- Email: [email protected]
Differentiating Veins From Arteries on Optical Coherence Tomography Angiography by Identifying Deep Capillary Plexus Vortices
Differentiating Veins From Arteries on Optical Coherence Tomography Angiography by Identifying Deep Capillary Plexus Vortices XIAOYU XU, NICOLAS A. YANNUZZI, PEDRO FERNA´NDEZ-AVELLANEDA, JOSE J. ECHEGARAY, KIMBERLY D. TRAN, JONATHAN F. RUSSELL, NIMESH A. PATEL, REHAN M. HUSSAIN, DAVID SARRAF, AND K. BAILEY FREUND PURPOSE:
To introduce a simple method for differentiating retinal veins from arteries on optical coherence tomography angiography (OCTA). DESIGN: Cross-sectional pilot study. METHODS: Four default en face slabs including color depth encoded, grayscale full-thickness retina, superficial plexus, and deep capillary plexus (DCP) from nine 333-mm and nine 636-mm OCTA scans were exported and aligned. Nine ophthalmologists with minimum OCTA experience from 2 eye institutions were instructed to classify labeled vessels as arteries or veins in 3 stages. Classification was performed based on graders’ own assessment at stage 1. Graders were taught that a capillary-free zone was an anatomic feature of arteries at stage 2 and were trained to identify veins originating from vortices within the DCP at stage 3. Grading accuracy was analyzed and correlated with grading time and graders’ years in practice. RESULTS: Overall grading accuracy in stages 1, 2, and 3 was (50.4% ± 17.0%), (75.4% ± 6.0%), and (94.7% ± 2.6%), respectively. Grading accuracy for 333-mm scans in stages 1, 2, and 3 was (49.9% ± 16.3%), (79.2% ± 9.6%), and (96.9% ± 3.1%), respectively. Accuracy for 636-mm scans in stages 1, 2, and 3 was (51.4% ± 20.8%), (72.3% ± 7.9%), and (93.2% ± 3.3%), respectively. Grading performance improved significantly at each stage (all P < .001). No significant correlation was found between accuracy and time spent
Supplemental Material available at AJO.com. Accepted for publication Jun 8, 2019. From the Vitreous Retina Macula Consultants of New York (X.X., P.F-A., K.B.F.); LuEsther T. Mertz Retinal Research Center, Manhattan Eye, Ear, and Throat Hospital (X.X., P.F-A., K.B.F.), New York, New York, USA; State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University (X.X.), Guangzhou, China; Department of Ophthalmology, Bascom Palmer Eye Institute, University of Miami Miller School of Medicine (N.A.Y., J.J.E., K.D.T., J.F.R., N.A.P., R.M.H.), Miami, Florida, USA; Department of Ophthalmology, Basurto University Hospital (P.F-A.), Bilbao, Spain; Stein Eye Institute, University of California (D.S.), Los Angeles, Los Angeles, California, USA; Department of Ophthalmology, New York University of Medicine (K.B.F.); Edward S. Harkness Eye Institute, Columbia University Medical Center (K.B.F.), New York, New York, USA. Inquiries to K. Bailey Freund, Vitreous Retina Macula Consultants of New York, 950 Third Avenue, New York, NY 10022, USA; e-mail: [email protected] 0002-9394/$36.00 https://doi.org/10.1016/j.ajo.2019.06.009
©
2019
grading or between accuracy and years in practice (r [ –0.164 to 0.617, all P ‡ .077). CONCLUSIONS: We describe a simple method for accurately distinguishing retinal arteries from veins on OCTA, which incorporates the use of vortices in the DCP to identify venous origin. (Am J Ophthalmol 2019;-:-–-. Ó 2019 Elsevier Inc. All rights reserved.)
B
ECAUSE RETINAL BLOOD VESSELS ARE OPTICALLY
accessible in vivo, they are easily imaged with noninvasive techniques.1 As the arterial and venous systems are differently affected in many systemic and retinal vascular diseases, classification of retinal vessels as arteries or veins is of high medical interest.2–5 In high-quality color fundus photographs, arteries can be distinguished from veins by using various characteristics such as size, shape, vessel crossing patterns, color, brightness, and optical reflexes. Understanding that arteries and veins usually alternate after they originate from the optic nerve head also aids in accurate vessel classification.6,7 In clinical practice, fundus photographs are often inadequate to classify the smaller-caliber vessels, particularly in eyes with media opacities or other pathology obscuring differentiating features. To reliably distinguish arteries from veins on grayscale fluorescein angiography, a review of transit-phase images is often needed, but these may not be available for the nonstudy eye.8 As fluorescein angiography is an invasive procedure, it is rarely used as a primary method for classifying retinal vessels. A number of optical coherence tomography (OCT)based strategies for distinguishing arteries and veins have been developed. Initially, laboratory-based Doppler Fourier-domain OCT was used to measure and compare the flow velocity in arterial and venous systems. This technique was found to be time-consuming and of limited utility in routine clinical care.9 Later, by measuring the vessel diameter and wall thickness, and assessing the presence or absence of the hyperreflective lower border reflectivity feature using commercially available spectral domain OCT (SD-OCT), retinal vessel classification became less cumbersome yet still needed additional scans targeting the vessels of interests.10 Iwase
ELSEVIER INC. ALL
RIGHTS RESERVED.
1
and associates combined laser speckle flowgraphy, a technique used to measure relative retinal and choroidal blood flow velocities, with an adaptive optics camera imaging of fine retinal vasculature structure, to determine retinal vessel types.11 Using custom-built high-speed adaptive optics near-confocal imaging, Gu and associates showed that imaging erythrocyte motion in living human eyes could facilitate accurate classification of vessels at the level of retinal microcirculation.12 However, none of these methods allow for rapid visual evaluation of retinal vessel classification without additional scans, and many of them cannot currently be used in clinical practice because of delivery barriers. Optical coherence tomography angiography (OCTA) enables noninvasive, depth-resolved, en face visualization of retinal vascular flow by using motion of flowing blood as the intrinsic contrast. To our knowledge, the only attempt for artery/vein (A/V) identification based on OCTA was reported by Alam and associates in diabetic retinopathy.13 Two quantitative OCTA features, artery-vein ratio of blood vessel caliber and artery-vein ratio of blood vessel tortuosity were introduced in their color fundus image–guided OCTA A/V differentiation algorithm. The technique was intended to help improve the sensitivity of OCTA detection and staging of diabetic retinopathy.13 A key advantage of OCTA in comparison to dye-based angiography is the ability to visualize discrete capillary plexuses.14 Three major capillary plexuses, the superficial vascular plexus (SVP), the intermediate capillary plexus, and the deep capillary plexus (DCP) can be resolved by current commercially available OCTA devices.15,16 With improvements in segmentation algorithms, OCTArendered capillary layers have been anatomically correlated to histologic sections.17,18 By studying the retinal collaterals formed after the occurrence of retinal vein occlusion, Freund and associates19 noted that all collateral vessels coursed through the DCP, whereas no collaterals were localized exclusively to the SVP, suggesting that the venous outflow predominantly originates in the intermediate capillary plexus and DCP, which is consistent with previous investigations both in animal models (mice,20 rats,21 and pigs22) and in humans.23,24 Balaratnasingam and associates confirmed that a physiologic avascular area, termed a capillary-free zone, was evident adjacent to retinal arteries in both histology and in OCTA images; however, histology correctly localized this finding to the SVP whereas OCTA often erroneously included portions of these more superficial vessels in the deeper layers.25 The featured vortices arrangement of capillaries in the DCP, therefore, could serve as a potential anatomic biomarker of venous origin. The current study was aimed to develop a simple OCTA-based method for visual classification of retinal vessels by recognizing that all vortices within the DCP are connected to veins in the more superficial retinal layers.
2
METHODS THIS CROSS-SECTIONAL PILOT STUDY FOLLOWED THE
tenets of the Declaration of Helsinki, complied with the Health Insurance Portability and Accountability Act of 1996, and was approved by Western Institutional Review Board (Olympia, Washington, USA). Written informed consent was obtained from all subjects. Participants enrolled in this study were recruited from volunteers who agreed to undergo the examinations and met the eligibility criteria described below at the Vitreous Retina Macula Consultants of New York (New York, USA) between February 2018 and October 2018. IMAGE ACQUISITION:
Retinal imaging for use in testing and training was collected retrospectively from scans acquired on normal eyes of healthy participants. Inclusion _ 20/20; (2) criteria were (1) best-corrected visual acuity > _ 21 mmHg, and (3) a spherical equivintraocular pressure < alent refractive error between 3 diopters (D) and þ1 D. Exclusion criteria were (1) history of any form of vitreoretinal diseases, high myopia, uveitis, glaucoma, or optic neuropathy; (2) media opacities that might prevent successful imaging; (3) prior intraocular surgery, laser treatment, or ocular trauma; (4) systemic or neurologic diseases that could affect retinal health, including diabetes, hypertension, dementia, or multiple sclerosis; and (5) current or prior use of systemic medications known to affect the retinal circulation. Subjects underwent a complete ophthalmic evaluation, including manifest refraction, uncorrected and bestcorrected visual acuity, intraocular pressure measurement, slit lamp biomicroscopy examination, ophthalmoscopic examination, high-resolution true color confocal color fundus photography (EIDON, CenterVue, Padua, Italy), and swept-source OCTA (SS-OCTA; PLEX Elite 9000, Carl Zeiss Meditec, Inc, Dublin, California, USA). Scan patterns acquired on the SS-OCTA device were the default 333-mm and 636-mm OCTA cube scans. Only scans with _8 were used in this study. a signal strength of >
TRAINING AND GRADING:
Nine 333-mm and nine 636-mm OCTA scan volumes centered on the fovea from 18 eyes of 14 healthy subjects (8 men and 6 women; 39.8 6 17.1 years of age, range 15-71 years) were used to create training and testing image sets for use in this study. A total of 147 vessels from the nine 333-mm image sets and 193 vessels from the nine 636-mm image sets were labeled with numbers for later assignment as artery or vein by study participants. Image sets for training and testing were created by exporting the automatically segmented default en face slabs (color retina depthencoded, grayscale full-thickness retina, SVP, and DCP with projection removal) from each OCTA volume and pasting them into a PowerPoint (Microsoft Corporation,
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
Redmond, Washington, USA) file, one slab per slide. Blood vessels were labeled with numbers in each slide. High-resolution color fundus photographs were used for identification of arteries and veins based on the following characteristics: arteries are brighter in color than veins; arteries are thinner than neighboring veins; the central light reflex is wider in arteries and smaller in veins; and arteries and veins usually alternate around the optic disc before branching out (Figure 1). Three of the authors (X.X., P.F.A., and K.B.F.) each independently reviewed and characterized each vessel used in the training and grading image sets. Subsequently, these authors reviewed their results together. Discordant gradings were rare. In those instances, following review, a consensus determination was achieved for each vessel used for training and testing. Nine ophthalmologists with minimal OCTA experience and no formal training in OCTA from 2 institutions (5 from Bascom Palmer Eye Institute, University of Miami Miller School of Medicine, Miami, Florida, USA; and 4 from Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangzhou, China), were instructed to classify all labeled vessels as arteries or veins in 3 stages. Three image sets, each composed of 1 randomly selected 333-mm and 636-mm study, were presented to each participant during each stage. Each scan was used only once during grading. In stage 1, untrained classification based on graders’ own assessment was performed. Prior to stage 2, graders were taught that a capillary-free zone was an anatomic feature of arteries. Prior to stage 3, graders were further trained to identify veins by recognizing their origin in vortices within the DCP (characterized as green convergence of capillaries on color depth-encoded slabs) and to consider that arteries and veins typically alternate as each vein drains capillary beds perfused by adjacent arteries in stage 3 (Figure 2). The training time prior to stage 2 and stage 3 was no more than 15 minutes. During training, typical examples of capillary-free zone/vortices within the DCP were shown to the graders using the image sets they had previously graded. Then each participant was asked to demonstrate an understanding of the technique by correctly identifying the vascular features on a new image set. Data including grading performances and the average grading time spent on each vessel in each stage regardless of the scanning patterns of the images used for grading were deidentified and recorded in a Microsoft Excel 2016 database (Microsoft Corporation). FIGURE 1. Examples of 333-mm (A) and 636-mm (B) optical coherence tomography (OCTA) image sets acquired from normal eyes and used for grading and training exercises. Vessels needed to be classified into arteries and veins are numbered on each slab: color retina depth-encoded image (A1 and B1), grayscale full-thickness retina (A2 and B2), superficial plexus (A3 and B3), and deep plexus with projection removal (A4 and B4). When preparing these image sets, the color
VOL. -
STATISTICAL
ANALYSIS: Descriptive results were presented as mean 6 standard deviation with range
depth-encoded OCTA images were aligned to high-resolution color fundus photographs using the retinal vessels as landmarks, which were then used to confirm the identity of each numbered vessel as either artery or vein (A5 and B5).
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
3
FIGURE 2. On a set of 333-mm optical coherence tomography (OCTA) images centered on the fovea, arteries (A), and veins (V) were classified and labeled on color retina depth-encoded (A), grayscale full-thickness retina (B), and superficial plexus (C) slabs. On the deep capillary plexus (DCP) slab with projection removal (D), vortices are marked with cyan dots showing the centers of capillaries convergence. The vessels are classified based on the following criteria: superficial arteries have adjacent capillary-free zones (E, magnification of the white solid line box in A, the arrowheads show the adjacent gaps free of capillaries), and the vortices in the DCP drain into veins (F, magnification of the white dotted box in A, magenta dots are put on the center of the convergence of capillaries). The classifications are confirmed using the high-resolution color fundus photograph (F) and its enlargement (G) centered on the fovea with the size of 333 mm.
4
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
TABLE 1. Average Grading Accuracy in Each Stage Subdivided Into Scanning Patterns (Mean 6 Standard Deviation; %) and Comparisons
TABLE 2. Sensitivity and Specificity in Classifying Retinal Veins in Each Stage Subdivided Into Scanning Patterns (%) Stage 1
Scanning Pattern
333 mm 636 mm Total P valueb
Stage 1
Stage 2
Stage 3
P Valuea
Scanning Pattern
49.9 6 16.3 51.4 620.8 50.4 6 17.0 .850
79.2 6 9.6 72.3 6 7.9 75.4 6 6.0 .020
96.9 6 3.1 93.2 6 3.3 94.7 6 2.6 .029
<.001 <.001 <.001 —
333 mm 44.96 636 mm 44.77 Total 44.85
Stage 2
Stage 3
Sensitivity Specificity Sensitivity Specificity Sensitivity Specificity
56.73 60.00 58.37
72.77 74.93 74.14
87.68 67.88 78.22
94.37 91.71 92.62
94.12 95.52 94.82
a
Comparisons of the grading accuracy between the 3 stages in each scanning pattern using generalized estimating equations. b Comparisons of the grading accuracy between 333-mm scan and 636-mm scan in each stage using generalized estimating equations.
(minimum, maximum). The Kolmogorov-Smirnov test and the Levene test were conducted to test the normality and homogeneity of variance, respectively. The correlation between accuracy and grading time, and between accuracy and graders’ year of practicing ophthalmology were analyzed using Pearson correlation coefficient. The accuracy between different grading stages and the accuracy between the 2 scanning patterns in each grading stage were compared using generalized estimating equations. Statistical analysis was conducted using SPSS version 22.0 (IBM Corp, Armonk, New York, USA) and a P value of <.05 was considered statistically significant.
RESULTS
TABLE 3. Correlations of Grading Accuracy With Grading Time in Each Stage
r 95% confidence interval of r P value
Stage 1
Stage 2
Stage 3
–0.013 –0.624, 0.679
0.518 0.057, 0.970
–0.164 –0.798, 0.741
.974
.189
.674
from arteries facilitate analyses of how the disease affects different vessel subtypes (Figure 3). Identifying the nature of a prior retinal vascular occlusion. A 61-year-old woman presented with the asymptomatic finding of localized thinning of the inner nuclear layer (INL) on OCT B-scans of the right eye. The OCT pattern was consistent with resolved paracentral acute middle maculopathy. Using our methods, it is possible to determine that a prior episode of cilioretinal artery hypoperfusion was responsible for producing OCT findings consistent with a previous ischemic event of the deep vascular complex (ie, paracentral acute middle maculopathy) (Figure 4).
OVERALL GRADING ACCURACY AND THE SEPARATE ACCU-
racies for the 333-mm and 636-mm image sets at each stage are summarized in Table 1. The sensitivity and specificity for identifying retinal veins at each stage are shown in Table 2. The average time spent on each vessel was (16.5 6 5.4 seconds), (8.5 6 3.6 seconds), and (15.6 6 8.1 seconds) in stages 1, 2, and 3, respectively. The average years in practice of 9 ophthalmologists was 6.4 6 4.2 years. Correlation between grading accuracy and mean time spent on each vessel during each stage are displayed in Table 3. Correlations between grading accuracy at each stage and graders’ mean years in practice were subdivided by scan pattern size and are displayed in Table 4. CLINICAL EXAMPLES:
Evaluating diabetic retinopathy. A 71-year-old woman was seen for follow-up of longstanding nonproliferative diabetic retinopathy associated with diabetic macular edema. She had received focal laser in the past, and more recently, intravitreal anti– vascular endothelial growth factor therapy as needed. The methods described herein for distinguishing veins
VOL. -
Exploring the vascular effects of macular telangiectasia type 2. A 64-year-old woman was referred for evaluation of bilateral metamorphopsia and reduced visual acuity to 20/150 in the right eye and 20/50 in the left eye. The diagnosis of macular telangiectasia type 2 was made based on characteristic clinical and imaging findings. Using our methods, it is easy to determine that the vascular damage is mostly concentrated around a vein draining the temporal aspect of the perifoveal capillary ring (Figure 5).
DISCUSSION IN THIS STUDY, WE DEMONSTRATE A SIMPLE METHOD TO
distinguish retinal arteries from veins in standard en face images acquired on a commercially available OCTA device. This technique uses the novel approach of identifying the capillary vortices in the DCP as an anatomic biomarker of venous origin. Ophthalmologists lacking OCTA
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
5
TABLE 4. Correlations of Grading Accuracy With Graders’ Years in Practicing Ophthalmology in Each Stage Subdivided Into Scanning Patterns Stage 1
r 95% confidence interval of r P value
Stage 2
Stage 3
333 mm
636 mm
333 mm
636 mm
333 mm
636 mm
0.587 –0.498, 0.926
0.608 0.322, 0.902
0.275 –0.354, 0.663
0.005 –0.670, 0.706
0.579 –0.008, 0.920
0.617 0.032, 0.930
.097
.082
.474
.990
.103
.077
FIGURE 3. Multimodal imaging of a 71-year-old woman with nonproliferative diabetic retinopathy in her right eye. (A) ‘‘True color’’ scanning laser ophthalmoscopy image shows microvascular abnormalities of diabetic retinopathy and evidence of prior focal laser for diabetic macular edema. (B) Fluorescein angiography (FA) at 1 minute 30 seconds shows scattered hyperfluorescent microaneurysms, staining laser scars, and minimal vascular leakage. A 9x9-mm color retina depth-encoded optical coherence tomography angiography study (C) is aligned to the FA image with veins labeled as ‘‘V’’ and arteries labeled as ‘‘A.’’ Some of the deep venous vortices used to identify veins are marked with magenta dots.
experience required only a brief training exercise to significantly improve their accuracy in grading retinal vessel type from a mean of 50.4% to a mean of 94.7%. Distinguishing arteries from veins on OCTA is of high clinical relevance because findings specific to each vessel type can add important information for detection and differentiation of many systemic and retinal diseases (Figures 3-5). The ability of OCTA to provide depthresolved images of retinal and choroidal vascular blood flow without the need for intravenous dye has made it a widely used imaging tool for studying retinal and optic nerve disorders including neovascular age-related macular degeneration, retinal vascular diseases, macular telangiectasia, pathologic myopia, inflammatory chorioretinal diseases, and glaucoma.26–28 Fast scanning speeds, improved retinal layer segmentation, and projection artifact removal algorithms29 have enabled current OCTA devices to resolve the discrete capillary plexuses within the retina. Using our training module, we found that grading accuracy increased from 50.4% in stage 1, to 75.4% in stage 2, and to 94.7% in stage 3. Postassignment interviews with graders expressed great difficulty grading vessels in stage 1 6
as the OCTA images lack visual clues, such as color, contrast, width, and shape, they would typically use to distinguish arteries from veins when evaluating other forms of retinal imaging. In addition, the graders could not trace to optic nerve because neither scan pattern includes the disc when it is centered on the fovea. In stage 2, graders used information from histologic studies of human retinas demonstrating that a capillaryfree zone is present around retinal arteries that develops during embryogenesis. This area exists where transmural oxygen diffuses to satisfy the metabolic demands of cells immediately adjacent to oxygen-rich superficial retinal arteries.30,31 Grading performance in stage 2 indicated that awareness of the periarterial capillary-free zone could improve accuracy to 75.4%. Graders had difficulty applying this technique to smaller arteries as the presence or absence of a capillary-free zone was more difficult to discern, and when they tried to follow the course of the larger arteries distal to arteriovenous crossings. In stage 3, graders were taught to identify veins by their origin in the vortices within the DCP which they were then able to trace back to arteriovenous crossings. When graders
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
FIGURE 4. Multimodal imaging of a 61-year-old woman with localized thinning of the inner nuclear layer (INL) in her right eye due to a prior episode of cilioretinal artery hypoperfusion. Near-infrared reflectance (A) shows no detectable lesion. Optical coherence tomography B-scans (B, C) corresponding to the green lines in A show INL thinning with preserved outer retinal bands indicating resolved paracentral acute middle maculopathy (PAMM). The green vortices in the 333-mm color retina depth-encoded optical coherence tomography angiography (OCTA) image (D) are used to distinguish veins (labeled ‘‘V’’) from arteries (labeled ‘‘A’’). Some of the vortices used to identify veins are marked with magenta dots. Flow is better preserved in the superficial vascular plexus (E) than in the deep capillary plexus (F). The OCTA study is useful to show that the reduced flow is in a retinal area supplied by a small cilioretinal artery indicated with blue arrows on color fundus photography (G) and fluorescein angiography (H).
VOL. -
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
7
FIGURE 5. Multimodal imaging of the left eye of a 56-year-old woman with macular telangiectasia type 2. ‘‘True-color’’ scanning laser ophthalmoscopy imaging (A) with enlargement of the foveal center (inset) shows loss of retinal transparency in the temporal fovea and a few small superficial retinal crystals. Fluorescein angiography (FA) at 1 minute 18 seconds (B) shows diffuse hyperfluorescence from the telangiectatic capillaries in the temporal fovea. The 333-mm color retina depth-encoded optical coherence tomography angiography image (C) aligned to the FA image (blue rectangle in B has veins labeled as ‘‘V’’ and arteries labeled as ‘‘A,’’ with some of the vortices used to identify veins marked with magenta dots. The vascular damage is mostly concentrated around a vein draining the temporal aspect of the perifoveal capillary ring. The green line in the near-infrared reflectance image (D) shows the location of a horizontal optical coherence tomography B-scan (E). The OCT B-scan shows central loss of both inner and outer retinal layers, hyporeflective intraretinal cavities, and preservation of a foveal depression.
had separate complementary strategies for identify arteries and veins, grading accuracy increased to 94.7%. In stages 2 and 3, grading accuracy for 333-mm scans was significantly higher than for 636-mm scans. This was likely related to different A-scan densities used in these 2 grid patterns. The 333-mm study uses 10-mm spacing between A-scans, whereas 12-mm spacing is used for the 636-mm scan pattern. Also, the 333-mm scan pattern uses 43 averaging for each B-scan whereas the 636-mm uses only 23 averaging/B-scan.32 Therefore, both the periarterial capillary-free zone and the DCP vortices were more easily recognized in the smaller 333-mm scan pattern. Also, because all studies were centered on the fovea, vessels at the superior, inferior, and temporal margins of 636-mm scans were often smaller than those on 333-mm scans, making it more difficult to recognize these important scan features. One technique for enhancing vascular detail in 636-mm and larger scan patterns is to average multiple tracked scan acquisitions (Supplemental Figure 1).33 Graders uniformly reported that among the 4 slabs provided in each image set, the color depth-encoded slab was the most useful in all aspects of grading. Color coding of the superficial and deep vessels made identifying the connection of veins to the deep capillary vortices simply 8
a matter identifying the convergence of thin green linear structures. After completing the training exercises, most graders felt that this single visualization would be sufficient for grading, whereas the 3 additional slabs were typically ignored in stage 3, as they seemed to provide no additional distinguishing information. Interestingly, grading accuracy was not statistically correlated with grading time or graders’ years in practice. For all participants, training grading sessions were completed in less than 15 minutes. These results suggest that the vessel classification method described herein can be used to train nonophthalmologists, including graders in imaging reading centers. We are currently exploring whether our methods can be applied to an artificial intelligence or a machine learning based automatic vessel classification algorithm. Potential limitations of our methods include the lack of a separate control group asked to grade the 3 successive image sets without training, but with some feedback regarding grading accuracy at each stage. It is possible that some of the improvement in grading accuracy could result from corrective feedback alone, not the specific techniques taught during the training exercises. To minimize this concern, graders were not provided complete answers for
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
the prior image sets, and training sessions for the next stage included just a few vessels that had been graded during the prior session. Another important limitation is that this study used scans obtained on only one of several commercially available OCTA instruments. Currently, the review software provided by other OCTA device makers does not include an en face color depth-encoded visualization, which was instrumental for the technique described in this study. However, we have successfully applied our method to standard acquisitions obtained on other OCTA devices by color coding either the superficial or deep slab and combining both into a composite image using ImageJ (Rasband, W.S., ImageJ, US National Institutes of Health, Bethesda, Maryland, USA, https://imagej.nih.gov/ij/) or Adobe Photoshop (Adobe Inc, San Jose, California, USA) (Supplemental Figure 2). Although this extra step would likely be too cumbersome to use in general practice, other OCTA device manufacturers could potentially provide a color retina depth-encoded image in future versions of their review software.
Finally, this study did not formally evaluate grading accuracy in eyes with retinal pathology or in scans with poor quality due to motion artifacts, low signal strength, or segmentation errors. The application of our technique in lower-quality images or diseased eyes is more challenging and limits its implementation in certain situations. As our training and testing subjects were limited to junior ophthalmologists from only 2 eye institutes having little OCTA experience, larger cohorts, with more diverse career backgrounds, would be necessary for a more hierarchical performance assessment. Finally, grading time in each stage was not subdivided by scanning patterns. Because the differences of grading accuracy of 333-mm scans and 636-mm scans were of statistical significance in stage 2 and stage 3, it would have been preferable to compare the time used for grading with each scanning pattern. In conclusion, we describe a novel method for accurately distinguishing retinal arteries from veins on OCTA that incorporates the use of vortices in the DCP to identify venous origin.
ALL AUTHORS HAVE COMPLETED AND SUBMITTED THE ICMJE FORM FOR DISCLOSURE OF POTENTIAL CONFLICTS OF INTEREST and the following were reported: D.S. is a consultant to Genentech, Heidelberg Engineering, Amgen, Bayer, Novartis, Optovue, Regeneron. K.B.F. is a consultant for Optovue, Heidelberg Engineering, Zeiss, Allergan, and Novartis and he receives research support from Genentech/Roche. This work was supported by the Macula Foundation, New York, NY; National Natural Science Foundation of China (81800879); Fundamental Research Funds of the State Key Laboratory of Ophthalmology, China (2018KF04 & 2017QN05); and Natural Science Foundation of Guangdong Province (2017A030310372). The authors thank Drs. Shasha Yang, Yao Yang, Linlin Hao, Shufen Lin and Hui Xiao from Zhongshan Ophthalmic Center, Sun Yat-sen University, China, for their help and contribution to this study. All authors attest that they meet the current ICMJE requirements to qualify as authors.
REFERENCES 1. Suzuki Y. Direct measurement of retinal vessel diameter: comparison with microdensitometric methods based on fundus photographs. Surv Ophthalmol 1995;39(Suppl 1): S57–S65. 2. Qiu C, Cotch MF, Sigurdsson S, et al. Microvascular lesions in the brain and retina: the age, gene/environment susceptibility-Reykjavik study. Ann Neurol 2009;65(5): 569–576. 3. Wong TY, Klein R, Sharrett AR, et al. The prevalence and risk factors of retinal microvascular abnormalities in older persons: the Cardiovascular Health Study. Ophthalmology 2003;110(4):658–666. 4. Ikram MK, De Jong FJ, Van Dijk EJ, et al. Retinal vessel diameters and cerebral small vessel disease: the Rotterdam Scan Study. Brain 2006;129(Pt 1):182–188. 5. Shin YU, Lee DE, Kang MH, et al. Optical coherence tomography angiography analysis of changes in the retina and the choroid after haemodialysis. Sci Rep 2018;8(1):17184. 6. Kondermann C, Kondermann D, Yan M. Blood vessel classification into arteries and veins in retinal images. Medical Imaging: SPIE 2007;9. 7. Welikala RA, Foster PJ, Whincup PH, et al. Automated arteriole and venule classification using deep learning for retinal images from the UK Biobank cohort. Comput Biol Med 2017; 90:23–32.
VOL. -
8. Shao Q, Heussen FM, Ouyang Y, Hager A. Retinal vessel diameter changes in different severities of diabetic retinopathy by SD-OCT. Eur J Ophthalmol 2016;26(4):342–346. 9. Wang Y, Fawzi AA, Varma R, et al. Pilot study of optical coherence tomography measurement of retinal blood flow in retinal and optic nerve diseases. Invest Ophthalmol Vis Sci 2011;52(2):840–845. 10. Ouyang Y, Shao Q, Scharf D, Joussen AM, Heussen FM. An easy method to differentiate retinal arteries from veins by spectral domain optical coherence tomography: retrospective, observational case series. BMC Ophthalmol 2014;14:66. 11. Iwase T, Ra E, Yamamoto K, Kaneko H, Ito Y, Terasaki H. Differences of retinal blood flow between arteries and veins determined by laser speckle flowgraphy in healthy subjects. Medicine (Baltimore) 2015;94(33):e1256. 12. Gu B, Wang X, Twa MD, Tam J, Girkin CA, Zhang Y. Noninvasive in vivo characterization of erythrocyte motion in human retinal capillaries using high-speed adaptive optics nearconfocal imaging. Biomed Opt Express 2018;9(8):3653–3677. 13. Alam M, Toslak D, Lim JI, Yao X. Color fundus image guided artery-vein differentiation in optical coherence tomography angiography. Invest Ophthalmol Vis Sci 2018;59(12): 4953–4962. 14. Spaide RF, Klancnik JM Jr, Cooney MJ. Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. JAMA Ophthalmol 2015; 133(1):45–50.
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
9
15. Zhang M, Hwang TS, Campbell JP, et al. Projection-resolved optical coherence tomographic angiography. Biomed Opt Express 2016;7(3):816–828. 16. Campbell JP, Zhang M, Hwang TS, et al. Detailed vascular anatomy of the human retina by projection-resolved optical coherence tomography angiography. Sci Rep 2017;7:42201. 17. Spaide RF, Curcio CA. Evaluation of segmentation of the superficial and deep vascular layers of the retina by optical coherence tomography angiography instruments in normal eyes. JAMA Ophthalmol 2017;135(3):259–262. 18. Gattoussi S, Freund KB. Correlating structural and angiographic optical coherence tomography in the intermediate and deep retinal capillary plexuses. Exp Eye Res 2017;165: 96–98. 19. Freund KB, Sarraf D, Leong BCS, Garrity ST, Vupparaboina KK, Dansingani KK. Association of optical coherence tomography angiography of collaterals in retinal vein occlusion with major venous outflow through the deep vascular complex. JAMA Ophthalmol 2018;136(11): 1262–1270. 20. Paques M, Tadayoni R, Sercombe R, et al. Structural and hemodynamic analysis of the mouse retinal microcirculation. Invest Ophthalmol Vis Sci 2003;44(11):4960–4967. 21. Genevois O, Paques M, Simonutti M, et al. Microvascular remodeling after occlusion-recanalization of a branch retinal vein in rats. Invest Ophthalmol Vis Sci 2004;45(2):594–600. 22. Foulds WS, Kek WK, Luu CD, Song IC, Kaur C. A porcine model of selective retinal capillary closure induced by embolization with fluorescent microspheres. Invest Ophthalmol Vis Sci 2010;51(12):6700–6709. 23. Foreman DM, Bagley S, Moore J, Ireland GW, McLeod D, Boulton ME. Three dimensional analysis of the retinal vasculature using immunofluorescent staining and confocal laser scanning microscopy. Br J Ophthalmol 1996;80(3):246–251.
10
24. Garrity ST, Paques M, Gaudric A, Freund KB, Sarraf D. Considerations in the understanding of venous outflow in the retinal capillary plexus. Retina 2017;37(10):1809–1812. 25. Balaratnasingam C, An D, Sakurada Y, et al. Comparisons between histology and optical coherence tomography angiography of the periarterial capillary-free zone. Am J Ophthalmol 2018;189:55–64. 26. Kashani AH, Chen CL, Gahm JK, et al. Optical coherence tomography angiography: a comprehensive review of current methods and clinical applications. Prog Retin Eye Res 2017; 60:66–100. 27. Hagag AM, Gao SS, Jia Y, Huang D. Optical coherence tomography angiography: technical principles and clinical applications in ophthalmology. Taiwan J Ophthalmol 2017; 7(3):115–129. 28. Tan ACS, Tan GS, Denniston AK, et al. An overview of the clinical applications of optical coherence tomography angiography. Eye (Lond) 2018;32(2):262–286. 29. Spaide RF, Fujimoto JG, Waheed NK. Image artifacts in optical coherence tomography angiography. Retina 2015; 35(11):2163–2180. 30. Michaelson IC, Campbell ACP. The anatomy of the finer retinal vessels, and some observations on their significance in certain retinal diseases. Trans Ophthalmol Soc UK 1940; 60:40. 31. McLeod D. Krogh cylinders in retinal development, panretinal hypoperfusion and diabetic retinopathy. Acta Ophthalmol 2010;88(8):817–835. 32. Zheng F, Zhang Q, Shi Y, et al. Age-dependent changes in the macular choriocapillaris of normal eyes imaged with swept-source OCT angiography. Am J Ophthalmol 2019;200: 110–122. 33. Pappuru RR, Briceno C, Ouyang Y, Walsh AC, Sadda SR. Clinical significance of B-scan averaging with SD-OCT. Ophthalmic Surg Lasers Imaging 2012;43(1):63–68.
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
To introduce a simple method for differentiating retinal veins from arteries on optical coherence tomography angiography (OCTA). DESIGN: Cross-sectional pilot study. METHODS: Four default en face slabs including color depth encoded, grayscale full-thickness retina, superficial plexus, and deep capillary plexus (DCP) from nine 333-mm and nine 636-mm OCTA scans were exported and aligned. Nine ophthalmologists with minimum OCTA experience from 2 eye institutions were instructed to classify labeled vessels as arteries or veins in 3 stages. Classification was performed based on graders’ own assessment at stage 1. Graders were taught that a capillary-free zone was an anatomic feature of arteries at stage 2 and were trained to identify veins originating from vortices within the DCP at stage 3. Grading accuracy was analyzed and correlated with grading time and graders’ years in practice. RESULTS: Overall grading accuracy in stages 1, 2, and 3 was (50.4% ± 17.0%), (75.4% ± 6.0%), and (94.7% ± 2.6%), respectively. Grading accuracy for 333-mm scans in stages 1, 2, and 3 was (49.9% ± 16.3%), (79.2% ± 9.6%), and (96.9% ± 3.1%), respectively. Accuracy for 636-mm scans in stages 1, 2, and 3 was (51.4% ± 20.8%), (72.3% ± 7.9%), and (93.2% ± 3.3%), respectively. Grading performance improved significantly at each stage (all P < .001). No significant correlation was found between accuracy and time spent
Supplemental Material available at AJO.com. Accepted for publication Jun 8, 2019. From the Vitreous Retina Macula Consultants of New York (X.X., P.F-A., K.B.F.); LuEsther T. Mertz Retinal Research Center, Manhattan Eye, Ear, and Throat Hospital (X.X., P.F-A., K.B.F.), New York, New York, USA; State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University (X.X.), Guangzhou, China; Department of Ophthalmology, Bascom Palmer Eye Institute, University of Miami Miller School of Medicine (N.A.Y., J.J.E., K.D.T., J.F.R., N.A.P., R.M.H.), Miami, Florida, USA; Department of Ophthalmology, Basurto University Hospital (P.F-A.), Bilbao, Spain; Stein Eye Institute, University of California (D.S.), Los Angeles, Los Angeles, California, USA; Department of Ophthalmology, New York University of Medicine (K.B.F.); Edward S. Harkness Eye Institute, Columbia University Medical Center (K.B.F.), New York, New York, USA. Inquiries to K. Bailey Freund, Vitreous Retina Macula Consultants of New York, 950 Third Avenue, New York, NY 10022, USA; e-mail: [email protected] 0002-9394/$36.00 https://doi.org/10.1016/j.ajo.2019.06.009
©
2019
grading or between accuracy and years in practice (r [ –0.164 to 0.617, all P ‡ .077). CONCLUSIONS: We describe a simple method for accurately distinguishing retinal arteries from veins on OCTA, which incorporates the use of vortices in the DCP to identify venous origin. (Am J Ophthalmol 2019;-:-–-. Ó 2019 Elsevier Inc. All rights reserved.)
B
ECAUSE RETINAL BLOOD VESSELS ARE OPTICALLY
accessible in vivo, they are easily imaged with noninvasive techniques.1 As the arterial and venous systems are differently affected in many systemic and retinal vascular diseases, classification of retinal vessels as arteries or veins is of high medical interest.2–5 In high-quality color fundus photographs, arteries can be distinguished from veins by using various characteristics such as size, shape, vessel crossing patterns, color, brightness, and optical reflexes. Understanding that arteries and veins usually alternate after they originate from the optic nerve head also aids in accurate vessel classification.6,7 In clinical practice, fundus photographs are often inadequate to classify the smaller-caliber vessels, particularly in eyes with media opacities or other pathology obscuring differentiating features. To reliably distinguish arteries from veins on grayscale fluorescein angiography, a review of transit-phase images is often needed, but these may not be available for the nonstudy eye.8 As fluorescein angiography is an invasive procedure, it is rarely used as a primary method for classifying retinal vessels. A number of optical coherence tomography (OCT)based strategies for distinguishing arteries and veins have been developed. Initially, laboratory-based Doppler Fourier-domain OCT was used to measure and compare the flow velocity in arterial and venous systems. This technique was found to be time-consuming and of limited utility in routine clinical care.9 Later, by measuring the vessel diameter and wall thickness, and assessing the presence or absence of the hyperreflective lower border reflectivity feature using commercially available spectral domain OCT (SD-OCT), retinal vessel classification became less cumbersome yet still needed additional scans targeting the vessels of interests.10 Iwase
ELSEVIER INC. ALL
RIGHTS RESERVED.
1
and associates combined laser speckle flowgraphy, a technique used to measure relative retinal and choroidal blood flow velocities, with an adaptive optics camera imaging of fine retinal vasculature structure, to determine retinal vessel types.11 Using custom-built high-speed adaptive optics near-confocal imaging, Gu and associates showed that imaging erythrocyte motion in living human eyes could facilitate accurate classification of vessels at the level of retinal microcirculation.12 However, none of these methods allow for rapid visual evaluation of retinal vessel classification without additional scans, and many of them cannot currently be used in clinical practice because of delivery barriers. Optical coherence tomography angiography (OCTA) enables noninvasive, depth-resolved, en face visualization of retinal vascular flow by using motion of flowing blood as the intrinsic contrast. To our knowledge, the only attempt for artery/vein (A/V) identification based on OCTA was reported by Alam and associates in diabetic retinopathy.13 Two quantitative OCTA features, artery-vein ratio of blood vessel caliber and artery-vein ratio of blood vessel tortuosity were introduced in their color fundus image–guided OCTA A/V differentiation algorithm. The technique was intended to help improve the sensitivity of OCTA detection and staging of diabetic retinopathy.13 A key advantage of OCTA in comparison to dye-based angiography is the ability to visualize discrete capillary plexuses.14 Three major capillary plexuses, the superficial vascular plexus (SVP), the intermediate capillary plexus, and the deep capillary plexus (DCP) can be resolved by current commercially available OCTA devices.15,16 With improvements in segmentation algorithms, OCTArendered capillary layers have been anatomically correlated to histologic sections.17,18 By studying the retinal collaterals formed after the occurrence of retinal vein occlusion, Freund and associates19 noted that all collateral vessels coursed through the DCP, whereas no collaterals were localized exclusively to the SVP, suggesting that the venous outflow predominantly originates in the intermediate capillary plexus and DCP, which is consistent with previous investigations both in animal models (mice,20 rats,21 and pigs22) and in humans.23,24 Balaratnasingam and associates confirmed that a physiologic avascular area, termed a capillary-free zone, was evident adjacent to retinal arteries in both histology and in OCTA images; however, histology correctly localized this finding to the SVP whereas OCTA often erroneously included portions of these more superficial vessels in the deeper layers.25 The featured vortices arrangement of capillaries in the DCP, therefore, could serve as a potential anatomic biomarker of venous origin. The current study was aimed to develop a simple OCTA-based method for visual classification of retinal vessels by recognizing that all vortices within the DCP are connected to veins in the more superficial retinal layers.
2
METHODS THIS CROSS-SECTIONAL PILOT STUDY FOLLOWED THE
tenets of the Declaration of Helsinki, complied with the Health Insurance Portability and Accountability Act of 1996, and was approved by Western Institutional Review Board (Olympia, Washington, USA). Written informed consent was obtained from all subjects. Participants enrolled in this study were recruited from volunteers who agreed to undergo the examinations and met the eligibility criteria described below at the Vitreous Retina Macula Consultants of New York (New York, USA) between February 2018 and October 2018. IMAGE ACQUISITION:
Retinal imaging for use in testing and training was collected retrospectively from scans acquired on normal eyes of healthy participants. Inclusion _ 20/20; (2) criteria were (1) best-corrected visual acuity > _ 21 mmHg, and (3) a spherical equivintraocular pressure < alent refractive error between 3 diopters (D) and þ1 D. Exclusion criteria were (1) history of any form of vitreoretinal diseases, high myopia, uveitis, glaucoma, or optic neuropathy; (2) media opacities that might prevent successful imaging; (3) prior intraocular surgery, laser treatment, or ocular trauma; (4) systemic or neurologic diseases that could affect retinal health, including diabetes, hypertension, dementia, or multiple sclerosis; and (5) current or prior use of systemic medications known to affect the retinal circulation. Subjects underwent a complete ophthalmic evaluation, including manifest refraction, uncorrected and bestcorrected visual acuity, intraocular pressure measurement, slit lamp biomicroscopy examination, ophthalmoscopic examination, high-resolution true color confocal color fundus photography (EIDON, CenterVue, Padua, Italy), and swept-source OCTA (SS-OCTA; PLEX Elite 9000, Carl Zeiss Meditec, Inc, Dublin, California, USA). Scan patterns acquired on the SS-OCTA device were the default 333-mm and 636-mm OCTA cube scans. Only scans with _8 were used in this study. a signal strength of >
TRAINING AND GRADING:
Nine 333-mm and nine 636-mm OCTA scan volumes centered on the fovea from 18 eyes of 14 healthy subjects (8 men and 6 women; 39.8 6 17.1 years of age, range 15-71 years) were used to create training and testing image sets for use in this study. A total of 147 vessels from the nine 333-mm image sets and 193 vessels from the nine 636-mm image sets were labeled with numbers for later assignment as artery or vein by study participants. Image sets for training and testing were created by exporting the automatically segmented default en face slabs (color retina depthencoded, grayscale full-thickness retina, SVP, and DCP with projection removal) from each OCTA volume and pasting them into a PowerPoint (Microsoft Corporation,
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
Redmond, Washington, USA) file, one slab per slide. Blood vessels were labeled with numbers in each slide. High-resolution color fundus photographs were used for identification of arteries and veins based on the following characteristics: arteries are brighter in color than veins; arteries are thinner than neighboring veins; the central light reflex is wider in arteries and smaller in veins; and arteries and veins usually alternate around the optic disc before branching out (Figure 1). Three of the authors (X.X., P.F.A., and K.B.F.) each independently reviewed and characterized each vessel used in the training and grading image sets. Subsequently, these authors reviewed their results together. Discordant gradings were rare. In those instances, following review, a consensus determination was achieved for each vessel used for training and testing. Nine ophthalmologists with minimal OCTA experience and no formal training in OCTA from 2 institutions (5 from Bascom Palmer Eye Institute, University of Miami Miller School of Medicine, Miami, Florida, USA; and 4 from Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangzhou, China), were instructed to classify all labeled vessels as arteries or veins in 3 stages. Three image sets, each composed of 1 randomly selected 333-mm and 636-mm study, were presented to each participant during each stage. Each scan was used only once during grading. In stage 1, untrained classification based on graders’ own assessment was performed. Prior to stage 2, graders were taught that a capillary-free zone was an anatomic feature of arteries. Prior to stage 3, graders were further trained to identify veins by recognizing their origin in vortices within the DCP (characterized as green convergence of capillaries on color depth-encoded slabs) and to consider that arteries and veins typically alternate as each vein drains capillary beds perfused by adjacent arteries in stage 3 (Figure 2). The training time prior to stage 2 and stage 3 was no more than 15 minutes. During training, typical examples of capillary-free zone/vortices within the DCP were shown to the graders using the image sets they had previously graded. Then each participant was asked to demonstrate an understanding of the technique by correctly identifying the vascular features on a new image set. Data including grading performances and the average grading time spent on each vessel in each stage regardless of the scanning patterns of the images used for grading were deidentified and recorded in a Microsoft Excel 2016 database (Microsoft Corporation). FIGURE 1. Examples of 333-mm (A) and 636-mm (B) optical coherence tomography (OCTA) image sets acquired from normal eyes and used for grading and training exercises. Vessels needed to be classified into arteries and veins are numbered on each slab: color retina depth-encoded image (A1 and B1), grayscale full-thickness retina (A2 and B2), superficial plexus (A3 and B3), and deep plexus with projection removal (A4 and B4). When preparing these image sets, the color
VOL. -
STATISTICAL
ANALYSIS: Descriptive results were presented as mean 6 standard deviation with range
depth-encoded OCTA images were aligned to high-resolution color fundus photographs using the retinal vessels as landmarks, which were then used to confirm the identity of each numbered vessel as either artery or vein (A5 and B5).
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
3
FIGURE 2. On a set of 333-mm optical coherence tomography (OCTA) images centered on the fovea, arteries (A), and veins (V) were classified and labeled on color retina depth-encoded (A), grayscale full-thickness retina (B), and superficial plexus (C) slabs. On the deep capillary plexus (DCP) slab with projection removal (D), vortices are marked with cyan dots showing the centers of capillaries convergence. The vessels are classified based on the following criteria: superficial arteries have adjacent capillary-free zones (E, magnification of the white solid line box in A, the arrowheads show the adjacent gaps free of capillaries), and the vortices in the DCP drain into veins (F, magnification of the white dotted box in A, magenta dots are put on the center of the convergence of capillaries). The classifications are confirmed using the high-resolution color fundus photograph (F) and its enlargement (G) centered on the fovea with the size of 333 mm.
4
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
TABLE 1. Average Grading Accuracy in Each Stage Subdivided Into Scanning Patterns (Mean 6 Standard Deviation; %) and Comparisons
TABLE 2. Sensitivity and Specificity in Classifying Retinal Veins in Each Stage Subdivided Into Scanning Patterns (%) Stage 1
Scanning Pattern
333 mm 636 mm Total P valueb
Stage 1
Stage 2
Stage 3
P Valuea
Scanning Pattern
49.9 6 16.3 51.4 620.8 50.4 6 17.0 .850
79.2 6 9.6 72.3 6 7.9 75.4 6 6.0 .020
96.9 6 3.1 93.2 6 3.3 94.7 6 2.6 .029
<.001 <.001 <.001 —
333 mm 44.96 636 mm 44.77 Total 44.85
Stage 2
Stage 3
Sensitivity Specificity Sensitivity Specificity Sensitivity Specificity
56.73 60.00 58.37
72.77 74.93 74.14
87.68 67.88 78.22
94.37 91.71 92.62
94.12 95.52 94.82
a
Comparisons of the grading accuracy between the 3 stages in each scanning pattern using generalized estimating equations. b Comparisons of the grading accuracy between 333-mm scan and 636-mm scan in each stage using generalized estimating equations.
(minimum, maximum). The Kolmogorov-Smirnov test and the Levene test were conducted to test the normality and homogeneity of variance, respectively. The correlation between accuracy and grading time, and between accuracy and graders’ year of practicing ophthalmology were analyzed using Pearson correlation coefficient. The accuracy between different grading stages and the accuracy between the 2 scanning patterns in each grading stage were compared using generalized estimating equations. Statistical analysis was conducted using SPSS version 22.0 (IBM Corp, Armonk, New York, USA) and a P value of <.05 was considered statistically significant.
RESULTS
TABLE 3. Correlations of Grading Accuracy With Grading Time in Each Stage
r 95% confidence interval of r P value
Stage 1
Stage 2
Stage 3
–0.013 –0.624, 0.679
0.518 0.057, 0.970
–0.164 –0.798, 0.741
.974
.189
.674
from arteries facilitate analyses of how the disease affects different vessel subtypes (Figure 3). Identifying the nature of a prior retinal vascular occlusion. A 61-year-old woman presented with the asymptomatic finding of localized thinning of the inner nuclear layer (INL) on OCT B-scans of the right eye. The OCT pattern was consistent with resolved paracentral acute middle maculopathy. Using our methods, it is possible to determine that a prior episode of cilioretinal artery hypoperfusion was responsible for producing OCT findings consistent with a previous ischemic event of the deep vascular complex (ie, paracentral acute middle maculopathy) (Figure 4).
OVERALL GRADING ACCURACY AND THE SEPARATE ACCU-
racies for the 333-mm and 636-mm image sets at each stage are summarized in Table 1. The sensitivity and specificity for identifying retinal veins at each stage are shown in Table 2. The average time spent on each vessel was (16.5 6 5.4 seconds), (8.5 6 3.6 seconds), and (15.6 6 8.1 seconds) in stages 1, 2, and 3, respectively. The average years in practice of 9 ophthalmologists was 6.4 6 4.2 years. Correlation between grading accuracy and mean time spent on each vessel during each stage are displayed in Table 3. Correlations between grading accuracy at each stage and graders’ mean years in practice were subdivided by scan pattern size and are displayed in Table 4. CLINICAL EXAMPLES:
Evaluating diabetic retinopathy. A 71-year-old woman was seen for follow-up of longstanding nonproliferative diabetic retinopathy associated with diabetic macular edema. She had received focal laser in the past, and more recently, intravitreal anti– vascular endothelial growth factor therapy as needed. The methods described herein for distinguishing veins
VOL. -
Exploring the vascular effects of macular telangiectasia type 2. A 64-year-old woman was referred for evaluation of bilateral metamorphopsia and reduced visual acuity to 20/150 in the right eye and 20/50 in the left eye. The diagnosis of macular telangiectasia type 2 was made based on characteristic clinical and imaging findings. Using our methods, it is easy to determine that the vascular damage is mostly concentrated around a vein draining the temporal aspect of the perifoveal capillary ring (Figure 5).
DISCUSSION IN THIS STUDY, WE DEMONSTRATE A SIMPLE METHOD TO
distinguish retinal arteries from veins in standard en face images acquired on a commercially available OCTA device. This technique uses the novel approach of identifying the capillary vortices in the DCP as an anatomic biomarker of venous origin. Ophthalmologists lacking OCTA
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
5
TABLE 4. Correlations of Grading Accuracy With Graders’ Years in Practicing Ophthalmology in Each Stage Subdivided Into Scanning Patterns Stage 1
r 95% confidence interval of r P value
Stage 2
Stage 3
333 mm
636 mm
333 mm
636 mm
333 mm
636 mm
0.587 –0.498, 0.926
0.608 0.322, 0.902
0.275 –0.354, 0.663
0.005 –0.670, 0.706
0.579 –0.008, 0.920
0.617 0.032, 0.930
.097
.082
.474
.990
.103
.077
FIGURE 3. Multimodal imaging of a 71-year-old woman with nonproliferative diabetic retinopathy in her right eye. (A) ‘‘True color’’ scanning laser ophthalmoscopy image shows microvascular abnormalities of diabetic retinopathy and evidence of prior focal laser for diabetic macular edema. (B) Fluorescein angiography (FA) at 1 minute 30 seconds shows scattered hyperfluorescent microaneurysms, staining laser scars, and minimal vascular leakage. A 9x9-mm color retina depth-encoded optical coherence tomography angiography study (C) is aligned to the FA image with veins labeled as ‘‘V’’ and arteries labeled as ‘‘A.’’ Some of the deep venous vortices used to identify veins are marked with magenta dots.
experience required only a brief training exercise to significantly improve their accuracy in grading retinal vessel type from a mean of 50.4% to a mean of 94.7%. Distinguishing arteries from veins on OCTA is of high clinical relevance because findings specific to each vessel type can add important information for detection and differentiation of many systemic and retinal diseases (Figures 3-5). The ability of OCTA to provide depthresolved images of retinal and choroidal vascular blood flow without the need for intravenous dye has made it a widely used imaging tool for studying retinal and optic nerve disorders including neovascular age-related macular degeneration, retinal vascular diseases, macular telangiectasia, pathologic myopia, inflammatory chorioretinal diseases, and glaucoma.26–28 Fast scanning speeds, improved retinal layer segmentation, and projection artifact removal algorithms29 have enabled current OCTA devices to resolve the discrete capillary plexuses within the retina. Using our training module, we found that grading accuracy increased from 50.4% in stage 1, to 75.4% in stage 2, and to 94.7% in stage 3. Postassignment interviews with graders expressed great difficulty grading vessels in stage 1 6
as the OCTA images lack visual clues, such as color, contrast, width, and shape, they would typically use to distinguish arteries from veins when evaluating other forms of retinal imaging. In addition, the graders could not trace to optic nerve because neither scan pattern includes the disc when it is centered on the fovea. In stage 2, graders used information from histologic studies of human retinas demonstrating that a capillaryfree zone is present around retinal arteries that develops during embryogenesis. This area exists where transmural oxygen diffuses to satisfy the metabolic demands of cells immediately adjacent to oxygen-rich superficial retinal arteries.30,31 Grading performance in stage 2 indicated that awareness of the periarterial capillary-free zone could improve accuracy to 75.4%. Graders had difficulty applying this technique to smaller arteries as the presence or absence of a capillary-free zone was more difficult to discern, and when they tried to follow the course of the larger arteries distal to arteriovenous crossings. In stage 3, graders were taught to identify veins by their origin in the vortices within the DCP which they were then able to trace back to arteriovenous crossings. When graders
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
FIGURE 4. Multimodal imaging of a 61-year-old woman with localized thinning of the inner nuclear layer (INL) in her right eye due to a prior episode of cilioretinal artery hypoperfusion. Near-infrared reflectance (A) shows no detectable lesion. Optical coherence tomography B-scans (B, C) corresponding to the green lines in A show INL thinning with preserved outer retinal bands indicating resolved paracentral acute middle maculopathy (PAMM). The green vortices in the 333-mm color retina depth-encoded optical coherence tomography angiography (OCTA) image (D) are used to distinguish veins (labeled ‘‘V’’) from arteries (labeled ‘‘A’’). Some of the vortices used to identify veins are marked with magenta dots. Flow is better preserved in the superficial vascular plexus (E) than in the deep capillary plexus (F). The OCTA study is useful to show that the reduced flow is in a retinal area supplied by a small cilioretinal artery indicated with blue arrows on color fundus photography (G) and fluorescein angiography (H).
VOL. -
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
7
FIGURE 5. Multimodal imaging of the left eye of a 56-year-old woman with macular telangiectasia type 2. ‘‘True-color’’ scanning laser ophthalmoscopy imaging (A) with enlargement of the foveal center (inset) shows loss of retinal transparency in the temporal fovea and a few small superficial retinal crystals. Fluorescein angiography (FA) at 1 minute 18 seconds (B) shows diffuse hyperfluorescence from the telangiectatic capillaries in the temporal fovea. The 333-mm color retina depth-encoded optical coherence tomography angiography image (C) aligned to the FA image (blue rectangle in B has veins labeled as ‘‘V’’ and arteries labeled as ‘‘A,’’ with some of the vortices used to identify veins marked with magenta dots. The vascular damage is mostly concentrated around a vein draining the temporal aspect of the perifoveal capillary ring. The green line in the near-infrared reflectance image (D) shows the location of a horizontal optical coherence tomography B-scan (E). The OCT B-scan shows central loss of both inner and outer retinal layers, hyporeflective intraretinal cavities, and preservation of a foveal depression.
had separate complementary strategies for identify arteries and veins, grading accuracy increased to 94.7%. In stages 2 and 3, grading accuracy for 333-mm scans was significantly higher than for 636-mm scans. This was likely related to different A-scan densities used in these 2 grid patterns. The 333-mm study uses 10-mm spacing between A-scans, whereas 12-mm spacing is used for the 636-mm scan pattern. Also, the 333-mm scan pattern uses 43 averaging for each B-scan whereas the 636-mm uses only 23 averaging/B-scan.32 Therefore, both the periarterial capillary-free zone and the DCP vortices were more easily recognized in the smaller 333-mm scan pattern. Also, because all studies were centered on the fovea, vessels at the superior, inferior, and temporal margins of 636-mm scans were often smaller than those on 333-mm scans, making it more difficult to recognize these important scan features. One technique for enhancing vascular detail in 636-mm and larger scan patterns is to average multiple tracked scan acquisitions (Supplemental Figure 1).33 Graders uniformly reported that among the 4 slabs provided in each image set, the color depth-encoded slab was the most useful in all aspects of grading. Color coding of the superficial and deep vessels made identifying the connection of veins to the deep capillary vortices simply 8
a matter identifying the convergence of thin green linear structures. After completing the training exercises, most graders felt that this single visualization would be sufficient for grading, whereas the 3 additional slabs were typically ignored in stage 3, as they seemed to provide no additional distinguishing information. Interestingly, grading accuracy was not statistically correlated with grading time or graders’ years in practice. For all participants, training grading sessions were completed in less than 15 minutes. These results suggest that the vessel classification method described herein can be used to train nonophthalmologists, including graders in imaging reading centers. We are currently exploring whether our methods can be applied to an artificial intelligence or a machine learning based automatic vessel classification algorithm. Potential limitations of our methods include the lack of a separate control group asked to grade the 3 successive image sets without training, but with some feedback regarding grading accuracy at each stage. It is possible that some of the improvement in grading accuracy could result from corrective feedback alone, not the specific techniques taught during the training exercises. To minimize this concern, graders were not provided complete answers for
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019
the prior image sets, and training sessions for the next stage included just a few vessels that had been graded during the prior session. Another important limitation is that this study used scans obtained on only one of several commercially available OCTA instruments. Currently, the review software provided by other OCTA device makers does not include an en face color depth-encoded visualization, which was instrumental for the technique described in this study. However, we have successfully applied our method to standard acquisitions obtained on other OCTA devices by color coding either the superficial or deep slab and combining both into a composite image using ImageJ (Rasband, W.S., ImageJ, US National Institutes of Health, Bethesda, Maryland, USA, https://imagej.nih.gov/ij/) or Adobe Photoshop (Adobe Inc, San Jose, California, USA) (Supplemental Figure 2). Although this extra step would likely be too cumbersome to use in general practice, other OCTA device manufacturers could potentially provide a color retina depth-encoded image in future versions of their review software.
Finally, this study did not formally evaluate grading accuracy in eyes with retinal pathology or in scans with poor quality due to motion artifacts, low signal strength, or segmentation errors. The application of our technique in lower-quality images or diseased eyes is more challenging and limits its implementation in certain situations. As our training and testing subjects were limited to junior ophthalmologists from only 2 eye institutes having little OCTA experience, larger cohorts, with more diverse career backgrounds, would be necessary for a more hierarchical performance assessment. Finally, grading time in each stage was not subdivided by scanning patterns. Because the differences of grading accuracy of 333-mm scans and 636-mm scans were of statistical significance in stage 2 and stage 3, it would have been preferable to compare the time used for grading with each scanning pattern. In conclusion, we describe a novel method for accurately distinguishing retinal arteries from veins on OCTA that incorporates the use of vortices in the DCP to identify venous origin.
ALL AUTHORS HAVE COMPLETED AND SUBMITTED THE ICMJE FORM FOR DISCLOSURE OF POTENTIAL CONFLICTS OF INTEREST and the following were reported: D.S. is a consultant to Genentech, Heidelberg Engineering, Amgen, Bayer, Novartis, Optovue, Regeneron. K.B.F. is a consultant for Optovue, Heidelberg Engineering, Zeiss, Allergan, and Novartis and he receives research support from Genentech/Roche. This work was supported by the Macula Foundation, New York, NY; National Natural Science Foundation of China (81800879); Fundamental Research Funds of the State Key Laboratory of Ophthalmology, China (2018KF04 & 2017QN05); and Natural Science Foundation of Guangdong Province (2017A030310372). The authors thank Drs. Shasha Yang, Yao Yang, Linlin Hao, Shufen Lin and Hui Xiao from Zhongshan Ophthalmic Center, Sun Yat-sen University, China, for their help and contribution to this study. All authors attest that they meet the current ICMJE requirements to qualify as authors.
REFERENCES 1. Suzuki Y. Direct measurement of retinal vessel diameter: comparison with microdensitometric methods based on fundus photographs. Surv Ophthalmol 1995;39(Suppl 1): S57–S65. 2. Qiu C, Cotch MF, Sigurdsson S, et al. Microvascular lesions in the brain and retina: the age, gene/environment susceptibility-Reykjavik study. Ann Neurol 2009;65(5): 569–576. 3. Wong TY, Klein R, Sharrett AR, et al. The prevalence and risk factors of retinal microvascular abnormalities in older persons: the Cardiovascular Health Study. Ophthalmology 2003;110(4):658–666. 4. Ikram MK, De Jong FJ, Van Dijk EJ, et al. Retinal vessel diameters and cerebral small vessel disease: the Rotterdam Scan Study. Brain 2006;129(Pt 1):182–188. 5. Shin YU, Lee DE, Kang MH, et al. Optical coherence tomography angiography analysis of changes in the retina and the choroid after haemodialysis. Sci Rep 2018;8(1):17184. 6. Kondermann C, Kondermann D, Yan M. Blood vessel classification into arteries and veins in retinal images. Medical Imaging: SPIE 2007;9. 7. Welikala RA, Foster PJ, Whincup PH, et al. Automated arteriole and venule classification using deep learning for retinal images from the UK Biobank cohort. Comput Biol Med 2017; 90:23–32.
VOL. -
8. Shao Q, Heussen FM, Ouyang Y, Hager A. Retinal vessel diameter changes in different severities of diabetic retinopathy by SD-OCT. Eur J Ophthalmol 2016;26(4):342–346. 9. Wang Y, Fawzi AA, Varma R, et al. Pilot study of optical coherence tomography measurement of retinal blood flow in retinal and optic nerve diseases. Invest Ophthalmol Vis Sci 2011;52(2):840–845. 10. Ouyang Y, Shao Q, Scharf D, Joussen AM, Heussen FM. An easy method to differentiate retinal arteries from veins by spectral domain optical coherence tomography: retrospective, observational case series. BMC Ophthalmol 2014;14:66. 11. Iwase T, Ra E, Yamamoto K, Kaneko H, Ito Y, Terasaki H. Differences of retinal blood flow between arteries and veins determined by laser speckle flowgraphy in healthy subjects. Medicine (Baltimore) 2015;94(33):e1256. 12. Gu B, Wang X, Twa MD, Tam J, Girkin CA, Zhang Y. Noninvasive in vivo characterization of erythrocyte motion in human retinal capillaries using high-speed adaptive optics nearconfocal imaging. Biomed Opt Express 2018;9(8):3653–3677. 13. Alam M, Toslak D, Lim JI, Yao X. Color fundus image guided artery-vein differentiation in optical coherence tomography angiography. Invest Ophthalmol Vis Sci 2018;59(12): 4953–4962. 14. Spaide RF, Klancnik JM Jr, Cooney MJ. Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. JAMA Ophthalmol 2015; 133(1):45–50.
OCTA-BASED RETINAL ARTERY/VEIN CLASSIFICATION
9
15. Zhang M, Hwang TS, Campbell JP, et al. Projection-resolved optical coherence tomographic angiography. Biomed Opt Express 2016;7(3):816–828. 16. Campbell JP, Zhang M, Hwang TS, et al. Detailed vascular anatomy of the human retina by projection-resolved optical coherence tomography angiography. Sci Rep 2017;7:42201. 17. Spaide RF, Curcio CA. Evaluation of segmentation of the superficial and deep vascular layers of the retina by optical coherence tomography angiography instruments in normal eyes. JAMA Ophthalmol 2017;135(3):259–262. 18. Gattoussi S, Freund KB. Correlating structural and angiographic optical coherence tomography in the intermediate and deep retinal capillary plexuses. Exp Eye Res 2017;165: 96–98. 19. Freund KB, Sarraf D, Leong BCS, Garrity ST, Vupparaboina KK, Dansingani KK. Association of optical coherence tomography angiography of collaterals in retinal vein occlusion with major venous outflow through the deep vascular complex. JAMA Ophthalmol 2018;136(11): 1262–1270. 20. Paques M, Tadayoni R, Sercombe R, et al. Structural and hemodynamic analysis of the mouse retinal microcirculation. Invest Ophthalmol Vis Sci 2003;44(11):4960–4967. 21. Genevois O, Paques M, Simonutti M, et al. Microvascular remodeling after occlusion-recanalization of a branch retinal vein in rats. Invest Ophthalmol Vis Sci 2004;45(2):594–600. 22. Foulds WS, Kek WK, Luu CD, Song IC, Kaur C. A porcine model of selective retinal capillary closure induced by embolization with fluorescent microspheres. Invest Ophthalmol Vis Sci 2010;51(12):6700–6709. 23. Foreman DM, Bagley S, Moore J, Ireland GW, McLeod D, Boulton ME. Three dimensional analysis of the retinal vasculature using immunofluorescent staining and confocal laser scanning microscopy. Br J Ophthalmol 1996;80(3):246–251.
10
24. Garrity ST, Paques M, Gaudric A, Freund KB, Sarraf D. Considerations in the understanding of venous outflow in the retinal capillary plexus. Retina 2017;37(10):1809–1812. 25. Balaratnasingam C, An D, Sakurada Y, et al. Comparisons between histology and optical coherence tomography angiography of the periarterial capillary-free zone. Am J Ophthalmol 2018;189:55–64. 26. Kashani AH, Chen CL, Gahm JK, et al. Optical coherence tomography angiography: a comprehensive review of current methods and clinical applications. Prog Retin Eye Res 2017; 60:66–100. 27. Hagag AM, Gao SS, Jia Y, Huang D. Optical coherence tomography angiography: technical principles and clinical applications in ophthalmology. Taiwan J Ophthalmol 2017; 7(3):115–129. 28. Tan ACS, Tan GS, Denniston AK, et al. An overview of the clinical applications of optical coherence tomography angiography. Eye (Lond) 2018;32(2):262–286. 29. Spaide RF, Fujimoto JG, Waheed NK. Image artifacts in optical coherence tomography angiography. Retina 2015; 35(11):2163–2180. 30. Michaelson IC, Campbell ACP. The anatomy of the finer retinal vessels, and some observations on their significance in certain retinal diseases. Trans Ophthalmol Soc UK 1940; 60:40. 31. McLeod D. Krogh cylinders in retinal development, panretinal hypoperfusion and diabetic retinopathy. Acta Ophthalmol 2010;88(8):817–835. 32. Zheng F, Zhang Q, Shi Y, et al. Age-dependent changes in the macular choriocapillaris of normal eyes imaged with swept-source OCT angiography. Am J Ophthalmol 2019;200: 110–122. 33. Pappuru RR, Briceno C, Ouyang Y, Walsh AC, Sadda SR. Clinical significance of B-scan averaging with SD-OCT. Ophthalmic Surg Lasers Imaging 2012;43(1):63–68.
AMERICAN JOURNAL OF OPHTHALMOLOGY
--- 2019