An OCTA based diagnosis system based on a comprehensive local features analysis for early diabetic retinopathy detection

In this work, we propose a comprehensive diagnosis system for detecting early signs of diabetic retinopathy (DR). The proposed system is based on extracting features that are describing both shape and appearance of the retinal vascular system from optical coherence tomography angiography (OCTA) scans. First, two various OCTA plexuses are segmented, which are retinal superficial and deep plexuses, to extract the retinal blood vessels from the other background tissues. Then, the developed system calculates the blood vessels density, blood vessels caliber, and distance map of the foveal avascular zone (FAZ) from both two segmented OCTA plexuses. Also, the large blood vessels are extracted from the segmented superficial plexus to retrieve the skeleton of the blood vessels. The skeleton of the large blood vessels is used to calculate bifurcation, crossover, and branch points. Finally, the extracted four features that are capturing shape and appearance of the segmented vessels and FAZ, which are blood vessels density, blood vessel caliber, the width of the FAZ, and different types of vascular bifurcation points are used to diagnosis the OCTA images by using two-stage random forest (RF) classifier. To measure the performance of the proposed system, 133 OCTA scans for different patients are used for training and testing based on k-fold cross-validation technique. The performance of the proposed system is measured by using five various metrics. A promising average result of overall accuracy (ACC) of 97% is obtained that can differentiate normal from mild DR cases.