Abstract
Purpose:
The purpose of this study was to compare concordance between ganglion cell-inner plexiform layer (GCIPL) data from the Cirrus optical coherence tomographer (OCT) Ganglion Cell Analysis (GCA) and visual fields (VFs), with and without Drasdo displacement.
Methods:
From 296 open-angle glaucoma participants, GCIPL deviation and raw thickness data were extracted over locations per the 10-2 VF test grid, with and without application of Drasdo displacement, with global and eccentricity-dependent sensitivities and specificities calculated for both. With OCT and VF data classified as within or outside normative limits, pattern deviation values were compared using paired t-tests and Spearman correlations. Regression models were applied to pattern deviation values as a function of GCIPL thickness, and differences in model performance with and without displacement were compared using extra sums-of-squares F tests.
Results:
There were small but significant improvements in global specificity without displacement (0.58–0.59 with displacement and 0.61 without displacement), without notable differences in sensitivity (0.77–0.78 with displacement and 0.76–0.78 without displacement). At abnormal VF locations and without displacement, a higher proportion of correct OCT classifications (P = 0.0008) and significant correlation with worsening pattern deviation values were observed (r = 0.50, P = 0.002). Regression models indicated significantly steeper slopes with Drasdo displacement centrally (P = 0.002–0.04).
Conclusions:
With GCA deviation maps, small improvements in structure-function concordance were observed without displacement, which are unlikely to be clinically meaningful. Using GCIPL thickness data, significantly better structure-function concordance was observed centrally with Drasdo displacement.
Translational Relevance:
Applying Drasdo displacement on probability-based reports is unlikely to alter clinical impressions of structure-function concordance, but applying displacement with GCIPL thickness data may improve detection of structure-function concordance.
For macular structure-function analyses using readily available clinical data, Ganglion Cell Analysis (GCA) printouts were generated for each participant from Cirrus OCT macular cube scans, and saved as portable networks graphic (PNG) files to preserve pixel-wise color information. The GCA displays ganglion cell-inner plexiform layer (GCIPL) measurements captured across the entire scan area, measuring 6 × 6 mm, with quantitative measurements and comparisons to the inbuilt normative database calculated over an elliptical annulus with outer dimensions of 4.8 × 4.0 mm horizontally and vertically, and inner dimensions of 1.2 × 1.0 mm horizontally and vertically. For preservation of location-specific data, the deviation map was chosen for analyses based on the GCA printout, where GCIPL thickness information is categorized based on percentiles of the normative distribution, with yellow and red pixels indicating data points lower than the fifth and first percentiles, respectively.
The displacement model applied in this study was derived from figure 6 in the study by Drasdo et al.,
8 depicting RGC displacement as a function of cone eccentricity averaged across all principal meridians, which is consistent with many previous studies investigating the macular structure-function relationship.
9–11,13,18 Degree coordinates for the 10-2 VF stimulus centers were converted to millimeters per Drasdo and Fowler,
19 and corresponding displacements were subsequently applied to displace the VF stimulus centers. Circles were drawn around displaced and nondisplaced centers to identify locations over which OCT information would be extracted, with the circle radius of 0.215 degrees matching projected Goldmann III stimulus sizes as used in 10-2 SITA Fast testing. For both displaced and non-displaced locations, only data from locations where entire circles fell consistently within the GCA elliptical annulus were flagged for extraction.
A custom algorithm to extract GCA deviation map probability values from displaced and non-displaced 10-2 VF locations were written using MATLAB version R2019a (Mathworks, Natick, MA;
Fig. 1). Each deviation map was flipped vertically to match VF results and converted to right eye format. To ensure accurate centration of the superimposed extraction locations, foveal centers were manually selected via visual inspection. The algorithm extracted averaged color pixel information in RGB format, with a minimum difference in pixel values of 30 required between the red/green and blue values and red and green/blue values for classification at less than the fifth percentile and less than the first percentile, respectively, with any location classified as at least less than the fifth percentile subsequently denoted as defective on OCT.
Although deviations from the normative distribution are commonly used to identify structural defects on OCT, these are not always reflective of underlying GCIPL thickness measurements. For precise extraction of raw GCIPL thickness data, Cirrus OCT macular cube scans were exported in image format (IMG), and GCIPL segmentation was performed using OCT Segmentation version 2.11, an open-source software tool available as part of the Automated Retinal Analysis (AURA) tool (
https://www.nitrc.org/projects/aura_tools/; provided in the public domain by Neuroimaging Tools and Resources Collaboration, University of Massachusetts Medical School, Worcester, MA).
20 An additional advantage is that raw GCIPL thickness data can be extracted over the entire OCT scanning area of 6 × 6 mm, enabling comparisons to close to all VF data from the 10-2 test grid. GCIPL thickness maps were flipped vertically to match VFs and converted to right eye format. Then, an additional MATLAB algorithm was written to extract averaged GCIPL thicknesses over displaced and nondisplaced locations, as described above, with the foveal location automatically detected from data in corresponding extensive markup language files (XML;
Fig. 2).
To compare sensitivity and specificity of displacement versus no displacement, data points were classified using the pattern deviation probability data from VFs and extracted deviation map data from OCT:
- True positive: VF location defective and OCT location defective.
- False positive: VF location not defective and OCT location defective.
- False negative: VF location defective and OCT not defective.
- True negative: VF location not defective and OCT location not defective.
Global sensitivities and specificities were computed from these classifications, and 95% confidence intervals were calculated as described by Ying et al.
21 to enable comparisons between displacement and no displacement methods. Data were also pooled as per VF clusters described by Choi et al.,
22 where VF locations falling within the same cluster demonstrate similar central tendency and distribution characteristics (see
Fig. 1), to visualize potential eccentricity-dependent variations in sensitivity and specificity. By avoiding a priori assumptions regarding which locations demonstrate similar statistical characteristics, this approach provides more robust justification for pooling certain data points together, improving discriminability of trends in datasets.
23–25 An advantage of this method is that it specifically describes clusters with similar functional change characteristics across the 10-2 test grid. Unlike more widely known methods clustering locations within 24-2 test grids, such as the Garway-Heath map,
26 due to the relative sparsity of test locations there would be some ambiguity in applying the Garway-Heath map to 10-2 locations.
Further statistical analyses were performed using GraphPad Prism version 8.4.3 (GraphPad, La Jolla CA) and RStudio version 1.2.5042 (RStudio, Boston MA). As pattern deviation probability data may not entirely characterize the depth of VF loss, particularly with more severe defects, analyses using quantitative pattern deviation data were performed. Normality was assessed using D'Agostino and Pearson tests, with parametric or nonparametric tests subsequently applied as per results of these tests. Data across the entire cohort were categorized as VF defective or not defective, and within each category and for each pattern deviation value, the proportion of corresponding OCT locations flagged as defective versus not defective were calculated for displacement and no displacement methods. Paired t-tests were performed to examine OCT classifications between methods across all pattern deviation values. Differences in proportions flagged between displacement and no displacement were also plotted for each pattern deviation value, with Spearman correlations applied to investigate variations in model performance with different pattern deviation values.
Finally, for comparison of GCIPL thickness measurements and quantitative pattern deviation data, GCIPL measurements from displaced and nondisplaced locations were converted to a decibel (dB) scale to match VF data.
23,27 As linear relationships between structural and VF data have been previously reported when both are expressed in logarithmic units,
28 linear regression models between GCIPL thickness and pattern deviation values were computed. Although the OCT measurement floor may limit the ability of linear regression models to describe this relationship,
10 the tendency for pattern deviation values to underestimate severity of VF defects in advanced disease may negate the advantages of more complex regression models in accounting for the floor effect.
29 As such, segmented linear regression analyses were applied, with Davies’ tests performed to determine whether significant differences in slopes on each side of the breakpoint existed, and therefore whether the structure-function relationship could be sufficiently described with a linear relationship.
For linear models, slopes were compared between displacement and no displacement using extra sums-of-squares F test, across all tested locations and separated into clusters as per Choi et al.,
22 per global sensitivity and specificity analyses. In clusters where Davies’ tests were significant for both displacement and no displacement, this process was repeated for slopes and breakpoints of the segmented linear regression models. For all analyses, statistical significance was set at
P < 5%.
In this study, comparing structural data from the Cirrus HD-OCT GCA and 10-2 VFs, there was slightly better macular structure-function concordance without the application of Drasdo displacement when using percentile-based OCT data in the included glaucoma cohort, as indicated by significantly higher specificities and a greater proportion of correct classifications at VF defective locations. This implies that when analyzing OCT results using normative database comparisons, as per reports readily available to clinicians, there is unlikely to be a notable clinical benefit of applying RGC displacement in interpretations of structure-function concordance. However, application of Drasdo displacement provides better concordance between GCIPL thickness and pattern deviation values at the central macula, suggesting that including RGC displacement at these locations in comparisons of OCT-derived GCIPL thickness data and VF data may aid clinical detection of macular structure-function concordance.
Although contrary to theoretical understandings of RGC receptive field locations relative to soma positions based on retinal anatomy, the findings of this study are largely in agreement with Hirasawa et al.,
14 who used a Nidek OCT and reported slightly higher correlation coefficients for structure-function comparisons without application of RGC displacement at the majority of locations across the macula. Considering that glaucomatous structural damage at the macula typically manifests as broad arcuate defects, even in early stages of disease,
4 if defects are wider than the magnitude of displacement then similar OCT measurements will be observed regardless of whether data are extracted with displacement applied or not. In contrast, the central 4 locations of the 10-2 VF test grid coincide with locations where RGC displacement is greatest, and without displacement GCIPL thickness is likely underestimated at these locations given encroaching upon the foveal pit. This has contributed to the larger spread in data within cluster 1 at higher pattern deviation values (see
Fig. 4) as well as significantly higher correlation coefficients at these locations, as previously reported.
14 Overall, the findings of the present study and Hirasawa et al.
14 appear to consistently indicate that applying RGC displacement to central macular locations confers greater concordance with VF data, whereas limited differences in structure-function concordance are observed with and without displacement beyond the central macula.
An alternative reason for the lack of overall improvement in structure-function concordance with displacement is that the Drasdo model may overestimate displacement, especially outside of the central 4 locations of the 10-2 VF grid. Other histological studies have measured smaller peak displacements and magnitudes of displacement with increasing eccentricity,
30,31 with differences in studies attributed to differences in samples and shrinkage. This possibility is suggested by slightly improved correlations with VF results when using the Sjostrand RGC displacement model over the Drasdo model, although no displacement still produced better correlation coefficients overall.
14 Another possibility is that these displacement models, based on histological data obtained from a few participants only, may not account for individual differences in RGC position. Customized displacement dependent on individual data, as suggested by Turpin et al.,
32 may be required to obtain better concordance between structure and function. Further investigations regarding different RGC displacement models and the role of customizing displacement on structure-function analyses may help to elucidate the most appropriate method to apply moving forward.
Interestingly, the better agreement between OCT data from GCA deviation maps without displacement and VF data appeared to be driven by the mid-peripheral macular locations, where relatively little displacement is observed compared to the central macula. Upon further investigation, this finding appeared to be driven by participants where the GCA deviation map did not flag locations as outside of normative limits, despite visible reductions in GCIPL thickness apparent on raw data (
Fig. 5). The notion that the apparently superior structure-function concordance without displacement is an artifact of errors in identifying abnormal OCT measurements is supported by the similarities in regression models between Drasdo and no displacement at the mid-peripheral macula (see
Table 3,
Fig. 4). Inbuilt normative databases do not incorporate a whole host of variables affecting inter-individual variations in GCIPL thickness, including but not limited to spherical equivalent refractive error and position of the optic disc relative to the fovea related to differences in trajectory of the papillomacular fibers,
33–35 which can translate to errors as highlighted in these examples. These findings highlight the pitfalls of overreliance on comparisons to normative databases, and reinforce the importance of interpreting these in conjunction with raw OCT data to arrive at accurate clinical assessments of pathological OCT findings.
Although 10-2 VF data were used to compare structure-function concordance with and without Drasdo displacement, these findings do not necessarily indicate the ability of VF testing to accurately identify glaucomatous damage. Indeed, in cases of pre-perimetric glaucoma, discordance between structural damage consistent with glaucoma and apparently clear VFs is often observed.
3,36 The findings from this study suggest that structure-function concordance can improve when using quantitative OCT and VF data at central locations, however, it is also possible that properties of central VF test parameters are suboptimal for more precise comparisons to structural measures. Previous studies have identified that using stimulus sizes scaled to spatial summation properties can improve concordance between structure and function in normal and glaucoma participants,
11,12,37,38 which may be more pertinent in the context of the macular structure-function relationship. Further investigations on optimizing VF testing parameters could improve methods of detecting structure-function concordance in clinical settings, which would be highly valuable.
As this study used retrospective clinical data, only a single VF result per participant was used for comparison with OCT data, and there is a possibility that these results would not be repeatable. However, as only 1 VF result per eye is generally obtained in clinical settings, the judgments of the presence or absence of structure-function concordance applied in this study would more closely approximate clinical decision making. Additionally, due to the patient demographic at the Centre for Eye Health, the majority of participants included in this study have early glaucoma per mean deviation criteria, and a more even spread across different severities would be ideal to more holistically describe effects across the spectrum of glaucoma. Given the current theory that little advantage is afforded with displacement due to the width of glaucomatous defects, there may be little difference, but nonetheless confirming the effects of displacement on structure-function analyses in more cases of moderate to advanced glaucoma would be valuable.
This study has demonstrated no clinically meaningful difference between GCIPL information extracted over locations with and without Drasdo displacement applied on comparisons with 10-2 VF data in a cohort of patients with glaucoma, when using OCT data categorized according to comparisons to normative data. However, with quantitative GCIPL thicknesses, significantly better relationships between structure and function were observed at the central macula. In turn, there appears to be little clinical value in applying displacement on probability-based reports of structure and function, however, with quantitative data applying displacement may improve detection of structure-function concordance. These findings may help guide frameworks to aid appraisals of the structure-function relationship in clinical settings.
Supported by the National Health and Medical Research Council of Australia Ideas Grant (NHMRC 1186915) and an Australian Government Research Training Program Scholarship (J.T.). Additionally, Guide Dogs NSW/ACT provided a PhD scholarship (J.T.), salary support (J.P. and M.K.) and support clinical service delivery at the Centre for Eye Health. The funding bodies had no role in the conceptualization or writing of the paper.
The authors would like to thank Barbara Zangerl for involvement in discussions on project direction and Daniel Rafla for assistance in data collection.
Disclosure: J. Tong, None; J. Phu, None; D. Alonso-Caneiro, None; S.K. Khuu, None; M. Kalloniatis, None