Evaluating different thresholding methods for vessel density calculations in optical coherence tomography angiography scans of diabetic macular edema eyes

Abstract

Purpose : Different thresholding methods to remove background noise in optical coherence tomography angiography (OCTA) scans have not been studied in diabetic macular edema (DME) eyes. We tested whether vessel density (VD) values from a built-in AngioVue Analytics software (Method 1) or a deep capillary plexus (DCP) vessel length density (VLD)-based thresholding method (Method 2) would be more accurate for DME eyes.

Methods : We examined 9 DME eyes, each with 5 repeated OCTA scans. For each eye, we performed image segmentation, registration, and averaging to generate an averaged scan for the full retina as well as the SCP, MCP, and DCP slabs. We calculated the parafoveal VD of each averaged scan using an automated method available in ImageJ and used this VD as the ground truth VD. Then, for the best quality scan of each layer, we calculated the VD using both Methods 1 and 2. For statistical analysis, we calculated the mean absolute error (MAE) of each of the two methods (using the averaged scan VD as the ground truth value) and performed the two-sided paired t-test.

Results : For the DCP layer, the MAE of Method 2 was smaller than that of Method 1 (p=0.042). For the SCP layer, the opposite was true (p=0.037). For the full retina and MCP layer, Method 1 had a smaller MAE than Method 2, but the difference between the two methods was not statistically significant (p = 0.46, 0.42 respectively for these two slabs).

Conclusions : For DME eyes, the DCP VLD-based thresholding method (Method 2) is significantly more accurate for the DCP slab, while the AngioVue VD (Method 1) more closely matches the ground truth VD for the SCP layer. The two methods perform similarly for the full retina and MCP slab. This study highlights the importance of validating different thresholding methods for different retinal slabs in order to accurately calculate parameters such as VD for downstream analysis in eyes with different pathologies, including DME.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Image caption: a) Example of an averaged scan of the DCP layer for a DME eye. b) Best quality scan out of the 5 scans used to generate the average image in (a). Here, Method 2 matches the ground truth VD more closely than Method 1. c-d) Graph comparing Method 1 (c) and Method 2 (d) with the averaged scan VD. (MAE = mean absolute error; m = slope of the best fit line)