Detection of stages in diabetic retinopathy using computer aided ensemble network

Abstract

Diabetic retinopathy (DR) is one of the progressive micro vascular disorders of diabetes and a major cause of preventable blindness in the world. Manual ophthalmologist evaluation is costly in terms of time and more likely to have inter-observer error, whereas current automated methods tend to fail to differentiate between intermediate stages of DR; yet, proper assessment of DR severity is critical to its successful intervention. This paper suggests a framework of hybrid ensemble deep learning (DL) model that incorporates VGG16, InceptionV3 and ResNet50 based on a weighted feature fusion model and a feature-scaled parametric activation (FSPA) model to maximize inter-classes separability. The Kaggle EyePACS dataset containing 35,126 retinal fundus images was used. Image normalization, contrast enhancement, and data augmentation improved robustness to class imbalance. The overall accuracy of the proposed ensemble was 95.0% and the sensitivity and specificity were 92.0 and 94.0 respectively and the quadratic weighted Kappa (QWK) was 0.91, with up to +5% improvements in accuracy over individual convolutional neural network (CNN) baselines. Although there are still limitations to differentiating moderate versus severe DR, the findings demonstrate that the proposed framework enables consistent and reliable DR grading for large-scale clinical screening.