Automated Quantification of Nonperfusion in Three Retinal Plexuses Using Projection-Resolved Optical Coherence Tomography Angiography in Diabetic Retinopathy

Abstract

Purpose: The purpose of this study was to evaluate an automated algorithm for detecting avascular area (AA) in optical coherence tomography angiograms (OCTAs) separated into three individual plexuses using a projection-resolved technique.

Methods: A 3 × 3 mm macular OCTA was obtained in 13 healthy and 13 mild nonproliferative diabetic retinopathy (NPDR) participants. A projection-resolved algorithm segmented OCTA into three vascular plexuses: superficial, intermediate, and deep. An automated algorithm detected AA in each of the three plexuses that were segmented and in the combined inner-retinal angiograms. We assessed the diagnostic accuracy of extrafoveal and total AA using segmented and combined angiograms, the agreement between automated and manual detection of AA, and the within-visit repeatability.

Results: The sum of extrafoveal AA from the segmented angiograms was larger in the NPDR group by 0.17 mm2 (P < 0.001) and detected NPDR with 94.6% sensitivity (area under the receiver operating characteristic curve [AROC] = 0.99). In the combined inner-retinal angiograms, the extrafoveal AA was larger in the NPDR group by 0.01 mm2 (P = 0.168) and detected NPDR with 26.9% sensitivity (AROC = 0.62). The total AA, inclusive of the foveal avascular zone, in the segmented and combined angiograms, detected NPDR with 23.1% and 7.7% sensitivity, respectively. The agreement between the manual and automated detection of AA had a Jaccard index of >0.8. The pooled SDs of AA were small compared with the difference in mean for control and NPDR groups.

Conclusions: An algorithm to detect AA in OCTA separated into three individual plexuses using a projection-resolved algorithm accurately distinguishes mild NPDR from control eyes. Automatically detected AA agrees with manual delineation and is highly repeatable.

Figure 1

Overview of the proposed algorithm for quantification of capillary nonperfusion.

Figure 2

Reflectance-adjusted thresholding on a normal eye with a low signal strength region (yellow circle), caused by a vitreous opacity. The low signal region seen on the original angiogram (A) persists with vessel filter (B) and binary filter with fixed threshold (C), falsely simulating capillary dropout. An en face map of the GCL and IPL reflectance amplitude map (D) is filtered and scaled to create the reflectance-adjusted threshold (E). Applying this, the binary vessel mask (F) does not show an area of false capillary drop out in the low reflectance area.

Figure 3

Nonperfusion detection in the superficial plexus of an NPDR eye. The original angiogram (A) is enhanced by a vesselness filter (B) and processed with a binary vessel mask with reflectance-adjusted thresholding (C), which is used to create a vessel distance map (D) by applying distance transform. Then morphologic operations removed regions with vessel distance less than four pixels (40 μm) (E) and then they eroded by a five-pixel-wide square kernel and threw away areas with minimum area smaller than eight pixels or a minor axis smaller than two pixels (F). Then remaining regions were dilated by seven-pixel-wide square kernel pixels (G). (H) Resulting avascular area (light blue) overlaid on the enhanced angiogram.

Figure 4

Avascular area in two scans of a normal eye with high SSI. The AA is limited to the foveal avascular zone and is similar in size and shape in all scans whether in individual plexuses or combined inner retinal angiogram. The EAA is defined as AA outside the white 1-mm circle.

Figure 5

Two normal control eyes with relatively low SSI scans in individual plexuses and combined inner retinal angiogram. No avascular area outside the foveal avascular zone is detected.

Figure 6

Two eyes with mild NPDR. Individual plexuses show incongruent areas of capillary nonperfusion that are not detected with combined inner retinal angiogram.