In this study, we adapted the graph search algorithm to semiautomatically identify the choroidal layer in SD-OCT volume scans and reported its performance in normal eyes and eyes with non-neovascular AMD. Good thickness agreement was observed between the algorithm and manual segmentation of the macular choroid, in both normal eyes (
r = 0.91 for the thickness measurement for “normal group 1,” as shown in
Table 2, and
r = 0.93 for the thickness measurement for “normal group 2,” as shown in
Table 3) and those with non-neovascular AMD (drusen) (
r = 0.94 for the thickness measurement as shown in
Table 4).
Although this was not the primary purpose of this study, we observed that, compared with normal eyes, the choroid was thinner in the AMD eyes. The thickness measurement was lower in the AMD eyes, and this finding was consistent throughout the macular region sampled by the OCT. This observation of a thinner choroid in AMD is in agreement with previously published reports.
23 However, the mean age of the healthy subjects was also considerably lower than the AMD patients, and choroidal thickness is also known to decrease with age.
We hypothesized that the poor visibility of the choroid-sclera junction in some B-scans in “normal group 2” of the eight cases was due to a thicker choroid, limiting the penetration of light through its full extent. From the thickness measurement results in
Tables 2 and
3, the average choroid thickness of “normal group 2” was thicker than that in “normal group 1,” which supported this hypothesis.
The algorithm performance was best centrally, with greater segmentation errors at the temporal and nasal edges of the scans. We suspect this is due to two reasons. One reason is that the image quality of the B-scans is poorer at the edges of the scan compared with the center, possibly due to the nature of the tracking technique. In this case, the confidence in the segmentation may not be as high for both the algorithm and manual delineation, leading to a greater apparent discrepancy. The second explanation is that the graph search of the surfaces was constrained by the neighborhood smoothness, but these same constraints could not be applied at the edge of the image. Despite these peripheral failures in some cases, the overall performance of the algorithm was good. Thus, it may be useful for large-scale quantitative studies, particularly for central choroidal thickness calculation. In addition, any errors or failures of the algorithm in selected B-scans could potentially be corrected manually.
Despite the favorable performance of the choroidal segmentation algorithm, there are several limitations in this preliminary study. First, only anisotropic (1024 × 37 × 496 voxels) SD-OCT volumes were used in this study. In addition, only a limited number of B-scans were acquired for the SD-OCT volumes because of the additional time required to acquire volume scans when using a tracking OCT. As a result, the 37 B-scan volume cube (approximately 123 μm apart between adjacent B-scans) is the standard volume acquisition protocol used in the Doheny Imaging Unit. It is possible, however, that a more isotropic SD-OCT dataset could yield more accurate segmentation and resultant thickness measurements. A second limitation of the described approach is that a tuned cost penalty and smoothness constraints were applied for the drusen eye dataset. Application of these constraints could potentially affect the generalizability of this approach, as it will not be fully automated for all types of eyes and diseases. The true generalizability of the approach (or subsequent modifications) needs to be evaluated in much larger, heterogeneous datasets that will be the subject of our future studies. A third limitation of our study is the OCT scanning wavelength that was used. For the Spectralis OCT, the center wavelength is 870 nm. At this wavelength, there may be only partial penetration of the choroid even with enhanced depth imaging approaches, particularly in cases with thicker choroids. In fact, from the results presented in
Tables 2,
3, and
4, the segmentation in the group with non-neovascular AMD eyes (
Table 4) performed slightly better than that in the two normal groups (
Tables 2,
3). The major reason was probably that the thinner choroid in the eyes with non-neovascular AMD allowed a deeper penetration of the light signal and better visualization of the choroid. However, as described previously, in the two normal groups, especially for the cases in “normal group 2,” the thicker choroid limited the light penetration. Although, the graders excluded B-scans in which the outer border of the choroid was not identifiable, there may have been other cases in which the visibility was poor but grading was still deemed to be possible. The accuracy of the manual segmentation may not have been as good in these cases. Similarly, the poor signal at the outer choroid in these cases, may have affected the performance of the algorithm. In these cases, the algorithm performed a best fit of the choroid-sclera junction, which may have undermined the accuracy of the segmentation. This problem, however, may be addressed in the near future, given the recent availability of Fourier Domain OCT devices with longer wavelengths (1050 nm), such as the Topcon swept-source OCT instrument (Topcon, Oakland, NJ). The final limitation of our algorithm is that it was evaluated only in normal and non-neovascular AMD eyes. The performance in a variety of other retinal and choroidal diseases remains to be determined.
In summary, in this study, the graph-search algorithm was adapted to semiautomatically delineate and quantify the choroid in normal eyes and eyes with non-neovascular AMD. The algorithm showed excellent agreement with human expert manual segmentation, particularly in the central portions of the volume scans. The semiautomated choroidal thickness calculation may be useful for large-scale quantitative studies of the choroid in normal and diseased eyes in the future.