A Scalable Deep Learning Framework for Diabetic Retinopathy Classification with Web Integration and FPGA-Based Edge Deployment

Abstract: Diabetic retinopathy (DR) is a progressive microvascular complication of diabetes and a major cause of preventable vision loss. Early and accurate identification of DR severity is essential; however, conventional screening relies heavily on manual examination of retinal fundus images, making large-scale diagnosis time-consuming and dependent on specialist availability. This work presents an end-to-end framework for automated DR classification that combines deep learning, web-based accessibility, and FPGA-oriented hardware implementation. A customized residual convolutional neural network is developed to classify retinal fundus images into five severity levels: No DR, Mild, Moderate, Severe, and Proliferative DR. The model employs residual skip connections to improve feature propagation and support effective learning of pathological retinal characteristics. A preprocessing and augmentation pipeline is incorporated to standardize image representation, improve feature visibility, and address class imbalance. The proposed model achieves an overall classification accuracy of 97.2%, demonstrating strong performance in multi-class DR classification. For practical accessibility, the trained model is integrated into a Flask-based web application that enables users to upload retinal images and obtain automated predictions. To extend the framework toward edge-oriented deployment, the convolutional parameters are converted from floating-point to fixed-point integer representation and mapped to a custom FPGA accelerator on the Zynq-7000 XC7Z010-1CLG400 platform. The hardware architecture incorporates parallel 3×3 multiply–accumulate processing, bias addition, re-quantization, and ReLU activation, with parameter and data movement supported through the Zynq processing system, DDR memory, and AXI/DMA interfaces. The proposed framework therefore establishes a unified path from software-based DR classification to web-accessible screening and quantized FPGA implementation, providing a foundation for efficient edge-based medical image inference.

Keywords: Diabetic Retinopathy, Deep Learning, Residual Convolutional Neural Network, Fundus Image Classification, Multi-Class Classification, Quantization, FPGA, Zynq-7000, Edge AI, CNN Accelerator, Web-Based Screening.