Effect of OCTA Image Artifacts on Quantitation of Vessel Density and Perfusion

Abstract

Background and Objective:

To quantify the effects of four common Optical Coherence Tomography Angiography (OCTA) artifacts on metrics of perfusion.

Patients and Methods:

Data from the Framingham Heart Study were reviewed to select consecutive OCTA scans with and without representative image artifacts. Four artifact categories were defined (decentration, motion, shadowing, defocus) and graded using a severity scale of “Not Present,” “Mild,” or “Severe.” Multivariate linear mixed effects models assessed the impact of each artifact on vessel skeleton density (VSD), vessel area density (VAD), and flux.

Results:

141 OCTA scans from 20 participants were included. Severe shadowing or any severity of defocus significantly attenuated all OCTA parameters. Severe decentration significantly attenuated VAD and flux, but not VSD. Motion artifacts and mild decentration had no significant effects.

Conclusion:

Image defocus and shadowing significantly affect all OCTA measures. VSD is the most robust measure against most artifacts.

Introduction

Optical Coherence Tomography Angiography (OCTA) is a noninvasive imaging modality approved by the United States Food and Drug Administration that captures three-dimensional, in vivo blood flow at high spatial and temporal resolutions.^1–3 Clinical OCTA applications have been implemented for age-related macular degeneration (AMD), retinal occlusions, diabetes, and hypertension.^1,3–6 As OCTA use expands, the accurate and reliable interpretation of data is of increasing importance.

One significant challenge for OCTA is the presence of image artifacts that can misrepresent retinal morphology, impair scan quality, and ultimately impede analysis.^7–10 Previous work has characterized many artifact types including projection artifacts, segmentation errors, eye movement (motion artifact), signal attenuation through light blockage (shadowing), and improper imaging technique (decentration, defocus).^7–18 The prevalence of such artifacts can be as high as 97% of OCTA scans.^13–21 Early development of projection artifact removal software and improved manual and automatic layer segmentation has minimized the effects of these two artifacts.^{3,9,13,15,22–25} However, it is unclear how decentration, motion, shadowing, and defocus impact OCTA quantitation.

Inconsistencies in defining and assessing these four artifacts have complicated the interpretation of existing literature. For instance, prevalence studies have opted for different operational definitions for motion artifacts or do not clearly distinguish shadowing from defocus or “tilt.”^7,18,21,26 Established artifact severity grading schemes also either involve the evaluation of all B-scans or are specific per artifact type.^13,17,27 Neither method is practical for large datasets. Lastly, studies have reported lower vessel density (VD) due to media opacities and induced defocus.^{12,20,28–30} Therefore, it is necessary to characterize the effects on widely used OCTA parameters including vessel skeleton density (VSD), vessel area density (VAD), and flux.^31,32

We previously reported a systematic image grading scheme and quantified the impact of projection artifacts in the Framingham Heart Study (FHS), a community-based longitudinal study of nearly 1000 subjects with a low prevalence of retinal pathology.³³ This current paper implements a generalized severity grading scheme to evaluate the effect of other prevalent image artifacts (decentration, motion, shadowing, and defocus) on OCTA quantification among FHS participants without visually significant ocular disease. These methods and findings are highly relevant for studies using OCTA-derived metrics.

Patients and Methods

Participant Recruitment

This is a nested study within the FHS involving participants from the FHS Offspring Generation and OMNI-1 cohorts who were evaluated between November 2020 and August 2022. Adult human participants were recruited through a study protocol approved by the Institutional Review Board of Boston University Medical Center and in compliance with the Declaration of Helsinki. All participants provided written informed consent.

Data Collection

Four foveal-centered 3×3 mm OCTA scans were acquired per undilated eye of each participant³³ using a commercially available spectral-domain Cirrus Angioplex 6000 (Carl Zeiss Meditec, Inc). Axial length (AL) and keratometry data were acquired using IOL Master 500 (Carl Zeiss Meditec, Inc). Exact imaging protocols, image rating methods, and whole-cohort demographics have been described.³³ In previous publications, a systematic image grading paradigm was implemented to rate scans as “Excellent,” “Good,” or “Suboptimal.”³³ Among each set of 3×3 mm OCTA quadruplicates, one previously determined “Excellent” scan was used as the reference ground truth for this current manuscript and all other scans were termed as “Duplicates.” Twenty consecutive participants with one Excellent scan containing no or minimal image artifacts and at least one Duplicate containing any image artifact in both eyes were chosen for analysis.

OCTA Image Artifact Grading Scheme

OCTA scans were graded for decentration, motion artifact, shadowing, and defocus using a grading scheme adapted from previous methods.²⁸ Grades were assigned by deliberation and agreement between two trained graders (AT and ACM) with discrepancies arbitrated by a final third trained grader (AHK). Each category received a separate grade of “Not Present,” “Mild,” or “Severe” in order of progressive severity (Table 1). Figure 1 illustrates exemplary images.

Table 1.

OCTA Image Artifacts Grading Scheme

Artifact	Definition	Severity Grade
		Not Present	Mild	Severe
Decentration	Magnitude of image offset from FAZ center	Image is centered on FAZ	Image is decentered by ≤1/2 diameter of FAZ	Image is decentered by >1/2 diameter of FAZ
Motion Artifact	Artifacts due to eye motion which appear as displaced vessels by >1 vessel width and/or motion lines	No motion artifacts	1–3 motion artifacts present	≥4 motion artifacts present
Shadowing	Areas of reduced signal on structural OCT due to media opacities, refractive shift, or z-axis motion	No shadowing	≤1/3 of image area affected by shadowing	>1/3 of image area affected by shadowing
Defocus	Decreased vessel definition due to poor machine focus	Well-defined vessels throughout image	Reduced capillary definition in ≤1/3 of image area	Reduced capillary definition in >1/3 of image area

OCTA = Optical Coherence Tomography Angiography, FAZ = Foveal Avascular Zone, OCT = Optical Coherence Tomography.

Figure 1: Exemplary OCTA Image Artifacts by Severity Grade.

All images are 3×3 mm en face OCTA scans of the superficial retinal layer (defined as from the inner limiting membrane to the boundary between the inner plexiform and inner nuclear layers. Full-thickness slabs demonstrate the same artifacts and can be used interchangeably). While multiple artifacts may be present, only one artifact is evaluated per image set. (A-C) Examples of decentration. The “slice navigator” magenta and teal lines from the commercial reader software intersect at the image center and facilitate assessment of decentration. (D-F) Examples of motion artifacts. Arrows and brackets denote single motion lines or large sections of motion artifacts, respectively. (G-I) Examples of shadowing. The corresponding structural images are displayed to the upper left of each scan. Red ovals outline areas of shadowing. (J-L) Examples of defocus. Red rectangles indicate broad regions of defocus where capillary definition is lost or substantially worse than other parts of the images.

Grading was performed by visual inspection without specialized software or laborious manual measurements. Therefore, these methods are practical for large datasets. Decentration was semi-quantitatively determined as displacement between the center of the foveal avascular zone (FAZ) and the center of the image window (Figure 2). FAZ diameter was defined as the longest linear diameter of the FAZ upon visual inspection. The image center was defined by the default intersection of the slice navigator lines on commercially available OCTA viewer software.

Figure 2: Evaluation of Decentration.

All images are en face OCTA scans. The image center is landmarked by the intersection of the “slice navigator” magenta and teal lines from commercial reader software. Decentration (dashed line) was defined as the offset from the image center to the foveal avascular zone (FAZ) center. The longest FAZ diameter for each scan is measured with double-headed arrows. The equivalent lengths of decentration and FAZ diameter (bisected) are displayed below each image for comparison. (A) A sample scan with a shorter FAZ diameter graded as “Severe” for decentration >1/2 FAZ diameter. (B) A sample scan with a longer FAZ diameter graded as “Mild” for decentration <1/2 FAZ diameter.

Motion artifacts on OCTA scans were defined as lateral displacement of non-capillary vessels (arterioles or venules) by >1 vessel width and/or horizontal bar-coding patterns (white, gray, or black; Figure 1).⁷ Multiple vessel breaks along one B-scan, indicating a single eye shift, were counted as one artifact. To differentiate shadowing from impaired vascular perfusion, we cross-referenced focal areas of OCTA signal loss to the structural Optical Coherence Tomography (OCT) scan (Figure 1). Shadowing was defined as corresponding areas of reduced signal intensity on both scans due to media opacities or refractive shifts (banding).¹⁷ Defocus was defined as decreased vessel definition in the absence of reduced signal intensity on the structural scan that was likely due to suboptimal machine focus or z-axis motion.

OCTA Image Analysis

OCTA quantification was conducted using a custom semiautomated and previously validated program (MATLAB R2019a; The MathWorks, Inc) to output VSD, VAD, and flux (Figure 3). VSD was calculated as the proportion of skeletonized vessel length to total image area.^31,32 VAD was calculated as the proportion of total vessel area to total image area.³¹ Flux was computed by averaging the nonbinarized pixel intensities within and including blood vessel walls.³² All OCTA metrics were derived from en face scans of the superficial retinal layer (SRL). The SRL was defined as extending from the internal limiting membrane to the boundary between the inner plexiform and inner nuclear layers. The FAZ area was included in quantitation and the large vessel areas were removed from analysis during MATLAB post-processing.³³ SRL VSD, SRL VAD, and SRL flux were the primary outcomes of this study.

Figure 3: Representative OCTA Quantification Parameters.

(A) An en face image before MATLAB processing. The corresponding structural image is displayed to the upper right inset. (B-D) Pseudocolored vessel skeleton density, vessel area density, and flux maps. Hot colors indicate higher values and cool colors indicate lower values along each respective scale bar. The FAZ was included and the large vessels were removed from quantification as described in the methods.

Statistical Analysis

Statistical analyses were performed using R version 4.4.1 (R Foundation for Statistical Computing). Summary statistics of means with standard deviations (SD) and range were used to summarize participant characteristics. Multivariate linear mixed effects models were used to evaluate differences in VSD, VAD, and flux between severity grades. Models were adjusted for age, sex, and multiple images per participant. Additional analyses included adjustment for AL as a model covariate and by using the adapted Littman Formula.³⁴ The classification of “Not Present” was set as the reference category for all image artifacts. Statistical significance was defined as p<0.013 when adjusted for multiple comparisons using the Bonferroni correction.

Results

Twenty consecutive FHS participants (40 eyes and 141 OCTA scans) meeting the inclusion criteria were included in this study (Table 2). Participants were all non-Hispanic Whites and were 73±5 years old. Table 3 summarizes the results of OCTA image artifact grading. When stratified by artifact type and severity, mild motion artifact had the greatest representation in this dataset (48%) while severe decentration and severe defocus tied for the least (6%). Among artifact category, motion artifact was the most common overall (59%) while defocus was the least (13%). Supplementary Table 1 tabulates the coincidence of artifacts in this dataset. Of all scans, 25 (18%) had no artifacts, 44 (31%) had one artifact, and 72 (51%) had two or more artifacts.

Table 2.

Participant Demographics

Characteristic	Value
No. of participants included	20
No. of eyes included	40
Mean age, years (SD) [range]	73 (5) [62–81]
Female, n (%)	9 (45%)
Non-Hispanic White	20 (100%)
Mean axial length, mm (SD) [range]
Right eyes	23.74 (0.84) [22.69–26.02]
Left eyes	23.63 (1.06) [22.63–26.57]
No. of OCTA scans included	141
Excellent	40
Duplicates	101
No. of subjects with retinal findings	10
No. of eyes with retinal findings on OCTA	13
Epiretinal membrane	4
Retinal distortion/disruption	3
Vitreomacular adhesion	3
Other*	5

OCTA = Optical Coherence Tomography Angiography, SD = standard deviation.

^*Other: posterior vitreous detachment, tortuous vessels, abnormal fovea, drusen. Eyes may contain more than one retinal finding.

Table 3.

Results of OCTA Image Artifact Grading

	Image Artifact Severity Grade
Category, n (%)	Not Present	Mild	Severe
Decentration	78 (55)	54 (38)	9 (6)
Motion	58 (41)	67 (48)	16 (11)
Shadowing	97 (69)	29 (21)	15 (11)
Defocus	122 (87)	10 (7)	9 (6)

OCTA = Optical Coherence Tomography Angiography.

Counts and percentages were calculated out of a total of 141 OCTA images.

Table 4 summarizes the main results of the linear mixed effects models. Two categories of artifacts had strikingly consistent findings. First, neither mild nor severe motion artifacts significantly affected any outcome despite being the most common artifact in this dataset. In contrast, defocus of any degree attenuated all parameters (Mild defocus: VSD: β=−0.0048± 0.0016, p=0.004; VAD: β=−0.0113± 0.0045, p=0.013; flux: β=−0.0119± 0.0032, p<0.001. Severe defocus: VSD: β=−0.0076±0.0017, p<0.001; VAD: β=−0.0176± 0.0048, p<0.001; flux: β=−0.0176± 0.0035, p<0.001).

Table 4.

Multivariate Linear Mixed Models for Vessel Skeleton Density, Vessel Area Density, and Flux

Artifact Type	Vessel Skeleton Density			Vessel Area Density			Flux
Decentration	Beta	SE	p-value	Beta	SE	p-value	Beta	SE	p-value
Mild	0.0001	0.0008	0.944	−0.0005	0.0024	0.842	−0.0008	0.0017	0.649
Severe	−0.0020	0.0018	0.256	−0.0138	0.0050	0.006	−0.0101	0.0035	0.005
Motion
Mild	−0.0004	0.0009	0.693	−0.0018	0.0026	0.488	−0.0010	0.0019	0.587
Severe	−0.0010	0.0016	0.515	−0.0041	0.0044	0.346	−0.0022	0.0031	0.475
Shadowing
Mild	−0.0016	0.0010	0.129	−0.0056	0.0029	0.057	−0.0057	0.0021	0.007
Severe	−0.0069	0.0014	<0.001	−0.0245	0.0039	<0.001	−0.0212	0.0028	<0.001
Defocus
Mild	−0.0048	0.0016	0.004	−0.0113	0.0045	0.013	−0.0119	0.0032	<0.001
Severe	−0.0076	0.0017	<0.001	−0.0176	0.0048	<0.001	−0.0176	0.0035	<0.001

All linear mixed models are adjusted for age, sex, and random effects of multiple images per participant. The “Not Present” category was used as the reference for all comparisons. N = 141 scans (20 participants). Significance by p<0.013.

Among decentration and shadowing, only severe gradings had a substantial impact on OCTA outcomes. Mild decentration had no significant effects while severe decentration impaired VAD (β=−0.0138±0.0050, p=0.006) and flux (β=−0.0101±0.0035, p=0.005). Similarly, mild shadowing was associated with a significant decrease in flux (β=−0.0057±0.0021, p=0.007) while severe shadowing attenuated all OCTA parameters (VSD: β=−0.0069± 0.0014, p<0.001; VAD: β=−0.0245±0.0039, p<0.001; flux: β=−0.0212±0.0028, p<0.001).

Table 5 describes the results after AL correction as a covariate in the model or by applying the adapted Littman formula,³⁴ respectively. After adjusting for multiple comparisons, these corrections did not alter the main results described above.

Table 5.

Multivariate Linear Mixed Models Adjusted for Axial Length

Axial Length Adjusted as Covariate in Multivariate Model
Parameter	Vessel Skeleton Density			Vessel Area Density			Flux
Decentration	Beta	SE	p-value	Beta	SE	p-value	Beta	SE	p-value
Mild	0.0001	0.0009	0.903	−0.0002	0.0024	0.937	−0.0004	0.0017	0.819
Severe	−0.0021	0.0018	0.242	−0.0142	0.0049	0.005	−0.0103	0.0035	0.004
Motion
Mild	−0.0004	0.0009	0.681	−0.0018	0.0026	0.472	−0.0009	0.0018	0.637
Severe	−0.0011	0.0016	0.503	−0.0043	0.0043	0.328	−0.0021	0.0031	0.488
Shadowing
Mild	−0.0016	0.0010	0.123	−0.0057	0.0029	0.048	−0.0059	0.0021	0.005
Severe	−0.0068	0.0014	<0.001	−0.0243	0.0039	<0.001	−0.0207	0.0028	<0.001
Defocus
Mild	−0.0048	0.0016	0.004	−0.0109	0.0045	0.016	−0.0115	0.0032	<0.001
Severe	−0.0077	0.0017	<0.001	−0.0177	0.0048	<0.001	−0.0177	0.0034	<0.001
Mean AL	−0.0005	0.0010	0.616	−0.0030	0.0026	0.259	−0.0027	0.0014	0.053
Axial Length Adjusted using Littman Formula in Multivariate Model
Parameter	Vessel Skeleton Density			Vessel Area Density			Flux
Decentration	Beta	SE	p-value	Beta	SE	p-value	Beta	SE	p-value
Mild	0.0006	0.0009	0.529	0.0009	0.0024	0.701	−0.0003	0.0016	0.868
Severe	−0.0027	0.0018	0.131	−0.0162	0.0049	0.001	−0.0106	0.0034	0.021
Motion
Mild	−0.0006	0.0009	0.497	−0.0026	0.0025	0.307	−0.0010	0.0018	0.570
Severe	−0.0014	0.0016	0.389	−0.0052	0.0044	0.238	−0.0020	0.0030	0.495
Shadowing
Mild	−0.0019	0.0010	0.071	−0.0064	0.0028	0.028	−0.0058	0.0020	0.004
Severe	−0.0069	0.0014	<0.001	−0.0240	0.0039	<0.001	−0.0200	0.0027	<0.001
Defocus
Mild	−0.0042	0.0016	0.012	−0.0091	0.0044	0.044	−0.0104	0.0031	<0.001
Severe	−0.0077	0.0018	<0.001	−0.0178	0.0048	<0.001	−0.0172	0.0033	<0.001

AL = Axial length, SE = standard error.

Linear mixed model is adjusted for age, sex, and random effects of multiple images per subject. The “Not Present” category was used as the reference for all comparisons. N=141 scans (20 subjects). Significance by p<0.013.

Discussion

Image artifacts are an inherent challenge in OCTA imaging and can confound the interpretation of data.^7,9 While some artifacts can be mitigated with additional scans or improved imaging technique, it is in many cases impossible to capture “perfect” scans. Thus, well-defined methods to assess artifact severity and a clear understanding of how they interfere with OCTA metrics are essential. In this paper, we used a grading scheme to evaluate four common artifacts (decentration, motion, shadowing, and defocus) that were observed in data collected from a large epidemiologic study under controlled conditions and well-developed protocols. Notably, OCTA data were acquired by trained technicians who were not certified photographers or ophthalmologists. Therefore, our methods and findings are likely applicable broadly and to settings without significant ophthalmologic infrastructure. We report that two common artifacts, motion or mild image decentration (within half a FAZ diameter), demonstrated no statistically significant effect on OCTA quantification. In contrast, defocus, shadowing, and severe image decentration significantly attenuated at least one OCTA parameter.

As one the most common image artifacts,¹⁷ motion can be difficult to eliminate due to eye blinks, microsaccades, or body movement.^7,9,18 Many commercial vendors, including the one used in this study, employ automatic but imperfect eye tracking to minimize motion artifacts.^35,36 Our results demonstrate that motion artifacts within the range of severity studied here will rarely cause significant changes in vessel density and perfusion measures. Therefore, clinicians and researchers can more confidently use OCTA data with motion artifacts as defined in this study for quantitation.

Decentration can often be adjusted with subject instruction or by directing fixation of the fellow eye with an external light.¹⁷ We semi-quantitatively assessed decentration by normalizing to the FAZ size (Figure 2), which is more practical than exact or laborious manual measurements. When the fovea is decentered, extrafoveal locations with dissimilar vascular morphology may be captured in the en face OCTA, making inter-scan comparisons unreliable. Decentration has been previously shown to confound retinal thickness measurements in both diabetic and healthy eyes.^37,38 In the present study, severe decentration demonstrated a greater effect on VAD and flux than VSD. In contrast, mild decentration appears acceptable for OCTA quantification as it demonstrated no statistically significant impact on any outcome.

Shadowing, if resulting from biological sources such as cataracts or floaters, may also be inevitable in OCTA scans. Previous studies have demonstrated that media opacities can reduce vessel density in a machine-specific manner.^20,29,30,39 Our results were consistent with these findings and emphasize the highly significant effect of shadowing on other OCTA parameters independent of defocus and non-perfusion (Table 1, Supplementary Table 1). Thus, OCTA shadowing due to corneal, lens, or vitreous opacities could potentially mimic retinal pathology and needs to be considered in future OCTA studies, especially among elderly cohorts with high prevalence of cataracts.⁴⁰

Defocus stems primarily from suboptimal operation of the OCT machine or from cases of unusual refractive error.¹⁵ Although defocus was less common overall in prevalence studies, it was more common among severe artifacts^17,21 and was negatively correlated with VD.^12,29 Previously, it has been suggested that this association may be driven by signal strength, and that defocus artifactually increases OCTA parameters due to the broadening of vessels.²⁹ In our dataset where all defocused images had the highest intensity signal, we demonstrate that any defocus was sufficient to impair all OCTA outcomes independently of signal strength. This discrepancy may be due to other papers using different OCT devices or including extreme defocus²⁹ that did not occur organically during our imaging. Signal strength detection may also be less sensitive to regional defocusing where there is otherwise no light attenuation. Overall, our highly significant associations reveal that defocus should be a principal consideration for quality control and underscore the importance of technician training and subject compliance to minimize artifacts during scan acquisition.

VSD, VAD, and flux represent properties of vessel morphology and capillary perfusion of the retinal vasculature. Although VSD and VAD both measure the density of blood vessels, VAD calculates total pixel area of the vascular network whereas VSD incorporates a binarization process to reduce vessel diameter to the width of one image pixel.^31,32 Previous work has explored comparisons between VSD and VAD.^{31,32,41–43} We show that VSD was less sensitive than VAD against severe decentration and may therefore be a more reliable measure under most conditions. Flux, a non-binarized measure thought to be correlated with red blood cell content, appeared to be the most susceptible OCTA parameter and was attenuated by almost all artifacts. Therefore, the most stringent quality control may be needed for flux.

Earlier papers have also shown that OCTA scans capture a larger transverse area in eyes with longer AL due to two-dimensional projection of a curved retina,^42,44,45 which can significantly alter density measurements for high myopes.^45–47 Therefore, we explored two methods of AL correction: statistical adjustment of AL as a covariate and through the adapted Littmann formula.³⁴ Our main results remained the same using either method (Table 5), suggesting that AL correction does not alter the effects of artifacts on OCTA measures in relatively normal eyes. Nevertheless, AL corrections are essential for the accurate quantification of OCTA parameters across the full range of ocular biometry.

We acknowledge several limitations of this study. First, there exists inherent but unavoidable subjectivity in OCTA grading. To reduce potential bias, we used clearly defined criteria and a well-characterized dataset from a large epidemiologic study.³³ Additionally, we exclusively evaluated 3×3 mm scans in relatively healthy, non-Hispanic White adult eyes using one commercially available device. There are differences in vascular metrics and artifact prevalence between OCT machines due to dissimilar scan rates, flow signal calculation, and artifact correction algorithms.^{1,9,17,27,29,30,36,48} Larger scan sizes also exhibit differences in artifact prevalence and scan resolution.^17,49 Thus, we cannot comment on other OCT devices, scan sizes, or populations with different anatomy and behavioral patterns such as in pediatric eyes.⁵⁰ However, our methods can be adapted and our findings likely apply broadly, perhaps at varied effect magnitudes. Lastly, while we observed a surprising lack of significance for motion artifacts towards OCTA parameters, extreme levels of any artifact, including motion, will undoubtedly influence OCTA outcomes.

In conclusion, the current study substantiates the need for image artifact assessment as part of OCTA image quality control, as emphasized in recent consensus papers.¹⁰ We demonstrate a significant effect of decentration, shadowing, and defocus but not motion artifacts on three OCTA metrics. We also introduce a practical and effective framework to evaluate artifact severity in OCTA scans. More research is needed to refine the proposed grading scheme for different OCT devices, scan sizes, potentially AL correction, and automated artifact scoring through machine learning. As OCTA becomes more integrated with clinical practice, precision and accuracy in interpreting OCTA data despite confounding artifacts will be paramount.