Abstract
Purpose :
To evaluate the impact of using different image processing algorithms to calculate commonly reported quantitative metrics in optical coherence tomography angiography (OCTA) images in patients with various stages of diabetic retinopathy.
Methods :
Single center, retrospective, observational study. Patients with diabetes from September 2017 to December 2018 were included. Complete ophthalmological exams and OCTA imaging with the Cirrus HD-OCT 5000 AngioPlex (Carl Zeiss Meditec, Inc., Oberkochen, Germany) were performed at each visit. Patients with coexisting chorioretinal disease were excluded. Scans with poor signal strength or significant motion or segmentation artifact were excluded. Demographic and clinical variables including age, gender, visual acuity, stage of diabetic retinopathy (DR), and presence of diabetic macular edema (DME) were recorded. 8 x 8 mm superficial slab images were thresholded using the Huang, Otsu, or Niblack algorithms in ImageJ (NIH, Bethesda, MD). The vessel density (VD) and skeletonized VD (SVD) were calculated for each image. Mixed-effect uni- and multivariate linear regressions were performed using the Stata statistical package (StataCorp LLC, College Station, TX). P-values < 0.05 were considered statistically significant.
Results :
144 scans from 48 patients were included. 54 were excluded for poor signal strength or significant artifact. Of the remaining 90, 26 had no DR, 47 had nonproliferative diabetic retinopathy (NPDR), and 17 had proliferative diabetic retinopathy (PDR). 24 of 90 scans had DME. The thresholding algorithm used significantly impacted VD and SVD even when controlling for age, DME, and DR stage (p-values < 0.001). The Otsu and Niblack algorithms gave significantly lower measurements of VD and SVD than the Huang algorithm (p-values < 0.001). DME was significantly associated with lower VD and SVD across all algorithms (p-values < 0.015). PDR was borderline significant for lower VD (p = 0.056) and significant for lower SVD using the Huang algorithm (p = 0.010) but not significant using Otsu and Niblack.
Conclusions :
Caution must be taken when quantitatively analyzing OCTA images, as the specific thresholding algorithm used may impact the results of any given study. There is a need for standardization of image processing algorithms to ensure robust and consistent analysis of OCTA imaging.
This is a 2021 ARVO Annual Meeting abstract.