Classification of Diabetic Retinopathy disease levels by extracting topological features using Graph Neural Networks

Diabetic retinopathy happens due to damage in blood vessels and is the prominent reason for blindness worldwide. Clinical experts observe the fundus images to diagnose the disease, but it is often an error-prone and tedious task. Computer-assisted techniques will help clinicians to detect the disease severity levels. In medical imaging, experiments of automated diagnosis using CNN produce impressive results. Even though disease classification tasks in retinal images via CNN face difficulty in retaining high-quality information at the output. A new deep learning methodology is proposed based on a graph convolutional neural network (GCNN). The proposed model aims to extract the essential retinal image features effectively. The work focuses on extracting the features using a Variational autoencoder and identifying the underlying topological correlations using GCNN. The experiments are carried out using two datasets: Kaggle and EyePACS datasets. The performance of the proposed model is evaluated using accuracy, U-kappa, sensitivity and specificity metrics. The results outperform when compared with other state-of-the-art techniques.