Deep Learning for Diabetic Retinopathy: A SqueezeNet-Based Approach with Data Augmentation for Enhanced Detection

Diabetic Retinopathy (DR) is a progressive retinal disease and a leading cause of blindness, necessitating early detection and precise classification for effective treatment. Manual diagnosis through fundus images is time-consuming and requires expert clinicians to identify intricate retinal features such as micro aneurysms, exudates, and hemorrhages. This study proposes a Convolutional Neural Network (CNN)-based approach leveraging SqueezeNet for automatic DR detection and severity classification. Our model is trained on a publicly available diabetic retinopathy dataset, utilizing real-time data augmentation techniques such as random rotations, horizontal and vertical flips, and random translations to improve feature generalization and enhance the model's ability to detect intricate retinal abnormalities. The training process follows a stochastic gradient descent (SGD) optimization strategy, with an initial learning rate of 0.0001, later increased to 0.0003 for improved convergence. Evaluation on a 5,000-image validation set demonstrates more than 85% accuracy, 89% specificity, and 80% sensitivity, with inference completed in 188 seconds. Additionally, a confusion matrix is generated to analyze model performance across different DR severity levels. While the model exhibits strong specificity, improvements in sensitivity remain a key focus for better early-stage DR detection. The results highlight the potential of deep learning-based automated DR screening systems in reducing clinical workload and enabling large-scale, cost-effective diagnosis, particularly in resource-limited settings.