Development of OCTA Automated Segmentation and Measurement Tools for Assessment of Cerebrovascular and Neural Health in Aging Populations

Abstract

Cerebral small vessel disease (CSVD) is one of the most commonly occurring vascular disease in the older population. Research has shown that the emergence of cerebrovascular disease has a profound effect on the changes that occurs in retinal vasculature. Optical coherence tomography angiography (OCTA) is a non-invasive, high resolution imaging modality used to extract images of retinal vasculature. However, manual segmentation of these images requires experienced clinical expert and is highly subjective to the individual's background. Therefore, there is a need of developing an automated segmentation technique for OCTA images. This research proposes a novel approach for automatic segmentation of OCTA images. The OCTA scans in this study involved individuals from two groups, older individuals whose scans were obtained from a publicly available dataset ROSE-1 [1], and OCT images of younger individuals that were acquired in the lab at University of Ottawa called NCM images. The proposed approach comprises of pixel-level and centerline level segmentation of OCTA images using Attention UNet model of different depths and merging these images to form a fully segmented superficial vascular complex (SVC) OCTA image. The experimental results demonstrates that the proposed approach (for fully segmented SVC OCTA) provides an accuracy of 0.9188, average dice of 0.7853, kappa score of 0.7354, G-mean score of 0.847 and balance accuracy of 0.85518. The trained pixel-level segmentation model was again used to segment NCM images, and the resulting segmented image was overlapped over the original image and studied. Subsequently, pixel-level segmented images were used to extract vascular features from retina such as vessel density, vessel length density, vessel perimeter index and vessel mean diameter for both the groups. It was found that there is a decrease in mean diameter of blood vessels (p-value < 0.05) in older individuals as compared to the younger group while the other parameters were not statistically significant (p-value > 0.05).