Deep learning-based automated detection of diabetic retinopathy for early diagnosis

Early detection and accurate staging of diabetic retinopathy (DR) are essential to prevent irreversible vision loss. In this work, a systematic and practical evaluation of deep learning models for automated DR detection and staging using retinal fundus images is presented. Rather than proposing a new network architecture, this study investigates the performance of three widely used deep learning models Convolutional Neural Network (CNN), AlexNet, and SqueezeNet within a unified experimental framework. The models are evaluated for both binary classification (normal vs. DR) and multi-class DR staging using identical preprocessing steps, dataset partitioning strategies, and evaluation metrics. A comprehensive performance analysis is conducted across multiple training ratios (70%, 80%, and 90%) and optimization algorithms (SGDM, RMSprop, and Adam) to examine model stability and generalization behavior. Experimental results demonstrate that AlexNet and SqueezeNet consistently outperform conventional CNNs in terms of classification accuracy and F1-score. In particular, SqueezeNet achieves diagnostic performance comparable to deeper architectures, indicating its potential suitability for practical diabetic retinopathy screening applications. The findings provide a deployment-oriented perspective on model selection for deep learning based DR diagnosis and staging.