This study examined the effects of preventable operational and analytic aspects of the SD-OCT on the overall accuracy of ETDRS retinal thickness plots. Scan density, position of the scan with respect to the foveal center, and magnitude of subject axial length differential all contribute to significant error in computing retinal thickness from SD-OCT volumes. An important point to consider is the cumulative nature of the errors reported here; these parameters should all be accounted for when developing normative databases or analyzing specific retinal features within individual patient data. While the errors were estimated using a single SD-OCT device (Cirrus HD-OCT), they are generic to SD-OCT imaging in general. The issue of scan positioning is typically something that can be addressed by the operator by repositioning the ETDRS grid (either manually or using an automatic function like Autofovea). Currently, correcting the lateral scale of OCT data/images requires offline correction by the user.
In comparing our results to previously published data, we find similarities and differences. In an examination of B-scan density, Sadda et al. concluded that 32 B-scans result in only a minimal change in retinal thickness [18
]. Our data also show that when examining maps of retinal thickness that are based on spatially integrating individual thickness values (i.e., ETDRS), reduced B-scan sampling has minimal impact. However, if interested in deriving absolute measures of retinal thickness at any given point, reduction to 32 B-scans (a value suggested to provide accurate retinal thickness maps), results in an average error of around 3μ
m per pixel. While this average error is within the system resolution on commercial SD-OCT systems, it is worth keeping in mind that the error at any one pixel can be much larger, since not all pixels will contribute equally to the total error (which is implicit in computing an average error). We feel this more accurately reflects the “real” cost of undersampling, and this would significantly limit the ability to make precise measurements of retinal features (e.g., drusen). This highlights the importance of considering how the SD-OCT data is going to be used when deciding how densely to sample the retina.
It is well documented that differences in axial length result in different ocular magnification of retinal images and thus can affect the accuracy of measurements of retinal features [27
]. With respect to OCT, axial length has been shown to influence measurements of retinal nerve fiber layer (RNFL) thickness [28
]. This of course is based on the fact that RNFL measures are presumed to be taken at a fixed distance from the optic nerve; thus individual differences in ocular magnificent would result in the RNFL being measured at the wrong location. Here we demonstrate that individual differences in ocular magnification also affect the accuracy of macular thickness maps. If the distribution of axial lengths in a normative database does not match that of the subject population being studied, misinterpretation can occur. Perhaps more important than retinal thickness maps is the fact that not correcting the nominal scan length for differences in axial length will obviate making reliable measurements in the lateral dimension within a given OCT dataset. This could include measuring the area of geographic atrophy, the size of a macular hole, or the size of a druse. Despite this, some SD-OCT systems still output lateral scale bars on their images that are given in μ
m or provide calipers with which to make lateral measurements in μ
m, despite no correction for axial length having been made. One should avoid using such scale bars to report absolute length measurements, as they are simply not accurate without first taking into account ocular magnification.
There have also been previous examinations of the effect of fixation on the accuracy of OCT thickness measurements. In glaucoma, it has been shown that if the circular scan is not centered on the ONH, the RNFL thickness measurements are inaccurate [32
]. Campbell et al. [33
] examined how intentionally shifting the center of macular volume OCT scans (Stratus time-domain) affected central subfield thickness measurements for 10 normal subjects. They found that scan decentration of 0.50
mm resulted in foveal thickness measurements that were in error by about 45%. For our normal subjects, the average decentration of the SD-OCT volume with respect to the foveal center was 0.09
mm and the average error of foveal thickness measurements was about 35%. While this is roughly consistent with the finding of Campbell et al. [33
], some discrepancy would be expected given our use of SD-OCT (instead of time domain) and our ability to precisely determine the exact misalignment between the two scans being compared (whereas the previous study would have be confounded by errors due to normal fixational instability). Currently, the Cirrus HD-OCT will automatically position the ETDRS grid over the center of the fovea (after the scan is taken). While this results in a more accurate ETDRS map, it may not be valid to compare these maps to a database in which the ETDRS maps were not centered on the fovea, though in the case of the Cirrus database, good centration of the volume on the fovea was an inclusion criterion. It is generally important to ensure that the scan parameters used to develop the normative database match that of the on-board scan protocol. Moreover, the subject composition (race and gender) may also need to be considered when comparing a specific patient to a particular normative database [12
There are several limitations to the present study. First, in our examination of B-scan sampling, we used 128 B-scans as the “truth”. This was simply due to a limitation of the specific SD-OCT device being used. However, as we showed in , 128 B-scans (at 512 A scans/B-scan) only sample 29% of the nominal 6
m × 6
mm volume. Thus these volumes are likely already in error compared to an isotropic volume of 512 B-scans. With the expected availability of even faster OCT systems, it will be important to quantify the level of inaccuracy systematically across more densely sampled volumes. In addition, we likely underestimate the real effect of undersampling, as we used simulated thickness maps. If one were to really only acquire 32 B-scans, this could affect the accuracy of segmentation as many OCT devices use 3D approaches to make correct assignment of layers. A second limitation is that we corrected for ocular magnification using a linear scaling based on axial length. There are other methods to correct for ocular magnification [26
], and the exact method used for the correction would influence the measured differences in retinal image magnification. Finally, we did not subanalyze different pathologies. It seems likely that different retinal pathology would suffer more (or less) than others. Intuitively, one can conclude that the more uniform the retinal thickness contoured (as might occur in retinitis pigmentosa, where the retina is uniformly thin), the less impact the B-scan sampling, axial length, and scan position would have. Likewise, retinal pathology that results in significant peaks and troughs in retinal thickness (macular holes, AMD, diabetic macular edema) might be more significantly influenced by these parameters. A more detailed, disease-specific analysis is required to clarify this issue.
It is important to keep in mind that the relevance of these errors of course ultimately depends on the clinical application. For monitoring patients over time, relative differences in retinal thickness would be generally unaffected by axial length, though comparing populations of patients (such as in a clinical trial) where there may be differences in axial length between the groups could result in significant error. If one uses the same sampling density, then the accuracy of these longitudinal measurements of retinal thickness will be on the order of that reported for previous repeatability and reproducibility studies. However, in instances where one is interested in correlating a measure of retinal thickness over a specific retinal area (e.g., central subfield thickness) with some other measure of vision (such as treatment response) these errors could reveal correlations that do not exist or hide ones that do exist. Moreover, where one is interested in making absolute measurements in the lateral dimension, such as foveal pit morphology [12
] or drusen volume [35
], it is critical that these sources of error be removed.