysis.

Conclusion

This work introduces an efficient system for detecting and classifying Diabetic Retinopathy into five severity levels using a transfer learning-based CNN model. The proposed solution combines image preprocessing, model prediction, and result visualization within a Django-based web application, enabling a smooth and user-friendly workflow for clinical use. The model demonstrates stable performance across multiple classes while maintaining low computational requirements, making it suitable for execution on standard hardware without the need for specialized resources. By providing clear prediction outputs along with confidence visualization, the system enhances interpretability and supports decision-making. Overall, the proposed approach offers a practical and scalable tool that can aid early detection of Diabetic Retinopathy and support screening efforts, especially in environments with limited medical infrastructure.

References

[1] International Diabetes Federation, IDF Diabetes Atlas, 2021. [2] World Health Organization, “Global report on diabetes,” 2016. [3] T. Y. Wong et al., “Diabetic retinopathy,” Lancet, 2016. [4] Y. LeCun et al., “Deep learning,” Nature, 2015. [5] A. Krizhevsky et al., “ImageNet classification with deep CNNs,” NIPS, 2012. [6] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE TKDE, 2010. [7] T. Walter et al., “Automatic detection of exudates,” IEEE TMI, 2002. [8] M. Osareh et al., “Automated detection of diabetic retinopathy,” IEEE TMI, 2003. [9] C. Gardner et al., “Neural network approach,” Br J Ophthalmol, 1996. [10] V. Gulshan et al., “Development of deep learning algorithm,” JAMA, 2016. [11] K. Simonyan and A. Zisserman, “Very deep CNNs,” 2014. [12] K. He et al., “Deep residual learning,” CVPR, 2016. [13] C. Szegedy et al., “Going deeper with convolutions,” CVPR, 2015. [14] A. Howard et al., “MobileNets,” 2017. [15] J. Pratt et al., “CNNs for diabetic retinopathy,” 2016.

Copyright

Copyright © 2026 Aravinda Karri, M. Radhika Mani. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.