UNIConv: An enhanced U‐Net based InceptionV3 convolutional model for DR semantic segmentation in retinal fundus images

Abstract

SummarySegmentation of diabetic retinopathy is a significant requirement for correct severity estimation level. Due to dissimilarity in shapes, size, and quantity of retinal lesions, the procedure of labor‐intensive type of grading becomes a more time‐consuming and challenging task. It imposes the prerequisite for an automated segmentation framework that can accurately identify the regions with margins and aid ophthalmologists in the early detection and monitoring of the severity level. The proposed work introduces a customized U‐Net model based on the InceptionV3 pre‐trained CNN network. It uses a periodic shuffling approach along with sub‐pixel convolution initialized to convolution nearest‐neighbor resize. The model performed training and validation for non‐proliferative diabetic retinopathy that is, microaneurysms, hemorrhages, hard and soft exudates on two standard datasets such as IDRiD and DIARETDB1. The proposed model achieves 99.82% accuracy, 88.39% sensitivity, and 99.94% specificity for hemorrhages, 99.95% accuracy, 70.76% sensitivity, and 99.98% specificity for microaneurysms, 99.83% accuracy, 85.9% sensitivity, and 99.95% specificity for hard exudates, along with 99.62% accuracy, 85.38% sensitivity, and 99.71% specificity for soft exudates segmentation on the IDRID dataset. In comparison with both datasets, the proposed system on the IDRID dataset performed accurately with superior state‐of‐the‐art outcomes for segmentation of the retinal lesions.