Microvascular damage assessed by optical coherence tomography angiography for glaucoma diagnosis: a systematic review of the most discriminative regions

Abstract A growing number of studies have reported a link between vascular damage and glaucoma based on optical coherence tomography angiography (OCTA) imaging. This multitude of studies focused on different regions of interest (ROIs) which offers the possibility to draw conclusions on the most discriminative locations to diagnose glaucoma. The objective of this work was to review and analyse the discriminative capacity of vascular density, retrieved from different ROIs, on differentiating healthy subjects from glaucoma patients. PubMed was used to perform a systematic review on the analysis of glaucomatous vascular damage using OCTA. All studies up to 21 April 2019 were considered. The ROIs were analysed by region (macula, optic disc and peripapillary region), layer (superficial and deep capillary plexus, avascular, whole retina, choriocapillaris and choroid) and sector (according to the Garway–Heath map). The area under receiver operator characteristic curve (AUROC) and the statistical difference (p‐value) were used to report the importance of each ROI for diagnosing glaucoma. From 96 screened studies, 43 were eligible for this review. Overall, the peripapillary region showed to be the most discriminative region with the highest mean AUROC (0.80 ± 0.09). An improvement of the AUROC from this region is observed when a sectorial analysis is performed, with the highest AUROCs obtained at the inferior and superior sectors of the superficial capillary plexus in the peripapillary region (0.86 ± 0.03 and 0.87 ± 0.10, respectively). The presented work shows that glaucomatous vascular damage can be assessed using OCTA, and its added value as a complementary feature for glaucoma diagnosis depends on the region of interest. A sectorial analysis of the superficial layer at the peripapillary region is preferable for assessing glaucomatous vascular damage.

As in many medical imaging technologies at their early development stage, a number of approaches for estimating the microvascular density based on different regions of interest (ROIs) have been proposed. However, data reported in these approaches are often conflicting and/or arising from smallscale studies, hindering the development of a general methodology to study glaucomatous vascular damage. Microvascular density measured from OCTA has shown to be device-dependent, artefact-dependent (e.g. eye motion, vitreous floaters, and media opacities) (Spaide et al. 2016; Sánchez Brea et al. 2019) and, more importantly, dependent on the imaged ROI. Since OCTA imaging is restricted to a narrow field of view, and the acquisition of a single image with good quality (i.e. no movement artefacts and good contrast) often requires a long exposure time (in patients known to have poor ocular surface and sometimes poor fixation capacities), it is important to ensure an efficient image acquisition, focusing first in the ROIs that yield more relevant information. Moreover, the distribution of the vascular glaucomatous damage among retinal, choriocapillaris and choroid layers is still under research. It is not clear yet whether the significant changes observed at the choriocapillaris and choroid are due to imaging artefacts or due to an actual disease mechanism (Sousa et al. 2019).
The aim of this systematic review was to contribute to the understanding of the role of vascular damage in glaucoma. To that end, the review focuses on the vascular density retrieved from the different ROIs that have been studied so far in the literature, reporting which ROIs have been found to be the most promising for studying glaucoma.

Methods
This research adhered to the Preferred Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines.

Study selection
A literature search was carried out in the PubMed database. The search query can be found in Appendix A. All studies that were published from the 1 January 2014 to the 21 April 2019 were included. The inclusion criteria for each study were as follows: (i) primary study, (ii) mention how the vessel density was computed, (iii) English language, (iv) conducted in humans, (v) investigate glaucomatous eyes in comparison with a healthy control group and (vi) reports at least: the area under the receiver operating characteristic curve (AUROC) or the statistical difference between the control and glaucoma groups. Four authors (A.B., N.B., V.E. and E.F.) screened all the titles and abstracts independently. A full-text screening was carried out by two authors (N.B. and E.F.) independently. In case of disagreement, a third author (A.B. or V.E.) was consulted to reach consensus.

Data collection
The extracted data included the following: study characteristics, AUROC values for different ROIs, microvascular density mean and standard deviation, and p-values from the statistical comparison between healthy and glaucoma groups. If no statistical comparison or p-values were provided, only the AUROC values were collected and vice versa.
For every study, the following characteristics were extracted: sample size including number of patients and eyes for each group, average age in years, and statistical difference (p-value) between groups, glaucoma severity, OCT device brand and respective light-source wavelength, cut-off value for the signal strength index (SSI) (or similar image quality measure) used to exclude patients/eyes and field of view of the OCTA image.
Data collection included the different layers: retina (including superficial, deep and avascular), choriocapillaris and choroid (Fig. 1A); the different regions: macula, optic disc (OD) and peripapillary or circumpapillary (when a circular band around the optic disc was considered instead of whole image) (  (Fig. 1C). The 'whole retina' was treated as a layer (Fig. 1A). Moreover, the 'whole region' (all the sectors combined) and the 'fovea' (centre of the macular region) were considered as sectors for the purpose of this review.

Data analysis
The collected data were used to find which ROIs (region, layer and sector) have been studied and their discriminating power between glaucoma and healthy controls. Since the data analysis was oriented towards the clinical interpretation, the microvascular density was treated as a generic feature, not taking into account the different mathematical approaches used to estimate it.
The AUROC was considered as the most relevant metric for evaluating which ROIs are the most promising for studying glaucoma, since it provides the performance measurement for classification problem at various thresholds settings. The AUROC of all ROIs included in all reviewed studies was averaged to determine a threshold for selecting the studies that would undergo a qualitative assessment (described in section Qualitative assessment). The statistically significant differences (given as p-values) were used to complement the information provided by the AUROCs and assess whether a ROI was relevant for differentiating glaucoma from healthy controls.
For those studies that did not report an AUROC, the decision to perform a qualitative assessment was based on the statistical comparison between the glaucoma and the healthy group. Hence, the ROIs that presented significant statistical differences (pvalue < 0.05) were also qualitatively evaluated.

Qualitative assessment
There are several characteristics in a study that have been reported in the literature as potentially impacting the outcome of OCTA-based glaucoma assessment and thus introducing bias. Thus, despite the high AUROC a method may have, it does not dismiss a careful qualitative analysis to identify these potential sources of bias. Hence, in this review, all the studies that reported an AUROC value above the threshold (mean AUROC values for all studied ROIs), and the ability to significantly differentiate glaucoma from healthy subjects, were qualitatively assessed. The qualitative assessment    to measure the risk of bias was performed independently by two authors (A.B. and E.F.). Criteria were composed in cooperation with experienced ophthalmologists (J.B.B. and I.S.; Appendix B). The following six aspects, ordered by relevance, were considered: 1 Age. The age should not differ significantly between the glaucoma and the healthy groups. If there is an age difference between groups, an adjustment should be executed.

Study selection
Ninety-six studies were identified using the search query in Appendix A. From those, 53 studies were considered eligible after screening the titles and abstracts.
Full-text screening resulted in 43 studies that met all inclusion criteria and, hence, were eligible for the data analysis (Fig. 2). All the included studies provided a statistical analysis of the quantitative vascular evaluation for different ROIs. Twenty-four studies provided AUROC as an outcome. The complete table with the characteristics of the reviewed studies can be found in Appendix B.

AUROC analysis
The AUROCs presented in the reviewed studies are summarized in Table 1 organized per layer, region and sector. All studies calculated AUROC values based on the microvascular density, despite using different image processing techniques for intensity quantification or binarization. Although the macular region showed the highest AUROC values (considering all studies individually), when taking the mean of all ROIs, the peripapillary region had the highest AUROC of 0.80 ± 0.09, whereas the macula and the optic disc both had AUROC of 0.74 ± 0.12. The mean AUROC values for all studied ROIs are shown in Fig. 3. The average of all AUROC values in Table 1 is 0.77, which was set as the threshold for deciding whether a study or ROI should be further analysed in the qualitative assessment. All three regions (optic disc, macular and peripapillary) yielded values above this threshold, as shown in Table 1.

p-value analysis
The results for the vascular density differed greatly between and within ROIs, as shown in Appendix C. Nevertheless, a statistically significant difference between control and glaucoma groups was observed for all the analysed ROIs. The number of statistically significant differences is summarized in Fig. 4 (and detailed in Table C1 in Appendix C).

Macular region
The whole image of the macula in the superficial layer included 15 out of 17 significant values. Five out of six values reported for the whole image in the deep layer were significant. Only one value, however significant, was reported for the choriocapillaris and none for the choroid.

Optic disc
The inside disc sector in the superficial layer included nine significant values and only one non-significant. The inferior segment in the superficial layer included only three values; however, all of them are significant. Only one out of 17 values reported a non-significant difference for the whole image of the optic disc in the superficial layer. No values were reported for the whole image in the whole retina.

Peripapillary and circumpapillary region
The whole image in the superficial layer included 17 out of 19 significant values. The superior and inferior sectors of the peripapillary region in the superficial layer were not represented as much in the literature. However, all the studies that analysed these regions reported a significant difference between the groups (four and three values, respectively). Seven out of eight values for the temporal inferior sector in the superficial layer were significant. No values were reported for the whole region in the choriocapillaris. Five out of five values were reported as significant for the whole region of the circumpapillary ROI in the superficial layer. One value was reported for the temporal superior, temporal inferior and the nasal inferior sectors in the superficial layer, and all three of them were significant.

Qualitative assessment
The bold font in Table 1 highlights the studies that provided one (or more) AUROC values above the threshold. The complete qualitative assessment was performed in the 22 studies that met the requirement of having an AUROC> 0.77 (Appendix D). From these study characteristics, it was possible to draw the following observations:

Discussion
This systematic review gives an insight into which ROIs have been studied so far in literature and which ones seem to contribute the most to an accurate diagnosis of glaucoma using microvascular density computed from OCTA. The ROIs in OCTA imaging were defined by three arguments: region of acquisition, layer and sector. The region of acquisition (macula, optic disc or the peripapillary region) should be the first argument to be considered in OCTA imaging, since it is related to the ability to detect glaucomatous vascular damage. Although the highest AUROCs (considering all studies individually) were observed at the macula, the peripapillary region showed the highest AUROCs when averaging all values per region of acquisition. As mean AUROC is a more reliable indicator than its maximum, we may conclude that the peripapillary region is the most relevant for studying glaucomatous vascular damage.
The second argument to be considered is the layer. Overall, the highest AUROCs were obtained for the superficial layer. Nonetheless, the deeper layers presented in some cases similar classification values to the superficial layer. However, the limited number of studies that have covered these deeper layers does not allow to draw conclusions on their added value for the diagnosis. These layers have been avoided due to the difficulty to explain the physical meaning of the imaged content. As light travels deeper through retinal tissue, it becomes more susceptible to refraction and diffraction. Moreover, given the heterogeneity of retinal tissue, light reflection and absorption occur at different levels depending on the region of acquisition and respective refraction index. As a consequence, shadows are projected to deeper layers, creating what is known as projection artefacts. Therefore, and despite the significant differences observed at the choriocapillaris and the choroid, it is difficult to conclude whether these differences arise from the pathology itself or are a consequence of imaging artefacts. Further research needs to be done in order to understand to what extent the information imaged by OCTA at deeper layers is reliable.
The last argument, and the smallest area, is the sector. A sectorial analysis is not always performed in glaucomatous vascular studies. A number of studies have opted for analysing the retinal layers, mainly the superficial vascular plexus, without any sector discrimination. However, for those that performed sectorial analysis, it was shown that microvascular density is affected differently depending on the sector. Taking the most studied region of acquisition and layer as reference (the superficial layer of the peripapillary region), it can be concluded from this review that the inferior sector (AUROC = 0.86 ± 0.03) and the superior sector (AUROC = 0.87 ± 0.10) are the most promising at discriminating glaucoma. Moreover, Fig. 3 shows that a sectorial circumpapillary analysis (with a fixed distance from the optically hollow) seems to provide a better discrimination than a sectorial peripapillary configuration (which takes into account the entire scan). Such a difference may be explained by the reduced variability present in the circumpapillary region, a specific circular ROI with fixed dimensions around the optic disc.
Overall, looking at the number of studies that used OCTA information to infer glaucomatous vascular damage and the respective AUROCs, it can be concluded that the whole region at the superficial layer of the peripapillary ROI is the most accurate measurement for glaucoma assessment, which could be even further improved by a sectorial circumpapillary analysis. This result was somehow expected, since glaucoma is characterized by a loss of optic nerve axons, which traverse the retina superficially in an anatomical area included in the OCTA's superficial layer. Moreover, all the axons meet at the optic nerve which makes a circumpapillary analysis at the peripapillary ROI the best option to capture information from all of them at the same time. Nevertheless, a certain discrepancy and conflicting results have also been observed between sectors at different layers and regions of acquisition.
Possible reasons for such a variability are related to the data and respective study design, and were qualitatively evaluated. Although no significant differences were observed in terms of age (except for one study which did not provide information (Rolle et al All of these devices were Swept-Source OCT (SSOCT) which could potentially indicate that a SSOCT may provide a better OCTA image quality and, consequently, may result in higher AUR-OCs. Further research is recommended to confirm the advantages of using SSOCT for OCTA imaging in assessing glaucomatous microvascular damage. A high risk of bias was identified in eight studies that included images with an image quality below the threshold suggested by the manufacturer (see Appendix D). Two other studies did not report which threshold was used. Only one study performed a fovea-disc axis correction (Jesus et al. 2019). Due to eye motion or slight differences in position during image acquisition, OCTA images from different subjects might not match the same sectors at the same location. Therefore, sectorial analysis requires images to be previously corrected, for instance taking the fovea-disc axis into account. This way all subjects will have the same reference point for the sectorial analysis.
Another reason for the current variability between studies is related to the method employed to extract vascular density. Although it is not the focus of this review, different image processing approaches can lead to different vascular interpretations within the same subject data. A popular method among the community is the OCTA image binarization based on thresholding techniques. The ratio of white or black pixels over a specific area is used to estimate the microvascular dropout. In general, the threshold is chosen based on an empirical analysis using an image processing programme such as ImageJ (Abràmoff et al. 2004). The separation of micro-from macrovasculature is another source of variability between studies. In some studies, the macrovasculature is segmented and removed from the region of acquisition. Other authors have opted for estimating vascular density based on all the information presented on the OCTA image. Macrovasculature is not expected to be affected by glaucoma, and it is a subject-dependent anatomical feature. Thus, an analysis on image pixel intensity including macrovasculature is not desirable, as it may bias the results. Similarly, the optically hollow area inside the optic disc, as well as the foveal avascular zone (FAZ), is subject-dependent. Therefore, it is desirable to segment and exclude these areas from the ROI before the microvascular density estimation is performed. Nevertheless, further research is needed for a better understanding of the variability between mathematical approaches and to understand which is the most appropriated for glaucoma diagnosis. Although a few research lines have already considered more complex procedures, such as fractal analysis (Gadde et al. 2016), replication studies are still needed to evaluate such advanced/complex methods.
The superior and inferior sectors of the superficial layer of the peripapillary region may be suitable for the diagnosis. However, the averaged AUROC reported in the reviewed articles is still lower than the values obtained with retinal nerve fibre layer thickness (measured through standard OCT imaging) and lower than the optic disc features (extracted from fundus imaging (Hemelings et al. 2020)), which usually result in AUROC values higher than 0.9. Nevertheless, recent studies have shown that vascular density assessed by OCTA seems to perform better than the gold standard biomarkers at discriminating advanced cases of glaucoma (Barbosa-Breda et al. 2018; Van Melkebeke et al. 2018). Hence, follow-up of (advanced) glaucoma using OCTA imaging may be a window of opportunity to establish OCTA as a common practice in the clinical environment. Thus, new studies will be required to infer which OCTA ROI is the best at glaucoma follow-up.

Conclusions
This review provides a comprehensive summary of the research on glaucomatous microvascular damage based on the analysis of different ROIs imaged with OCTA. The collected data show that the superficial layer in the peripapillary region is the most informative to infer vascular damage. Furthermore, at this location and layer, the inferior and superior sectors have been found as the most discriminative ROIs to study glaucomatous vascular damage with OCTA.    In the work published by Lommatzsch et al. (2017), two studies were performed. In each study, a different number of eyes were examined using one of two devices (Optovue or the Zeiss Cirrus).