Abstract
Purpose :
Lipid emulsion-based optical coherence tomography angiography (LE-OCTA) allows for 2D and 3D visualization of the intrascleral aqueous outflow tract in porcine and human eyes. Prerequisite for further analysis of the acquired data is the segmentation of the intraluminal signal from unspecific background signal and artefacts. In this study we compared different segmentation approaches.
Methods :
Five LE-OCTA datasets were acquired from freshly enucleated porcine eyes using a swept source OCTA (PlexElite, Carl Zeiss Meditec, Dublin, USA). Enface (2D) images were exported and 6 different segmentation algorithms were applied to each image: Thresholding, region growing with manual seed point selection (RG), Frangi- or Jerman vesselness filter or a combination of vesselness filter and RG. Four reviewers rated the segmented images on a 5-point scale in the following categories: Sensitivity and specificity of aqueous vein detection, suppression of bulk extravenous signal, resolution of the smallest vein branches and amount of speckle background noise. Results were expressed as either cumulative rating (maximum 25 points) or ranks.
Results :
The combination of Jerman vesselness filter and RG reached the highest cumulative rating (22 ± 2.58 points). It was followed by Frangi filter and RG, which had a significantly lower rating (18.3 ± 1.7 points, p=0.015). Adding RG to Jerman or Frangi filter reduced their respective top ranks (1st and 2nd) to 3rd and 4th regarding the resolution of the smallest vein branches. However, it improved the suppression of bulk extravenous signal, such as leaked contrast agent, and the amount of remaining speckle noise (1st Jerman&RG, 2nd Frangi&RG). Segmentation purely based on thresholding performed worst in almost all categories.
Conclusions :
The combination of a global vesselness filter and a region growing algorithm resulted in good suppression of noise while largely preserving the complex and detailed structure of the intrascleral outflow system.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.