Fused information of DeepLabv3+and transfer learning model for semantic segmentation and rich features selection using equilibrium optimizer (EO) for classification of NPDR lesions

Diabetic retinopathy (DR) is the leading cause of blindness amongst diabetics. The incidence and development of vision impairment may be reduced if detected at an early stage. Deep learning (DL) algorithms are more effective in detecting DR. Early stage detection of DR is the main objective of this study. The proposed method comprises two phases: segmentation and classification. In phase I, the semantic segmentation model is named the Javeria Segmentation (JSeg) model, in which the ResNet-50 model is used as the backbone of Deeplabv3+. The proposed segmentation model is trained on selected hyperparameters, such as 32 batch size, 250 training epochs, and Adam optimizer solver with eight down-sampling factors. The model provided up to a 0.90 meanIoU. In phase II, N x 1000 features were extracted using ResNet-101, out of which N x 262 optimum features were selected based on the equilibrium optimizer (EO). The DR lesions are classified into grades 1, 2, 3, and 4 using selected kernels of the SVM and neural network classifiers. The proposed classification model provides a meanROC of 0.98 & PLUSMN; 0.02 on a Narrow kernel of the neural network and 0.97 & PLUSMN; 0.01 on the Fine Gaussian kernel of the SVM classifier. Compared with the most recently published works, the proposed method provides superior results. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Diabetic retinopathy (DR) is the leading cause of blindness among diabetics, and early detection can reduce vision impairment. This study proposes a deep learning-based method for early stage detection of DR, using segmentation and classification phases. The results demonstrate superior performance compared to recent studies.