Redefining Image Transformations with Context Aware Methodologies to Elevate Deep Learning Training Outcomes

Diabetic Retinopathy (DR) is the major reason for vision loss, necessitating early and precise detection for accurate diagnosis. In this study, we investigate the performance of NFNet and compare it against an ensemble of MobileNetV1 & MobileNetV2, as well as their standalone versions, for DR classification. While MobileNet architectures offer computational efficiency, ensembling them allows for a balanced trade-off between speed and accuracy. Additionally, we employ context-aware data augmentation to enhance model generalization across varying DR severity levels. Our results show that NFNet achieves the highest classification accuracy of 92%, outperforming the MobileNetV1 & V2 ensemble (80%), demonstrating its superior feature extraction capabilities. Due to the lower computational expense of MobileNet ensemble, it is the most suitable option for real-time applications. This study highlights the speed vs. accuracy trade-off in DR classification and lays the groundwork for future improvements through self-supervised learning and hybrid ensemble techniques