Abstract
Purpose :
OCT-Angiography (OCTA) is an imaging modality that produces high-quality images of the retinal vasculature; however, vessel density (VD) measurements vary depending on the method used. To address this, we obtained manual segmentation of OCTA scans and compared to automatic thresholding and an artificial-intelligence machine-learning classifier.
Methods :
IRB approved study to analyze OCTA images, 20 healthy subjects were enrolled. One eye per subject was included. 3x3mm scans of the superficial retinal capillary plexus were obtained by swept-source OCTA. Vessel segmentation maps were obtained by 3 methods: manual tracing by three graders utilizing Photoshop; automatic thresholding in FIJI (Global and Local Mean radius: 32, 64, 128, 256; and Local Phansalkar radius: 32, 64, 128, 256); and machine-learning (Weka Segmentation - FIJI). VD was calculated using the particle counter function within FIJI. Statistical analyses performed in JMP.
To simplify comparisons across the three methods; we averaged the VD data of the 2 manual graders with the highest correlation and selected the 2 automatic thresholding settings with the most agreement in their algorithms: Global Mean and Phansalkar radius 32.
Results :
Patients: 17 females; 15 Caucasian, 1 African-American, 2 Hispanic, 1 Asian, and 1 “Other.” Mean age of 51 years +/- 17. OCTA of 14 right and 6 left eyes.
Mean VD by the three manual graders was 54.0 +/- 2.7, 45.8 +/- 6.3, and 54.2 +/- 5.7, respectively (p<0.0001) although Graders 1 and 3, p=0.9211.
VD of the 5 Mean algorithms were significantly different (p= 0.0003); while the 4 Phansalkar settings were not (p= 0.1789).
When comparing auto-thresholding, manual tracing, and machine-learning in the Global Mean algorithm, the intraclass correlation coefficient was 0.068 (fourth-class); while in the Phansalkar radius 32 algorithm had an ICC of 0.0776 (fourth-class). In paired T-tests, manual grading was found to be significantly different from all alternative methods of segmentation (p<0.0001).
Conclusions :
Although agreement can be found within one modality, there are large discrepancies in the VD achieved when comparing different methodologies. Relative to manual tracing, our machine-learning classifier produced higher VD measures; conversely automatic thresholding produced lower VD readings.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.