Towards standardizing retinal optical coherence tomography angiography: a review

The visualization and assessment of retinal microvasculature are important in the study, diagnosis, monitoring, and guidance of treatment of ocular and systemic diseases. With the introduction of optical coherence tomography angiography (OCTA), it has become possible to visualize the retinal microvasculature volumetrically and without a contrast agent. Many lab-based and commercial clinical instruments, imaging protocols and data analysis methods and metrics, have been applied, often inconsistently, resulting in a confusing picture that represents a major barrier to progress in applying OCTA to reduce the burden of disease. Open data and software sharing, and cross-comparison and pooling of data from different studies are rare. These inabilities have impeded building the large databases of annotated OCTA images of healthy and diseased retinas that are necessary to study and define characteristics of specific conditions. This paper addresses the steps needed to standardize OCTA imaging of the human retina to address these limitations. Through review of the OCTA literature, we identify issues and inconsistencies and propose minimum standards for imaging protocols, data analysis methods, metrics, reporting of findings, and clinical practice and, where this is not possible, we identify areas that require further investigation. We hope that this paper will encourage the unification of imaging protocols in OCTA, promote transparency in the process of data collection, analysis, and reporting, and facilitate increasing the impact of OCTA on retinal healthcare delivery and life science investigations. Standardized imaging protocols, data analysis methods, reporting of metrics and trial outcome measures will enhance the impact of OCTA on improving retinal healthcare.


Introduction
Biomedical optics offers many promising tools for clinical medical imaging and diagnostics that are attractive because they are non-invasive, portable, and often low cost 1 . New advances in optics and photonics technologies will continue to shape the future of biomedical research and clinical applications 2 . To fully capitalize on the potential of these promising tools to contribute to better human health, much more effort is needed in introducing and applying standards: in instrumentation, imaging protocols, data analysis methods, and reporting of findings. Such standards would facilitate open data and software sharing, and cross-comparison and pooling of data from different studies [3][4][5][6] , which, in turn, would facilitate the building of large, annotated image databases derived from healthy and diseased subjects. The availability of such databases is a critical step in the study and improvement of clinical management of specific diseases to reduce their burden and improve human health 7 . The broad need has been widely recognized by regulatory agencies, including the U.S. Food and Drug Administration (FDA) and The European Medicines Agency, by subject-specific societies 6 , and by research consortiums, including the Lifetime Initiative 8

and The European
Institute for Biomedical Imaging Research 9 .
In this paper, we set out and make the case for the steps needed to standardize optical coherence tomography angiography (OCTA) imaging of the retinal microvasculature. Through review of imaging protocols, data analysis methods and metrics, reporting of findings, and clinical practice of retinal OCTA, we identify and offer guidance on the standards needed, as well as identify areas requiring further investigation before standards can emerge. The imaging and quantitative analysis of retinal microvasculature are not only important for the study, diagnosis, monitoring, and guidance of treatment of ocular conditions 10 , such as age-related macular degeneration, diabetic retinopathy, vascular occlusion, and glaucoma, but also for systemic conditions, such as cardiometabolic and neurodegenerative diseases 11 . OCTA is an emerging technique attractive to clinicians as it is quickly carried out, non-invasive-not requiring administration of dye-and, distinctively, can provide images of the vascular network from different retinal depths 5 . OCTA derives from optical coherence tomography (OCT), which has proven to be a game changer in ophthalmic care and become the most performed imaging procedure in ophthalmology 12,13 . OCTA has not yet become a standard imaging tool in ophthalmic care, but the field is growing strongly. A continuing increase in the number of publications year on year ( Fig. 1) highlights the importance of OCTA to the scientific and clinical research communities and makes it an ideal case study on how to introduce standards in biomedical optics more generally.
The importance of standardization of imaging data in eye healthcare has recently been highlighted by the American Academy of Ophthalmology: "Standardization would advance the needs and interests of ophthalmologists, their patients, and the quality of clinical care by promoting interoperability, enabling the creation of comprehensive datasets for research and big data analyses, and developing algorithms for machine learning and artificial intelligence" 14 . As well, standardization has been discussed in the specific context of OCTA 15 and there have been initial steps taken toward harmonization of the nomenclature and reporting of findings for common retinal diseases 5,[16][17][18][19] . We build on the evidence presented in these and other previous non-OCTA initiatives 4,5,7,18,[20][21][22][23] to advocate clear, rigorous, and easy-to-apply terminology and procedures for physicians, researchers, and OCTA device manufacturers toward the standardization of both technical and clinical aspects of retinal OCTA.
We first set the scene by briefly discussing the retina and its blood supply, the origins and technical background of OCTA, and its application in clinical practice. We then review current imaging protocols, data analysis methods, metrics, and clinical practice. We identify existing discrepancies and inconsistencies and, where possible, propose minimum standards for image acquisition, analysis, and reporting of findings. Where not possible, we identify areas requiring further investigation. We do not review OCTA signal generation and processing, as these topics have been well covered in other reviews 10,[24][25][26][27] . Due to space limitations and its distinctiveness, we also largely exclude choroidal circulation.
We hope that this paper will contribute to global momentum for the establishment of guidelines for Pubmed, with "optical coherence tomography angiography", "OCT-angiography" and "retina" as the search key words. Data retrieved on 28 January 2022 standardized and transparent data collection, analysis, and reporting in OCTA and in biomedical optics more generally. For OCTA, such standards will accelerate scientific discovery and its establishment as a standard of care in clinical medicine.

The retina, its blood supply, and clinical significance
The retina is a semi-transparent neural tissue that lines the internal rear (posterior) surface of the eye. Its function is to generate visual perception by converting the projected image of the world created by the optics of the eye into electrical signals that are transmitted via the optic nerve to the brain. Despite having an average thickness of only 300 µm, the retina has an enormous capacity for parallel processing of complex visual signals, due to the multilayered and pixelated organization of retinal neurons 28 . A schematic of the retinal blood supply and fovea is shown in Fig. 2. The foveola, circa 350 μm in diameter, is the avascular central region of the retina with maximum visual acuity. The high information processing demands of visual perception places a high energy demand on the retina 29,30 and, therefore, maintenance of adequate oxygen supply and waste exchange within the retina are critical for its function. Oxygen is delivered via a combination of retinal and choroidal microcirculation 29,31 . The inner retinal microcirculation (Fig. 2) is fed from arterioles, flows to capillaries in three layers, and drains through venules; with all vessels having a diameter of 100 µm or less 32 . The microvascular architecture of the inner retina is comparatively sparse transversally and axially because light must pass through vascular regions to the photoreceptors in the outer retina and vessels are highly scattering and disrupt image formation 32 .
The macula, defined as the central retinal region of circa 5.5 mm in diameter, is responsible for high-resolution and color vision, also contains the fovea. Mean capillary flow velocity in the macula is in the range 1.4-3.3 mm/s, which is comparable to that of the brain and higher than in the gut mesentery, for example, for which the mean velocity is 0.8 mm/s 33 . The velocity ranges of blood flow across the whole retina are 1.8-7.2 mm/s in the retinal arterioles and venules and 0.2-3.3 mm/s in the retinal capillaries 34 . By comparison, the ranges for choroidal arterioles and venules and choroidal capillaries (choriocapillaris) are 5.5-12 mm/s and 0.3-3.6 mm/s, respectively.
The outer retinal layers (Fig. 2) are avascular and depend on diffusion from the inner retinal and choroidal microcirculations for metabolic support 32 ). SCP, ICP, and DCP are thin vascular complexes that are connected to one another and receive the arteriolar supply (red) from SVP arterioles and drain (blue) to SVP venules. NFL, nerve fiber layer; GCL, ganglion cell layer; IPL, inner plexiform layer; INL, inner nuclear layer; OPL, outer plexiform layer; HFL, Henle fiber layer; ONL, outer nuclear layer; PR, photoreceptors; RPE, retinal pigment epithelium. The Bruch's membrane (BM) is just below the RPE. The inner retina is defined as the NFL to OPL, and the outer retina (which contains no vessels) as the HFL, ONL, PR, and RPE. Not shown: in the peripapillary retina (near the optic disc, not shown here but presented in Fig. 3), an additional vascular plexus is present: RPCP (radial peripapillary capillary plexus). RPCP and SVP together are commonly known as the SVC (superficial vascular complex). Adapted with permission from Nesper and Fawzi 189 microcirculation in the inner retinal layers vary between layers and across retinal locations to meet the distinct and varied metabolic demands of the local cellular environment ( Fig. 3) 30,32,35 . The high overall oxygen demands of the retina and relatively sparse nature of its microvasculature require dynamic, well-regulated blood flow. In contrast, the choroidal vasculature is denser and lacks this regulatory ability 30 . Figure 3 further illustrates the retinal microvascular network. Human volunteer fundus photography is correlated with OCT cross-sectional imaging.
As well, confocal microscopy shows the retinal microvascular network at different depths and eccentricities of perfused donor retinas (ex vivo).
The visualization and characterization of microvascular networks in different spatial domains of the retina is vitally important for improving the understanding, classification, and staging of a wide range of common diseases with a vascular component (either as cause or effect) 36 , including age-related macular degeneration (AMD), diabetic retinopathy, vascular occlusion, coronary heart disease, and Alzheimer's disease. Even small reductions in blood flow have been shown to have deleterious effects on retinal function 30 . Vessel density reduction in diabetic retinopathy occurs in multiple retinal locations but most commonly in the deep capillary plexus 37 . An abnormal superficial vascular complex in the optic nerve can be present in glaucoma, whereas, an abnormal macular deep capillary plexus can be present in retinitis pigmentosa 38 . The peripapillary circulation region, located at the margin of the optic nerve head, is a critical site in the pathogenesis of retinal vein occlusion, optic disc rim (i.e., Drance) hemorrhage, and glaucoma 39,40 . Vessel density reduction may be present in all retinal and choroidal layers in coronary heart disease 41 and present in the superficial vascular complex in Alzheimer'disease 42 .

Overview of optical coherence tomography angiography and retinal applications
We have identified over 300 clinical trials registered at www.clinicaltrials.gov evaluating OCTA as a diagnostic tool for screening of retinal and choroidal dysfunction in ocular and systemic diseases. In its most common configuration, OCT captures, without contrast agent, a volumetric image dependent on the scattering properties of the tissue by transversally scanning a focused beam and collecting the depth-encoded scattering profile (termed A-scan) through (spectral-domain) low-coherence interferometry. Transverse scanning captures a line of A-scans termed a B-scan, and defines the fast (x) axis. Slower scanning in the orthogonal transverse direction (slow y axis) then completes a volume scan. OCTA enables volumetric visualization of retinal vasculature from the motion contrast generated mainly by scattering from erythrocytes (red blood cells) in flowing blood causing the speckled OCT signal to fluctuate. OCTA images are extracted by analysis of the time-varying OCT signal (intensity and/or phase) between acquisitions, most commonly cross-sectional scans (B-scans) at the same or closely adjacent locations 38 . The processed volume created by the stack of OCTA B-scans is usually presented as a maximum intensity projection image in the transverse plane over a specified depth range.
The use of fluctuating optical speckle to characterize tissue substantially predates OCTA; even its use in OCT dates to the 1990s. Its use in characterizing flow dates back to at least Barton and Stromski in 2005 43 followed by the seminal work of Makita et al. 44 , Wang et al. 45 , and Fingler et al. 46,47 , in developing and applying OCTA for retinal imaging. Although, Makita et al. introduced the term "optical coherence angiography" and Wang et al. "optical angiography", the acronym OCTA has become widely adopted 48 . As yet, OCTA methods do not allow accurate blood flow quantification 49 , but there are ongoing attempts to solve this difficult problem 50 . The first commercial OCTA instrument, the spectral-domain (SD) OCTA from Optovue, was approved by the FDA in 2015 51 . In the same year, the first commercial sweptsource (SS) OCTA system was introduced by Topcon 10 . Currently established commercial OCTA device suppliers include Optovue, Topcon, Heidelberg Engineering, Carl Zeiss Meditech, Canon, Optopol, and Nidek 25 .
Commercial instruments usually acquire 70,000-100,000 A-scans per second using a light source centered at a wavelength of either 840 nm or 1050 nm, for SD-OCTA or SS-OCTA, respectively, with an interscan time (ΔT, the interval between B-scans) of 4-5 ms, and achieving optical axial resolution (product of physical resolution and group refractive index of the medium) of 5-10 µm and transverse resolution of circa 20 µm; these resolutions are the full-width at half-maximum in tissue 10 . Research instruments can operate at substantially higher acquisition speeds, up to megahertz A-scan rates 52 , ΔT of <1.5 ms 10 , and resolutions as high as circa 3 µm 53 .
In principle, more than two B-scans can be recorded at a single location and used to generate multiple interscan times, with associated data storage and processing overhead 54 . Commercial instruments usually collect between two and four B-scans at the same location 10 . The value of the interscan time determines the maximum flow velocity capable of generating speckle contrast and, therefore, which vessels are visible in the OCTA image 10,34 . As discussed, blood flow velocity in the retinal circulation can be anywhere in the range 0.2-7 mm/s. For retinal vessels oriented perpendicular to the imaging beam, the maximum distinguishable velocity can be found by dividing the transverse resolution by ΔT; therefore, for ΔT = 1.5 ms and ΔT = 5 ms and transverse resolution of 20 µm, maximum speeds are 13 mm/s and 4 mm/s,  respectively. These values are reduced for nonperpendicular vessels. On the other hand, the longer the interscan time, the better the sensitivity to low flow velocities, as more speckle decorrelation occurs between scans; thus, there is a trade-off to be struck. There remains strong continuing interest in quantifying blood flow with OCTA. Low decorrelation values are proportional to flow velocity but full decorrelation leads to a saturation velocity value. Thus, there is a further trade-off in selection of interscan times that is outside of the current scope of our discussion 10 .
The resolution of the OCTA images-in all three spatial dimensions-represents a trade-off between visibility of vessels, their accurate representation, and sensitivity to motion artifact. The standard transverse resolution of ∼20 µm of commercial systems enables imaging over several hundred micrometers in axial depth, typically covering an angular field of view (FOV) of between 10 × 10 and 30 × 30 degrees that corresponds to a linear FOV of between 3 × 3 mm and 9 × 9 mm on the retinal surface of a typical eye. (The exact correspondence depends on the subject: correction of transverse image magnification error is further discussed below.) At the standard transverse resolution, the caliber of some capillaries appears wider than their true width 38 . High transverse resolution (below 10 µm, only available from research systems) helps to resolve small features of the microvascular network 53 , which is especially important for visualizing the densely packed choriocapillaris. However, increased resolution is usually obtained over a reduced imaging depth and FOV and makes the measurement more sensitive to motion artifacts that degrade the quality of OCTA images. Any eye movement sufficient to cause OCT speckle to decorrelate between acquisitions (as a rule of thumb, requiring motion with magnitude on the order of the resolution) will generate OCTA contrast. Bulk motion artifact correction allows efficient removal of axial eye motion artifact; transverse eye motion correction is not so easily achievable.
There have been many algorithms introduced to generate OCTA maps and these can be divided into phase-, amplitude-and complex-signal methods 24,38,49,55 . Phasebased methods, although capable of generating excellent quality OCTA images 44,46,47,56 , require precise removal of background phase noise caused by eye bulk motion or from system instabilities, which is challenging 57,58 . Therefore, such methods have not yet been applied in commercial instruments. Amplitude-based methods 43,59 are applied in most commercial instruments (e.g., Optovue, Topcon, and Heidelberg) 38 . Complex-signal-based methods demonstrate high sensitivity to low-velocity flow; however, they are sensitive to eye movements too. Optical microangiography (OMAG) is one such prominent research example 45 also used commercially by Zeiss 38 .
Additional post-processing can be applied to OCTA datasets to enhance the quality of the final angiogram. The split-spectrum amplitude-decorrelation angiography (SSADA) algorithm was the first to be used in a commercial instrument (Optovue) 10 . SSADA splits the OCT optical frequency spectrum into bins to reconstruct multiple lower axial resolution images to reduce the effects of axial eye motion, before recombining them to regain the signal-to-noise ratio 60 . As well, averaging of multiple OCTA datasets 61 produces valuable improvements in OCTA image quality, as we discuss below.
The current commonly used clinical angiography techniques are fluorescein angiography (FA) for the retina and indocyanine green angiography (ICGA) for the choroid. OCTA enjoys the advantage over these that it does not require a contrast agent; thereby, avoiding the rare adverse patient reactions to the dye 62 . As well, OCTA acquisition times can be as short as a few seconds compared to the 5-10 min required (after dye administration) to capture the time-lapse sequences required in FA/ ICGA 58 . Perhaps the greatest advantage of OCTA is the resolution in depth from the volumetric images not available from the two-dimensional FA/ICGA images. In particular, OCTA provides better visualization of the deep capillary plexus and choroidal vasculature compared with FA and ICGA 63,64 (Fig. 4a, b). Conventional FA/ICGA images 30-50 degrees of the retina, however, ultra-widefield FA/ICGA imaging systems have been developed to image the retina spanning up to 200 degrees 65 . As mentioned, typical clinical OCTA angular FOVs cover 10-30 degrees. Lab-based OCTA instruments have achieved up to 70 degrees in a single measurement 66 and 100-degrees by applying a montage scanning protocol 67 . Whilst FOVs for FA/ICGA and OCTA have become very comparable, OCTA cannot visualize leakage or vascular permeability important for clinical diagnosis and management 10 .
OCTA demonstrates good correlation with histology ( Fig. 4c) 49 but tends to yield larger vessel area densities and larger capillary diameters 68 , which has been confirmed by retinal vessel diameter measurements comparing SS-OCT to adaptive-optics ophthalmoscopy 69 .
OCTA has already enabled novel discoveries in ophthalmology 55,58,70,71 . For example, OCTA has provided insight into the relative importance of the superficial versus deep capillary plexuses in macular telangiectasia 72 and diabetic retinopathy 73 . Promising early-stage labbased research includes the study of retinal vascular function, including monitoring of pulse wave velocity 74 , blood flow heterogeneity 32 , and response to external stimulation 75 . Vascular pulsatility is a field of the future for OCTA, with early work in the retina based on Doppler OCT, and early work using OCTA directed at skin imaging 76,77 . Looking ahead, OCTA may also be beneficial for neurological research 51,78 to advance understanding of pathophysiology of multiple sclerosis 79 , Alzheimer's disease 42,80-84 , various optic neuropathies 51 , and diagnosis of cerebral small vessel disease 85 .

Review and standardization of OCTA imaging protocols
In this section, we discuss key OCTA scanning protocols and imaging parameters to support standardizing image collection procedures and their reporting. We include the important emerging area of widefield OCTA and pay particular attention to the harmonization of terminology.

Scanning field of view and sampling density
We first consider the effect of the FOV on raster scanning, which is commonly used in both commercial and non-commercial instruments 86 .
The OCTA image construction process is illustrated in Fig. 5. The linear scan length defines the FOV and the combination of scan length, number of A-scans per Bscan, and number of OCTA B-scans enables calculation of the transverse scanning density (usually termed sampling density) in the x and y directions. Sampling density is calculated as: ρ x,y = Δx/scan separation x,y , where Δx is the OCT transverse resolution, scan separation x,y is the scan length divided by the number of samples (either number of A-scans along the x-axis or OCTA B-scans along the yaxis). A sampling density of 2 represents the Nyquist limit. A value of 2.5 or higher would ideally be used to provide a margin to accommodate non-ideal factors but, in practice, commercial systems routinely use sub-Nyquist sampling.
In practice, the choice of FOV and sampling density requires compromise. A larger FOV is often associated with lower sampling density to maintain the same OCTA volume acquisition time. The lower sampling density results in a lower image resolution and detection of fewer smaller vessels. Analysis of images produced at lower sampling density leads to lower image-grading agreement and reduced vessel density measures 55,[87][88][89] . Reduced sampling density may additionally affect the accuracy of vessel diameter assessment, already biased by the system's transverse resolution, at all FOVs 69 .
3 × 3-mm FOV is often chosen as a good compromise to maintain adequate sampling density and image acquisition time. For example, for OCTA parameters of 4 B-scans per y location, 100,000 A-lines per second and 300 × 300 samples per FOV, total acquisition time is 4 s. For lower numbers of B-scans at each y location, this time can be further reduced, but at the cost of OCTA image quality. Not all commercial platforms keep sampling density constant as FOV changes. For instance, RTVue XR Avanti (Optovue, Inc., Fremont, CA, USA) offers retinal scanning protocols: either 3 × 3-mm or 6 × 6-mm FOV with fixed 304 × 304 A-scan sampling, sampling densities of 2 and 1 (sub-Nyquist), respectively, for a transverse resolution of circa 20 µm. Plex Elite 9000 (Carl Zeiss Meditec, Dublin, CA, USA) provides a 3 × 3-mm FOV with 300 × 300 A-scans and a 6 × 6-mm FOV with 500 × 500 A-scans, partially compensating sampling density as the FOV increases, at 2 and then 1.7, also for circa 20 µm transverse resolution. Due to the limitation on acquisition speed of current OCTA systems, sampling densities of 1.7 and above represent a reasonable compromise for gross OCTA image quantification. The importance of adequate sampling density is highlighted in Fig. 6.
We recommend reporting the sampling density (currently rarely reported) in addition to the transverse resolution and image size, to enable interpretation of all analyses, especially quantitative ones.

Widefield OCTA
Widefield OCTA imaging extends the raster scanning protocol to achieve much larger FOVs. Images are captured as single shot, as demonstrated, e.g., by Polans 66 and Wei 90 , or by mosaicking/montaging of scans, as shown by Hendargo 91 and Zhang 67 . A wide variety of image sizes and montages have been reported, from a single scan of 12 × 12 mm to montage scans comprising: two 15 × 9-mm scans; five scans of 12 × 12-mm each; and five scans of 6 × 10-mm each 5 . The term "ultra-widefield" OCTA has been introduced to describe 100-degree angular FOV OCTA imaging comprising a 4 × 4 grid of 6 × 6-mm scans 67 . Recently, three slightly different definitions have been proposed for widefield and ultra-wide-wield OCTA by Munk et al. 5 , Choudhry et al. 92 , and Pichi et al. 18 . Based on their investigations and discussions, we recommend defining widefield and ultra-widefield OCTA, independent of the use of montaging, as producing an image in the angular FOV range 70-110 degrees and above 110 degrees, respectively; this is aligned with definitions commonly applied in retinal fundus photography 93 .

Review and standardization of OCTA data analysis procedures and terminology
In this section, we discuss a wide range of OCTA data analysis procedures and terminology applied to collected OCTA datasets to make them ready for clinical evaluation. We include OCTA image processing, retinal layer segmentation and definition, image artifacts, image averaging, image quality indicators and classification, and correction of transverse image magnification error.

En face OCTA image processing
En face OCTA maps (also termed C-scans) can be created from the volumetric OCTA dataset from any retinal layer of interest. These maps are usually calculated by mean projection (the average value of voxels at a given x, y position over a given depth range) or maximum intensity projection (MIP, the maximum value voxel in depth). The MIP, in effect, ignores the overlap of multiple vessels or projection (shadow) artifacts (described below), thereby, providing a cleaner angiogram 38,94 . Several processing steps are commonly applied to en face OCTA images before quantitative analysis is undertaken 95 , including image filtering (e.g., using a Frangi "vesselness" filter to improve vessel visualization) 96 , binarization (e.g., using Otsu, global, fuzzy, or other thresholding methods) [97][98][99] , and image skeletonization. As well, the correction of vessel discontinuity in the OCTA binary image, by filling the gaps between vessel segments, has been reported 23 . Because of the fine line between appropriate interpolation and inappropriate image manipulation, we recommend not to apply this procedure to avoid ambiguous manipulation of images but, if applied, a rigorous description of steps taken should be given. The recent review papers by Tan et al. 95 and Meilburger et al. 100 103 . These studies highlight the need for consensus within the community regarding the optimum pipeline for OCTA image processing 95 . More comparative studies, such as those of Gorczynska et al. 104 , Giarratano et al. 105 , and Dadkhah et al. 106 , are needed to benchmark the available algorithms and select the best candidates for image processing and segmentation. Better open access analysis code and better datasets of the same individual imaged on several platforms, to use as a baseline or benchmark, would be very helpful in realizing this objective. Well-defined, optimized pipelines for OCTA en face image processing and analysis will be critical precursors for the successful conduct of large-scale multicenter studies to pave the way for the development and validation of OCTA-based biomarkers 21 .

Retinal layer segmentation and definition
Several algorithms have been introduced enabling automated segmentation of retinal layers based on computer vision 91,107-109 and deep learning 110,111 . Commercial instruments usually have pre-set layers of interest segmented through an automated process; however, expert manual correction is often required, especially in diseased retinas 58 . More work is still required to find better, more universal methods for automated segmentation of OCTA images that do not require manual intervention 49,112,113 . The pre-set layers and even the definition of layers vary slightly between commercial instruments, which also holds true for non-commercial instruments and studies, and will likely cause bias in quantifying OCTA metrics 17 . Campbell et al. have introduced terminology for the retinal circulation 114 aligned to current knowledge of retinal anatomy gained from histological studies 28,115,116 . We note that this terminology has been recently further updated by Hormel et al., who have proposed replacing the term radial peripapillary capillary plexus with nerve fiber layer plexus (NFLP) and the term superficial vascular plexus with ganglion cell layer plexus (GCLP) 38 . Since Campbell et al.'s terminology has already been widely applied, we recommend adopting their layer names and boundary definitions (Fig. 3c) 114 , as described below, and reserving consideration of further updates. Such consistency will enable generation of a critical mass of studies of the same microvascular networks for which comparison between studies will be meaningful.
We summarize Campbell et al.'s definitions below (see Figs. 2 and 3c).
Superficial vascular plexus (SVP)-defined as the inner 80% of the ganglion cell complex (GCC) that includes the nerve fiber layer, ganglion cell layer and inner plexiform layer 114 . Characterized by a dense irregular lattice work of vessels composed of larger arteries, arterioles, capillaries, venules, and veins 28,117 .
Superficial capillary plexus (SCP)-that part of the SVP comprising only capillaries; capillaries in this network demonstrate looping.
Intermediate capillary plexus (ICP)-comprises capillaries composed of vertical and oblique segments and located 28 within the outer 20% of the GCC and the inner 50% of the inner nuclear layer (INL).
Deep capillary plexus (DCP)-comprises capillaries arranged in a laminar plane 28 and located within the outer 50% of the inner nuclear layer (INL) and the outer plexiform layer (OPL) (includes Henle's fiber layer; not shown in Fig. 3 but visualized in Fig. 2).
Deep vascular complex (DVC)-is composed of the intermediate capillary plexus (ICP) and the deep capillary plexus (DCP).
Radial peripapillary capillary plexus (RPCP)-is present in the peripapillary retina and characterized by long capillary segments predominantly oriented parallel to the nerve fiber layer axons. RPCP plays an important role in supplying the densely packed nerve fiber layer 28,117 .
Superficial vascular complex (SVC)-groups the RPCP and SVP.

Image artifacts
OCTA suffers from various image artifacts that impact image quality and interpretation and are important to avoid or recognize and mitigate. Artifacts can be caused by the nature of OCTA data collection and processing, eye movement, and properties of the eye and light-matter interaction. There are many reports seeking to introduce standardized terms and definitions of artifacts 86 . Most often, artifacts have been defined and described as: • shadow artifacts caused by OCT signal attenuation during imaging and appearing as areas of decreased flow information in the image and due to, e.g., large vessels, cataracts, vitreous floaters, and pathology within the retina such as drusen 120 ; • field-dependent ocular aberrations that can introduce variable image quality in wide-field OCTA images and artifacts, such as double rendering of vasculature 66 ; and • projection artifacts, also known as decorrelation tails, which are particularly challenging and universal to all OCTA devices, as they unavoidably result from the light-matter interaction with vessels 38,86,121 . They cause the projection of the more superficial vascular network into the deeper networks. Without the removal of projection artifacts, it is difficult to obtain accurate metrics describing the microvascular network below the SVP. There have been various algorithms introduced for removing projection artifacts in research, such as slab subtraction (operating on en face images only) 122,123 and projection-resolved OCTA algorithms (operating on en face and cross-sectional images) 38,124,125 . Commercial instruments now also include proprietary projection artifact-removal algorithms. Projection-resolved OCTA algorithms generally produce better results than slab subtraction methods, however, can also produce images containing residual artifacts 121 . It is important to report in publications which algorithm has been used to mitigate projection artifacts, whether residual artifacts remain and, if so, how they impact on the assessment of the results. We advocate the use of the artifact terminology described here and strongly encourage the incorporation of projection artifact mitigation into both research and clinical OCTA imaging products. Further work is required to compare and select the best of the available algorithms.

En face image averaging
Averaging of multiple en face OCTA images (projections) improves image quality [126][127][128] by reducing noise that could otherwise be misinterpreted as flow, by improving the connectivity of vessels 98,129,130 , and by making the vessels sharper and thinner 127 . Averaging modifies quantitative automated measurements; for example, leading to decreased vessel density 98,127,129,130 . It may also improve microaneurysm detection 131 and diagnostic accuracy of metrics 127 . Overall, averaging enhances the visualization of vessels which makes qualitative and quantitative assessment of OCTA images more reliable 132 .
Uji et al. reported that an average of a minimum of six en face images was needed for ideal binarization of macular OCTA images from the deep capillary plexus and two images were needed from the superficial vessel plexus when using the Zeiss Cirrus 5000 (Carl Zeiss Meditec, Dublin, CA, USA) 98 . Nelson reported that averaging three en face images significantly improved vessel visualization of the radial peripapillary capillary layer when using the Zeiss Cirrus 5000 127  shown that for the Optovue instrument averaging 5-9 en face OCTA images significantly improves SNR and averaging 4-7 OCTA frames improves skeletonization metrics 128 . Furthermore, Nelson et al. has demonstrated that OCTA averaging has improved diagnostic accuracy of vessel area density and significantly improved vessel skeleton density assessment of glaucoma; indicating that the averaging procedure may help better identify clinically valuable microvascular network biomarkers 127 .
Averaging typically requires multiple volumetric scans to be collected which may lengthen total patient time in the clinic, which is an imposition on individuals undergoing OCTA examinations and a clinical overhead. More systematic studies on averaging are required before recommendations on standards can be made. Until then, we recommend not comparing metrics obtained on single images with those from averaged images as it has been demonstrated that they are not equivalent 98,133 .

OCTA image quality indicators and classification
Yu et al. have demonstrated that vessel density measurements correlate with OCTA image quality 134 . It is clear that standards for image quality are needed. Automated image assessment procedures are under development for OCTA image interpretation, seeking to distinguish both good and poor image quality, and to identify the type and severity of disease. Automated OCTA image quality indicators have been introduced by some OCTA manufacturers, e.g., Optovue's scan quality index (SQI). However, consensus on the choice of threshold values for acceptable image quality has not been reached. For example, Optovue has recommended that an OCTA image with an SQI below 6 should be excluded from analysis. Ali et al. have shown that an SQI of 7 or above should be considered for the Optovue OCTA and is insufficient on its own-visual inspection or imagegrading should still be applied 135 . Wang et al. 85 and Mirshahi et al. 136 accept image quality based on SQI above 5 for the Optovue OCTA. Lim et al. have recommended only using OCTA images with "Zeiss signal strength" of at least 9 for the Zeiss Cirrus 5000 with an AngioPlex OCTA (Carl Zeiss Meditec, Dublin, CA, USA) to obtain reliable microvascular metrics, whereas, the manufacturer recommends a threshold of 6 137 .
Lauermann et al. reported a deep learning algorithm to classify images as sufficient or insufficient 138 . Automated methods have also been investigated to identify the type and severity of diseases, including diabetic retinopathy 139,140 , AMD 141 , and glaucoma 142 . Overall, much more work is required to demonstrate how algorithms perform with images from different retinal depths and obtained with different OCTA instruments. Alternatively to image quality assessment, methods that are tolerant to a wide range of OCTA image quality could be developed, such as convolutional neural network-based deep learning models 143,144 . There is, as yet, no consensus on the best methods to either assess OCTA image quality or undertake reliable analysis that is "image-quality independent" and, thus, more work on these topics is needed 111 .

Correction of transverse image magnification error
Nominally, the linear OCTA scanning area (expressed in mm 2 ) is derived from the calculation of the linear distance on the surface of the retina subtended by a fixed angular FOV using a defined ocular axial length (e.g., for the Optovue RTVue XR Avanti system, this length is fixed at 23.95 mm). Thus, a fixed angular FOV of 10 × 10 degrees will correspond to a 3 × 3-mm FOV only for individuals with an axial length 23.95 mm (Fig. 7). To convert retinal images from angular units, an FOV correction for the variation in real axial length from the reference setting is required 145 . The most common method to correct transverse image magnification error is the Littmann-Bennett method that expresses 146 the relationship between the measured image diameter D m and the true diameter D t as D t = p·q·D m , where p is the magnification factor of the imaging system and q that of the eye. The factor q can be found from q = 0.01306·(axial length−1.82). The factor p can be calculated from the Bennett formula setting the axial length so that D t = D m thus p = 1/q. For example, for the RTVue XR Avanti, p = 3.46 since the axial length is 23.95 mm.
It would represent a significant advance if more OCTA instrument manufacturers were to make available the axial length for which their system's scanning protocols have been scaled. Sampson et al. have demonstrated that not compensating for axial length variation in OCTA can result in a relative difference before and after correction of up to 51% in foveal avascular zone (FAZ) area 147 . A similar observation has been made by Linderman et al. 102 . Thus, the impact on vascular parameters of not correcting for axial length is likely to be substantial. However, in a systematic review, Llanas et al. found that only 8% of 509 studies corrected OCTA images for magnification before quantitative analysis 148 .
One factor impeding researchers and clinicians from correcting magnification errors is access to the comparatively expensive biometry device required to measure axial length 149 . Tan et al. have proposed alternatively to correct the FOV by registering the images according to the measured distance between centers of the optic nerve head and fovea, assuming a small individual variation of the distance 150 . However, a study by Rohrschneider has demonstrated considerable variability in the distance between the optic nerve head and the fovea, suggesting this method may not be satisfactory 151 . Recently, Morgan et al. have explored the estimation of axial length from measured refractive error and corneal curvature to support cost-effective myopia management 149 and have shown these factors offer a good estimate of actual axial length. Therefore, the estimated axial length could potentially be used with the Littmann-Bennett formula to correct image magnification error. However, this potential approach to correction has not yet been investigated. Lal et al. recently empirically demonstrated the need to include both axial length and refractive error when correcting for transverse image magnification error 152 . However, such still more demanding requirements, in measuring more parameters related to eye biometry, run counter to what is required: a simple and low-cost method of correction that avoids measuring axial length. Thus, more research is needed to establish a simple, robust, standard method for image size correction because of its large impact on vessel parameters.
Review and standardization of metrics for quantitative analysis of the retinal microvascular network architecture and foveal avascular zone Microvascular structural and functional biomarkers in the retina can be mapped and quantified using specialized metrics (Table 1, Fig. 8) 33,153 . Since many diseases affect multiple markers, it is important in building a complete picture of a condition to characterize retinal vascular networks using multiple metrics 154 , even if ultimately a minimum set of biomarkers is desirable.
Currently, the quantitative analysis of the retinal microvascular network in OCTA is limited to the Axial length equal to defined ocular axial length Axial length shorter than defined ocular axial length Axial length longer than defined ocular axial length a b c n eye D t

Retina Cornea
Axial length FAZ axis ratio 1 Binarized Ratio between the major and minor axis of the ellipse best fit to FAZ outline shape.
utilization of 2D-projection maps. This approach facilitates direct comparison with other 2D techniques, such as FA and ICGA, and compresses projection artifacts discussed under the subsection on Image Artifacts above. The most-used metrics are vessel density, flow area, nonflow area and FAZ area. However, the way in which such metrics are defined and used can vary between instruments and/or studies. Munk et al. have reported a plethora of terminology around reduced flow. Altered retinal and choroidal capillary flow/perfusion in diabetic retinopathy has been described as "flow void", "capillary non-perfusion", or "impaired capillary perfusion" 5 . Other terms used include "greyish area", "no-flow area", "areas with decreased vascular perfusion", "flow deficit", "flow attenuation", "low and no flow", "reduced flow", and "flow abnormalities" 5 . Similarly, Fawzi reported that a whole range of terms have been used to describe the absence of flow, including "nonperfusion", "avascular area", and "absent flow" 17 . There are further inconsistencies among metrics, with definitions of vessel density referring to areas or lengths. The "vessel density" quantified by built-in software of the Zeiss Angioplex OCTA instrument is defined as the total length of skeletonized perfused vasculature per unit area of the measurement region. The Optovue AngioVue OCTA built-in software calculates vessel density as the total area of perfused vessels ("all white pixels") per image area. Durbin et al. defined "perfusion density" as the percentage area occupied by perfused binarized vessels and "vessel density" based on a map with vessels of 1-pixel width 155 .
After consideration of these and other works, we recommend using the metrics as named and defined in Table 1 and visualized in Fig. 8. Our selections appear most commonly in the literature and comprehensively characterize the important changes in the vascular network architecture due to vascular dysfunction 153 . Specifically, we suggest replacing the term vessel density with vessel area density (VAD) to better distinguish between VAD and vessel length density (VLD) ( Table 1).
The FAZ plays an important role in microvascular disease development and should be quantified along with the microvascular network. Most commercial instruments and published papers report FAZ area using the same definition. However, in some studies, FAZ area is analyzed within a single retinal layer; whereas, in others from the whole retina. Since such differences will impact on the analysis outcomes, comparisons between such studies would be invalid. We recommend adding other FAZ metrics, including FAZ perimeter length and acircularity index 156 . The strong correlation between FAZ acircularity and the presence of diabetic retinopathy has been demonstrated by Krawitz et al. 156 . The parameter "foveal vessel density (FD-300)" should be also considered, first introduced by Wang et al. 157 and adopted by Optovue. The parameter is defined as the percentage of the area occupied by vessels (white pixels) divided by the area in a 300-µm width rim surrounding the FAZ and was reported to be more sensitive to FAZ enlargement introduced by vascular pathologies than FAZ area alone 157 . A similar observation was made by Rosen et al., who have performed a similar analysis to Wang et al. but using a 200-µm width rim 158 .
We recommend at this stage in the development of OCTA using the widest applicable range of standardized OCTA metrics. The comprehensive characterization of the retinal microvascular network will provide the strongest prospects for identifying the best biomarkers for early diagnosis and monitoring of diseases. As yet, which metrics have the highest predictive value in clinical practice is unknown 159 ; therefore, this approach will  support the identification of the best metrics and, ultimately, the best diagnostic tools, including those based on artificial inteligence 111 . Further systematic research on biomarker refinement will eventually lead to identifying single or minimum combination biomarkers to precisely identify disease stages and response to treatment. For consistency, the definitions of the retinal layers proposed above should be used along with these metrics   Table 1 to clearly identify which part of the vascular network is being characterized. The retinal microvascular network architecture should be characterized in the whole image as well as in subfields defined by the Early Treatment Diabetic Retinopathy Study (ETDRS) grid commonly applied in ophthalmology. This grid divides the retina into nine regions defined by three rings: a central foveal ring 1 mm in diameter, an inner macular ring 3 mm in diameter, and an outer macular ring 6 mm in diameter. The inner and outer rings are divided into four quadrants: nasal, temporal, superior, and inferior 160 . An advantage of the ETDRS grid is that its layout accounts for natural differences in the shape of the retina 160 ; enables localization of important areas of the macula and identifies their changes due to disease; and is applied by many ophthalmic imaging modalities; thus, allowing direct comparison of localized changes between OCTA and other modalities.

Clinical practice
Standardization is needed across the whole OCTA measurement pipeline, thus, we recommend adoption of a standard protocol for imaging subjects: pupil dilation before OCTA examination (after careful assessment by the treating physician for potential harmful side effects) and lubricating eye drops (for subjects with an unstable tear film); imaging in a dark room with OCTA imaging head adjusted to ensure subject comfort; subject fixation on a central target and requested to refrain, where possible, from blinking; and imaging conducted by welltrained OCTA operators. OCTA images should be reviewed immediately post-measurement and repeated if determined to be inadequate. We recommend the combination of visual inspection (presence of blinking and motion artifacts) and an image quality indicator (scan quality index or equivalent) to confirm poor image quality or otherwise. Subjects attending single clinics should be imaged on the same device with the same protocols to increase the accuracy of comparison between visits 63 . Viewing en face OCTA maps should always be performed in concert with inspection of 3D and single B-scan OCT quality and segmentation results to ensure correct image interpretation 10,38,63 . Ideally, OCTA instrument vendors would harmonize training to follow standardized approaches to data collection, analysis and interpretation.
In the meantime, clinicians and researchers must keep in mind that their data might not be directly comparable: between institutions, instruments, and their own datasets analyzed using different manufacturers' software. Currently reported VAD in the superficial vascular complex can range from 28 to 60% between studies 49 162 . In general, there needs to be more open discussion of differences in quantitative metrics and their origins, and we repeat that more effort should be undertaken to harmonize imaging protocols and data analysis methods to minimize these differences.
When reporting intrasession (imaging on the same day) and intersession (imaging on different days) repeatability, we recommend measuring the coefficient of repeatability, CR, based on the within-subject standard deviation, S w , as introduced by Bland and Altman 88 . Here, the standard deviation of repeated measures for each subject should be calculated and squared to obtain the variance for each subject. S w is the square root of the average variance for all subjects. CR = 2.77·S w and 95% confidence intervals (CI) can be calculated as: CR ± 1:96 Á S w = ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2nðm À 1Þ p À Á , where n is the number of subjects and m the number of measurements for each subject. Furthermore, the inter-ocular symmetry (differences in metrics between the left and right eye) can be quantified by limits of agreement, LA, also introduced by Bland and Altman. Here, the mean, d, and standard deviation, SD, of the differences between both eyes is first calculated to then be able to calculate LA as: d ± 1:96SD. Some authors do not report the within-subject standard deviation necessary for estimating the threshold of test-retest variability 163 . Others define CR as 1.96 times the standard deviation of the differences, instead of 2.77 times the within-subject standard deviation, as recommended by Bland and Altman 164 . Such varied approaches to statistical analysis of results preclude meaningful comparisons between studies 88 .

Recommendations, future perspectives, and challenges
OCTA has greatly advanced in recent years, yet the journey toward clinical utility is just beginning. The prospects for OCTA and its applications are numerous and exciting if we can move beyond fragmented and piecemeal approaches toward large-scale coordinated application of a consensus methodology.
An exciting prospect is the building of reliable, multisourced retinal OCTA databases of healthy subjects and patients with key vascular diseases 165 . This work has commenced with Coscas et al., who first reported a singlecenter normative database of OCTA metrics 166 . Still more comprehensive databases are required not only to further the OCTA-based diagnosis of retinal disease but also support the discovery of better biomarkers with artificial intelligence and related technologies 25,167 . To be able to build reliable large-scale databases 58,168 , we have outlined changes in practice and the further research required. We summarize the most important recommendations of this review below.
1. Consistent anatomy: Adopt definitions and terminology for the retinal vascular layers as summarized in the section "Retinal layer segmentation and definition" above. Such consistency will minimize variations in data analysis and reporting of findings due to variable layer definition and in-depth location of en face OCTA maps and make easier comparison of findings between reports. 2. Report key imaging parameters: Producers of OCTA instruments and researchers should report the transverse resolution and transverse scanning density along with the angular FOV of each protocol.  105 studies. Such data sets should include data collected with OCTA but ideally also with histology; histology currently suffers from less artifacts and enables visualization of many more vessels 68 . We acknowledge that patient privacy concerns might stop researchers from building and sharing large databases between institutions 169 . However, large data sets are critical for the datadriven innovation that ultimately will improve patient care 169,170 . New tools and better regulations/policies need to be developed to better protect privacy and enable data sharing 169 . Possible tools for considerations are describing patterns of groups in the dataset rather than individuals, federated learning 171 -a machine learning paradigm in which algorithms are trained collaboratively without exchange of data-as well as generative adversarial networks 169,172 -a deep learning approach in which a model is trained to generate synthetic image examples that are different from the original images. 9. Create open-source software and standard metrics: Following Steps 1-8 above should lead to a suite of best practices suitable for implementation in opensource software that will also enable quantitative analysis of the retinal vascular network-based on a selection of the metrics defined herein thereby minimizing errors of using and comparing data analyzed by different algorithms 97,99 . In the meantime, OCTA users should focus their efforts on harmonizing imaging protocols, data analysis methods, and reporting of data and in ensuring they publish detailed descriptions of all procedures they apply.
We now move our attention to motivating continued research on promising new ways of quantifying the microvascular network architecture.
Visualization of retinal OCTA data through a 2Dprojection map is convenient as it enables direct comparison with other 2D techniques, such as FA and ICGA, and compresses projection artifacts as discussed. However, 2D-projection of the volumetric vascular network may lead to altered metrics, such as caused by false branch points generated by vessels appearing to overlap in 2D but in reality existing at distinct elevations in 3D, and introducing errors into other metrics such as segment length 153 . 3D vessel orientation has been reported to be important in understanding local tissue oxygenation 153,173 . In principle, full volumetric analysis would seem a laudable objective, but much more work is needed to advance methods and metrics for characterization of the retinal microvascular network in 3D before it could be seen as a viable clinical pathway 174,175 .
Further research on increasing the visibility of retinal vessels in the OCTA image is warranted. Imaging artifacts, image post-processing, and vessel orientation to the scanning beam all impact on the visualization of retinal vasculature 86,176 . As demonstrated by Zhu et al. 176 and further confirmed by Yu et al. 68 , vertically oriented (i.e., aligned with the optical beam) vessels are more difficult to visualize. Possible ways to improve visualization include: imaging after tracer injection, e.g., intralipid 176 ; image averaging 177 ; multi-angle illumination and detection 178 ; and expanding the detectable flow velocity range by imaging at various scanning rates and combining images obtained at different rates 50,[179][180][181] . Varying dynamic and static wavefront aberration may introduce variable image quality in OCTA images as well as artifacts, such as double rendering of vasculature, that could be overcome by applying adaptive optics correction 66 .
A range of other technical improvements merit further research. Increasing the angular FOV to more than 100 degrees, without sacrificing sampling density, would enable monitoring and assessment of vascular conditions at different eccentricities 182 . A recent study has shown that the early stage of diabetic retinopathy occurs in the peripheral retina so can be missed by standard FOVs 24 .
A dual-mode instrument incorporating standard (20 µm) and high (5-10 µm) transverse resolution would allow users to choose between screening and region-ofinterest imaging. A promising pathway to a dual-mode instrument is to combine OCTA with adaptive optics to reduce the impact of the optical aberrations of the human eye 53,[183][184][185] . More generally, higher-scan-speed and higher-resolution OCTA imaging, if artifacts could be overcome, would further improve our capacity for functional study of the retinal microvasculature 49 and further enhance early-stage diagnosis of retinal vascular diseases 186,187 .
Whilst we have promoted a wide range of metrics as potential biomarkers, we should not preclude future development of new metrics. In this vein, Ma et al. have recently proposed the retinal microvascular orientation pattern as a potential biomarker for diabetic retinopathy diagnosis 188 .

Conclusion
Our review highlights the challenges regarding the lack of comparability of OCTA findings between instruments and studies and the importance of drastically improving this to support the diagnosis, monitoring, and treatment of retinal and systemic vascular diseases. The OCTA retinal imaging community is urged to define best practices for data collection, analysis, and reporting of results, and to promote transparent data sharing to promote accuracy, reliability, and wider collaboration. We hope that the topics discussed here, in proposing minimum standards, as well as identifying areas of concern and opportunity for future research, will encourage researchers, clinicians, and systems manufacturers to come together as a community to continue harmonization of the current disparate terminology, methods, and practice and motivate progress in the field.
OCTA is a powerful and exciting technology that has the strong potential to change medical practice in relation to retinal disease. We second Spaide et al. who wrote: "OCTA development can and will be a group effort. Let's get to work" 10 . Indeed, together, we can do it.