When we initially conceived our individualized structure-function mapping approach, some relevant features of interest were not readily available in empirical datasets, hence prior comparisons with other mapping schema have been performed using simulation. Several features key to the model are now either routinely available outcomes from clinical OCT (the position of the optic nerve head relative to the fovea), or are readily measurable (temporal raphe). This is the first study to directly compare mapping schema using such an empirical dataset. On average, for the dataset used herein, the CUSTOM-MAP has benefits in some eyes, and demonstrates on average better concordance between structure and function than the POP-MAP approach. In particular in the arcuate regions, CUSTOM-MAP was more likely than POP-MAP to match with damaged OCT (
Fig. 6, green wedges), however, POP-MAP matched better with OCT damage in the inferior macular region (
Fig. 6, purple wedges).
Modeling studies can act as a driver to find empirical support to either prove or disprove assumptions of the model. For example, in our first detailed exploration of our CUSTOM-mapping model, our assumptions regarding the temporal raphe positioning were incorrect (we assumed continuation of the FoDi line). Subsequently, high-resolution transverse section OCT imaging of the raphe have enabled this assumption to be corrected,
11,13,34 and we have refined both the model and knowledge of population variance of this anatomic feature.
A key feature of our current model that similarly may deserve empirical investigation is the assumption that axons between the fovea and disc take a direct, straight path into the optic nerve head (rather than a longer curvilinear path). These straighter trajectories largely drive the differences between CUSTOM-map and POP-map in the superior and inferior visual field regions between the optic nerve head and the macula. In this superior visual field, CUSTOM-map appeared to show advantages over POP-map in some eyes (
Fig. 6). The straight path model arises from the intuition that axons would take the shortest possible path to the optic nerve head during development. There is no clear physical reason for these bundles to follow curved trajectories to avoid the fovea like their temporal counterparts. However, bundles in this area may fasciculate with temporal bundles and follow their curved path. Visualization of reflective surface bundles in retinal photography cannot help to answer this question, because photos do not allow visualization of trajectories below the retinal surface. Straighter overlapping retinal nerve trajectories have been reported in some eyes when there are RNFL defects in retinal photos
35; however, accurate understanding of the positioning of the trajectories in the Z-plane is difficult in the 2D image. Visualization of bundle trajectories in different depth planes may now be visible with high-resolution en-face OCT in select eyes with particular patterns of arcuate damage.
Figure 9 shows a case example of an eye with extensive inferotemporal retinal nerve fiber layer drop-out. The high-resolution OCT scan (Spectralis, 30 × 25°, 241 horizontal B-scans, 30 µm separation, ART 16) is shown en face at 21 µm below the inner-limiting membrane. In the superior retina, trajectories of the bundles appear curvilinear; however, the large amount of reflectile bundle tissue is problematic for visualization of bundles at different depths. Simply sectioning the OCT data en face at slight differences in depths through healthy tissue does not solve this problem. Inferiorly, because of the absence of surface bundles, a wedge of deeper bundles are visible entering the optic nerve head. These appear to be straighter in profile, suggesting that the trajectory of RNFL bundles may indeed differ in different depth planes. Clearly, a larger dataset of high-resolution en-face images in eyes with localized retinal nerve fiber bundle defects is required to explore this systematically; however, this case example is suggestive of differences in nerve fiber trajectories at different depth planes. Interestingly, previous reports of straighter fibers visualized in occasional retinal photos have also been in the inferior temporal retina.
36 Perhaps there is an anatomic asymmetry in the bundle trajectories between the superior and inferior retina that might contribute to the trend for POP-MAP to show some advantages over CUSTOM-MAP in the inferotemporal macular region of the visual field in a limited number of eyes (purple sectors in
Fig. 6A). Future high-resolution OCT imaging may assist in resolving these questions.
In choosing our analytical approach, we considered very carefully assumptions that are implicit in various presentations of structure-function relationships and settled on using a simple approach to validate point-wise concordance between visual field and OCT outcomes. In addition to the assumptions discussed, this type of approach shares similarities to that used by eye clinically. In addition to comparing mapping, there are other structure-function analyses and potential applications for which this type of concordance approach may be suitable, for example, if we simply wanted to direct the sampling of visual field testing into regions of interest that are spatially concordant with retinal nerve fiber layer damage.
37 However, a more quantitative approach that estimates the likely depth of visual field damage for a measured retinal nerve fiber layer thickness decrease is necessary if visual field algorithms are to seeded with more efficient starting estimates predicted from retinal nerve fiber layer damage,
1,2 or if structural data were to be used to predict the visual field.
38 We contest that for these approaches to have maximal benefit for individual eyes, a CUSTOM-MAP approach is warranted. It may also be the case that through the use of CUSTOM-MAP, in addition to other approaches to decrease noise in both OCT and visual field estimates, better models to enable quantitative prediction between OCT and visual field could be derived.
Here we defined a visual field defect as a total deviation of less than or equal to 6 dB. Our study inclusion criteria removed individuals with significant lens opacity, required all individuals to have reasonable visual acuity, and required the visual field defect to have a spatial pattern consistent with glaucoma. Consequently, we consider it unlikely that the visual field data is significantly contaminated by nonglaucomatous lesions (however, it may be contaminated by measurement noise). Importantly, because we are comparing the two mapping schema on the same visual field locations, any visual field location that was classified as defective but where the defect arose as a result of a reason other than glaucoma, should potentially be nonconcordant with cpRNFL thinning for both schema. For more general usage of the concordance approach for matching structural and functional damage, there may be advantages to using pattern deviation metrics to minimize potential contamination of the classification of visual field locations due to issues that affect the general height of the visual field.
To perform this analysis, we required a database of visual fields and OCT images that had estimates of the temporal raphe, in addition to the FoDi angle (angle between the fovea and the optic nerve head as illustrated in
Fig. 1A). Estimation of the FoDi is now commonplace in commercially available OCT. The temporal raphe can be visualized with improved precision with higher resolution scans. With rapidly increasing OCT scanning speeds, incorporating such measurements as standard in commercial “glaucoma” OCT protocols may be feasible. There are already automated approaches for detecting the raphe,
20,39 and information regarding changes to the apparent position, width, or reflectance of the raphe may provide useful, additional information for glaucoma management.
39,40 We have previously shown that the distribution of raphe positions is similar in those with glaucoma to age-matched controls
13; however, we are unaware of any literature describing substantial longitudinal follow-up of the raphe architecture in glaucoma. With more advanced disease, as the retinal nerve fiber bundles are lost in the raphe region, the precision of measurement of raphe position is expected to decrease. In the context of our mapping, if it is assumed that the position of remaining retinal ganglion cell bodies stays relatively fixed in the retina, then using a raphe position measured earlier in the disease process for an individual would seem appropriate.
Another recent advance in commercially available OCT for glaucoma is wide-field imaging. In this article we have concentrated analysis on cpRNFL scans because cpRNFL data is widely used in practice and has extensive normative data in a range of commercial instruments. Structure-function mapping for wide-field imaging can be performed largely by direct superimposition on wide-field imaging, if the region of interest in the OCT is the region immediately stimulated by the perimetric stimulus (with some correction in the macular region for Henle fiber displacement as described in the methods). However, any analysis that relates optic nerve head, Bruch's membrane opening, or cpRNFL data to the visual field will require mapping schema similar to ours (or previously published approaches). Our mapping schema is also relevant to retinal photography for environments where OCT is still an expensive and not-so-accessible technology.
It is worth noting that there are other individual anatomic factors that could be customized within the model but that we chose to keep as population estimates. One example is disc size in the horizontal and vertical direction. In the model there is an interplay between retinal ganglion cell density and disc size. For example, if the disc is made longer in the vertical direction, but the number of RGCs kept the same, then locations in the nasal field will have higher angles of insertion into the optic nerve head (toward “poles”) because the budget of allowable axons into the optic nerve head around the horizontal meridian is decreased and those at “the poles” is increased. Because we do not know how the total number of retinal ganglion cells relates to disc size and shape (if at all) in the population, we have chosen to avoid this complication and instead to use a population average for disc shape
23 and number of retinal ganglion cells.
26 We also chose to use a population average value for the foveal radius.
24,25 Our previous work on attempting to customize this value provided some benefit, but it was relatively limited.
41,42 Future comparisons of model performance in highly myopic eyes with glaucoma may reveal additional features such as disc tilt that may improve model performance. Our current dataset only included five eyes with myopia greater than −6D, which does not enable such analysis.
In summary, we present an updated model customized to an individual's eye that maps any location in the visual field to the optic nerve head, enabling structure-function analysis. This model allows individual maps to be derived based on a few, readily measurable anatomical features. Using an empirical database where data for these features was available, we demonstrate that the customized mapping approach results in improved concordance between structural and functional data in some eyes, when compared to a map based on population averages. We argue that structure-function mapping schema are best compared using a concordance approach. Our mapping schema is readily available for both clinical and research purposes.