A 2D-Structured Dilation Based Hierarchical CNN for the Detection of Diabetic Retinopathy Grade Levels

Abstract

This paper introduces a diabetic retinopathy severity grading approach thatuses a 2D-structured dilation-based hierarchical convolutional neural network (CNN). In this approach, the pre-processed fundus image is utilized to segment the regions such as the optic disc, blood vessels, and lesion regions. The optic disc, blood vessel, and lesion regions combine to form the region of interest. The proposed 2D-structured dilation-based Hierarchical CNN (2D-SDHCNN) has a parallel section of Lstages that use different dilated masks in a hierarchical structure. The dilation is also applied to the convolutional filters of each subnetwork and the region that corresponds to the dilated mask of the global feature is also utilized in the hierarchical network. The hierarchical network can able to extract deep features near the vessels and lesion candidates. Datasets such as Kaggle APTOS and Messidor-2 are utilized for evaluating the suggested 2D-SDHCNN approach. The suggested approach performance highly depends on the dilation factor used in the 2D-SDHCNN. The 2D-SDHCNN approach yields a precision, Mathews correlation coefficient (MCC), and accuracy of 97.61%, 97.03%, and 97.73% respectively when evaluated using the Kaggle APTOS dataset. Also, the suggested scheme when evaluated utilizing Messidor-2 provides precision, MCC, and accuracy of 93.30%,93.42%, and 95.39% respectively.