Abstract

This research proposes a hybrid strategy that combines light weight models of deep learning with extensive pre-processing methodologies to improve the categorization and identification of diabetic retinopathy (DR). DR images have been resized during the pr-processing stages and suitable filters algorithms have been used to eliminate both salt-and-pepper Gaussian noise. The selected filtering method ensures effective smoothing of homogeneous regions while preserving critical edge details, with its performance evaluated using Structure and Edge Preservation Indices. A lightweight pre-trained network is employed for classification, ensuring computational efficiency without compromising Accuracy. Extensive experimentation demonstrates that Shuffle Net achieves a remarkable classification accuracy of 96.33% on a combined dataset. These findings highlight the potential impact of the proposed hybrid strategy for enhancing and automating DR detection, paving the way for scalable and accurate diagnostic tools in medical imaging, and potentially improving the lives of millions affected by diabetic Retinopathy.