Abstract
Purpose :
To visualize 3-D organization of retinal vasculature in humans using single OCT/A scan.
Methods :
OCT/A scans (3×3-mm, 400×400 A-lines) were collected using a PLEX® Elite 9000 swept-source OCT (ZEISS, Dublin, CA) at various retinal regions from healthy subjects. OCTA tail artifacts were removed by an advanced projection-removal algorithm. A ResNet-based generative adversarial network was designed to reconstruct the single scan OCTA volume, using merged (N=10) volumes as ground truth. For the training process, we proposed a percent retina projection strategy, i.e., to segment retinal layers and project multiple en face images at each percent depth slab, to ensure the network can better learn the retinal vasculature pattern. After training, the network was applied to reconstruct every depth plane to obtain the high-definition OCTA volume, requiring no layer segmentations. The reconstruction performance of the proposed strategy was compared to conventional depth plane and B-scan training schemes using the same network. Retinal arterioles and venules in OCTA images were correlated to the fundus image to investigate the organization of retinal vasculature.
Results :
Thirty OCT/A scans from 10 subjects were included in this study. Five scans were used for training and the others were used for testing. Our proposed strategy achieved the best structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS) among three strategies by comparing the en face images of superficial capillary plexus (SCP), intermediate capillary plexus (ICP), and deep capillary plexus (DCP) (Fig. 1). From en face and volume visualization of the reconstructed OCTA volumes, we were able to confirm three features of retinal vasculature: 1) retinal arterioles typically have avascular zones and longer vascular bifurcation length; 2) capillary vortices frequently exist in DCP, especially at regions near fovea; and 3) retinal venules tend to extend to capillary vortices in DCP (Fig. 1 and 2).
Conclusions :
The proposed deep learning reconstruction method enables detailed 3-D visualization of retinal vasculature using single OCT/A scan, which may pave a way to study their complex organization, pathogenesis and treatment response in various retino-vascular diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.