Manual OCTA Segmentation Correction to Improve Image Quality and Visualization of Choroidal Neovascularization in AMD

Abstract

Purpose : OCTA segmentation errors can lead to false negative or false positive CNV diagnosis. We quantitated improvement in CNV visualization using different segmentation correction attempts

Methods : Retrospective, case series of 198 eyes with neovascular AMD (nAMD) patients imaged with an SD-OCT (Spectralis, Heidelberg). 73 of 198 OCTA scans required manual segmentation correction. We compared uncorrected 512 OCTA B- scan to software propagated minimal manual (2 scan corrections), moderate (10 scan corrections) to 50 (detailed correction) scans with falsely segmented Burch’s Membrane (BM).
Main outcome measures: 2 retina experts graded the improvement of image quality (scale -1 to 3). The background (vessel) signal was quantified using ImageJ and the CNV vessel pattern was analyzed using AngioTool categorized the improvement into two groups, 1. improvement of CNV visualization and 2. reduction of background noise.

Results : 40 of the 73 corrected OCTAs improved after manual correction (55%). In about 50% of these 40 OCTAs corrected for BM segmentation errors, the correction led to a significant improvement in expert image quality grading in both the CNV visualization and the reduction of background noise. The minimal scan correction (2) already led to these improvements and additional corrections (10 or 50) did not further improve expert grading. Expert grading revealed reduced background noise after minimal correction of BM, additional corrections led to a further mild reduction in background noise. 2 corrections showed the most significant improvement in CNV identification compared to baseline uncorrected image (2-correction median grade=1 versus baseline median of zero, p < 0.0001). In Angio-Tool analysis, we observed statistically significant changes in vessel density after 2, 10 and 50 corrections compared to baseline, uncorrected images (p values 0.0412, 0.0106, 0.0095 respectively).

Conclusions : Segmentation errors correction and propagation of these corrections enhance the image quality. Our data demonstrates the efficacy of our approach in neovascular AMD (nAMD). In the subset of OCTA images that underwent resegmentation, 55% exhibited enhanced quality following manual correction. This success translated into a notable improvement in the proportion of high-quality (usable) images, elevating it from 63% to an impressive 83%.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.