An Enhanced Segmentation and Deep Learning Architecture for Early Diabetic Retinopathy Detection
Renato R. Maaliw, Zoren P. Mabunga, Maria Rossana D. de Veluz, Alvin Sarraga Alon, Ace C. Lagman, Manuel B. Garcia, Luisito Lolong Lacatan, Rhowel M. Dellosa
Abstract
Diabetic retinopathy is a serious complication needing prompt diagnosis and medication to avert vision loss. Lesions caused by the condition are difficult to track because they are hidden behind the eye's structure in small and subtle forms. To extract relevant features., we created a robust pipeline using multiple preprocessing techniques., image segmentation architecture (DR-UNet) with atrous spatial pyramid pooling., and an attention-aware deep learning convolutional network with different modules based on ResidualNet. Empirical results show that our framework has segmentation accuracies of 87.10% (intersection over union) and 84.50% (dice similarity coefficient). Moreover., classification performance of 99.20% provided better results than existing schemes., as reinforced by the smooth convergence of training/validation loss and accuracy. This study has the potential to supplement traditional diagnosis to identify better the ailment in its early and advanced stages.