Diabetic Retinopathy Grading Using 3D Multi-path Convolutional Neural Network Based on Fusing Features from OCTA Scans, Demographic, and Clinical Biomarkers

Abstract

Diabetic Retinopathy (DR) is considered one of the major reasons for vision loss in the working-age population in most of the countries. DR is caused by high blood sugar levels (diabetes), which damages retinal blood vessels and leads to blindness. Both diagnosis and grading of DR require manual measurements and visual assessment of the changes that happen in the retina, which is a highly complex task. Thus, there is an unmet clinical need for a non-invasive and objective diagnostic system, which can improve the accuracy of both early signs and grading detection for DR. In this paper, we proposed a computer-aided diagnosis (CAD) system for detecting early signs as well as grading of DR. Four significant retinal vasculature features are extracted from optical coherence tomography angiography (OCTA) scans, which reflect the changes in the retinal blood vessels due to DR progress. The developed system fuses these four significant features with clinical and demographic biomarkers. The proposed system uses a 3D convolutional neural network (CNN) to segment blood vessels from both OCTA deep and superficial plexuses. Finally, these extracted features are classified by using the random forest (RF) technique to differentiate first between the DR from normal subjects. Then, grade the DR subjects into mild or moderate. Our preliminary results of grading DR in a cohort of patients (n == 100) demonstrated an average accuracy of 96.8%, sensitivity of 98.1%, and specificity of 88.8%. These results show the feasibility of the proposed approach in early detection as well as the grading of DR.