Autoencoder-based Gaussian mixture model for diagnosing early onset of diabetic retinopathy

Abstract

The current study presents a simplified yet innovative solution towards effective early diagnosis of diabetic retinopathy (DR) that leads to irreversible blindness. A review of current literature shows a considerable number of machine learning and deep learning approaches have been presented; however, there are significant issues with the early detection of DR. Hence, the proposed study deploys a novel architecture using an autoencoder that extracts a hidden representation of retinal images while binary classification is carried out using a Gaussian mixture model. The prime contribution is the joint integration of deep learning with statistical modelling towards efficient feature extraction and anomaly detection, supporting early determination of DR. The study outcome shows a proposed system to significantly exhibit 96.5% accuracy, 94.2% sensitivity, and 98.3% specificity on two standard benchmarked datasets in comparison to existing models frequently used for the diagnosis of DR.