Improved visualization of retinal vasculature using multi-layer segmentation of optical coherence tomography angiographic images

Visualization of vasculature from OCTA is typically done by generating en face angiography images between boundaries of interest segmented from the structural OCT data. The purpose of this study is to qualitatively and quantitatively evaluate a new multilayer segmentation algorithm in comparison to a commercial implementation that segments only the retinal pigment epithelium (RPE) and inner limiting membrane (ILM), and estimates the remaining layers based on predetermined rules.

Figure 2: Regression analysis found the two methods were well correlated, with an r-squared of 0.98 for the SRL (slope of 0.99, offset of 0.01, Figure 1, blue) and 0.89 for the DRL (slope of 1.1, offset of 0.01, Figure 1, yellow). As expected, the MLS segmented SRL density was slightly lower than CS while the MLS segmented DRL was higher than with the CS.

PURPOSE RESULTS

▪ One eye from each of 34 subjects with early diabetic retinopathy and no edema▪ CIRRUS™ HD-OCT 5000 with AngioPlex® OCT Angiography (ZEISS, Dublin, CA) ▪ 3x3 OCT Angio scan. ▪ Density of vasculature in angiography en face images was quantified using the

commercial method to estimate the percent area of perfused vasculature▪ En face angiography images generated using each of two methods:

• Agreement between MLS and CS by regression• Repeatability reported as coefficient of variation (intra-individual standard deviation

divided by group mean)

METHODS

Multilayer segmentation improves visibility of microvasculature in the deeper retinal layer on OCTA en face images with no cysts. The change affects the quantitative results derived from the angiograms, but the results are correlated to the original results in eyes with early diabetic retinopathy, and the reproducibility remains good.

CONCLUSIONS

Poster #: PB072

Email: mary.durbin@zeiss.com

Disclosures: MD (E), HB(E), AF(E), JCV (C), Carl Zeiss Meditec, Inc.; TS, MS, CN, ML (None)

Mary K Durbin, PhD1; Homayoun Bagherinia, PhD1; Ali Fard , PhD1; Torcato Santos, BA2; Mário Soares, BA2; Catarina Neves, BA2; Marta Lopes, BA2; Jose G Cunha-Vaz, MD, PhD2,3

Superficial Slab

CU

RR

ENT

Segm

enta

tio

nM

ult

ilaye

r Se

gmen

tati

on

Deep Slab Central fast B-scan

1 Carl Zeiss Meditec, Inc., Dublin, CA 2 Assoc. for Innovation and Biomedical Res. on Light and Image, Coimbra, Portugal. 3 Univ of Coimbra, Coimbra, Portugal

0.2

0.25

0.3

0.35

0.4

0.45

0.2 0.25 0.3 0.35 0.4 0.45

Perf

usi

on

Den

sity

fro

m M

ult

ilaye

r Se

gmen

tati

on

(u

nit

less

)

Perfusion Density from Commercial Segmentation (unitless)

Figure 1: Superficial and deep en face angiograms, and B-scans showing segmentation, for two segmentation approaches

Com

merc

ial seg.

Multila

yer

seg.

SRL DRL

N = 22N = 22

Perfusion Vessel Density Perfusion Vessel Density

CV: 9.7%

CV: 4.1%

CV: 5.3%

CV: 4.8%

CV: 4.8% CV: 8.4%

CV: 3.9%

CV: 5.2%

CV: 4.4%

CV: 4.5% CV: 41.8%

CV: 6.8%

CV: 8.1%

CV: 6.7%

CV: 7.3% CV: 48.1%

CV: 7.4%

CV: 8.2%

CV: 7.4%

CV: 7.6%

SRL DRL

Perfusion PerfusionVessel Density Vessel Density

N = 21N = 21

N = 21N = 21

Perfusion Vessel Density Perfusion Vessel Density

CV: 4.2%

CV: 5.4%

CV: 5%

CV: 4.9%

CV: 3.9%

CV: 5.4%

CV: 4.6%

CV: 4.6%

CV: 5.5%

CV: 7.4%

CV: 5.7%

CV: 7.4%

CV: 6.1%

CV: 7.7%

CV: 6.3%

CV: 7.6%

Figure 3: Reproducibility of Perfusion Density (left)

and Vessel Density(right) in each of four quadrants on 3x3mm scans, using the two methods (Top: Commercial, Bottom: Multilayer) for both the SRL (left grouping in green) and DRL (right grouping in grey).