Computer-aided diagnosis system based on a comprehensive local features analysis for early diabetic retinopathy detection using OCTA

Abstract

In this chapter, our aim is to generate a computer-aided diagnosis system to detect the early signs of diabetic retinopathy (DR). The system is based on extracting some features that describe the appearance and shape of the vascular system that can be found in optical coherence tomography angiography (OCTA) scans. In the beginning, the system segments two different retina plexuses, such as superficial and deep plexuses, to separate the blood vessels from the background tissues. Then, four different features were extracted from the segmented images. Blood vessel density, blood vessel caliber, distance map of the foveal avascular zone, and bifurcation and crossover points were extracted from both OCTA plexuses. Finally, these four features were used for the diagnosis of early signs of DR using two-stage random forest classifier. A total of 133 scans for normal and DR subjects were used to evaluate the performance of our system. Using twofold cross-validation, our system obtains an accuracy of 97%, which is a promising sign for the effectiveness of our proposed system.

keywords: Computer-aided diagnosis, Diabetic retinopathy, Feature extraction, Foveal avascular zone, Optical coherence tomography angiography, Random forest