In this study, we present an algorithm to automatically segment the NCO and optic cup at the RPE/BM plane in SD-OCT volumes. Using human expert planimetry on the stereo photographs as the RS, we showed that the NCO border positioning differences () between the algorithm and the RS for the 68 eyes were as good as interobserver differences. The linear cup-to-NCO area ratio () for the 68 eyes correlated reasonably well with the LCDR of the RS (r = 0.85). Other objectively derived 2D SD-OCT metrics () correlated reasonably well with those of the RS (r = 0.69~0.83). In addition, we qualitatively (, ) demonstrated the good match of the present algorithm and the RS. We concluded that in most eyes the NCO in SD-OCT was consistent with the clinically appreciable defined optic disc margin obtained by planimetry.
However, even though the NCO boundary and clinical disc margin corresponded reasonably well in most eyes, it is interesting to note the discrepancies (, ). These discrepancies were consistent with the findings reported by Strouthidis et al.12,13
In particular, we found that the clinical disc margin sometimes corresponded with the varying combinations of different structures other than the NCO, such as the border tissue of Elschnig. For instance, in the SD-OCT B-scan shown in , the RS defined the innermost termination of the border tissue as the temporal optic disc margin (blue arrow). In , the RS defined the border tissue as the temporal optic disc margin (blue arrow), which was obviously different from the NCO (yellow arrow) of the algorithm.
Because of such underlying differences, compared with the parameter correlations by planimetry between different experts, the relatively smaller correlations of the NCO-based metrics with those of the RS () were not surprising. However, the fact that the algorithm demonstrated smaller unsigned border positioning differences than those between the experts () yet had lower correlations () was surprising. This might have occurred in part because correlation (measuring the direction and noise of linear relationships) does not take into account any bias (e.g., consistent overestimation or underestimation of a parameter), whereas the unsigned border positioning errors were influenced by any bias between the measurements. Given that the experts tended to have larger biases than were found between the algorithm and the RS (as indicated by the signed errors in ), this might have contributed to the larger unsigned errors as well. In addition, it is important to note that the definition of the algorithm's cup boundary
is different from the traditional clinical definition of cup margin
. The algorithm simply defined cup as the crossing point of the reference plane with the retinal surface, whereas the human experts tended to delineate the cup margin at the inflection point of the surface slope of the cup as seen in stereo photographs.19
This may also explain our observation that cup area at the level of the RPE/BM plane was often smaller than the clinically visible cup area on stereo fundus photographs as seen by human experts ().
There are several advantages of the current automated segmentation approach over manual planimetry. First, although planimetry is the gold standard for quantifying glaucoma progression, it introduces great interobserver variability.5
Its subjective nature is one of the potential sources of interobserver variability. However, the present automatic algorithm based on SD-OCT is completely objective and, therefore, should be more reproducible (assuming the NCO is a relatively stable landmark) compared with subjective, manual segmentation by human experts, though this has yet to be rigorously demonstrated. Second, as reported,5
manual segmentation by planimetry is cumbersome and time-consuming and remains a research tool. However, the algorithm, when properly implemented, should need only a few minutes to produce the analysis and would be compatible with that used in routine clinical use. Third, as found by our automated and others' manual 12,13,20
studies, the clinical optic disc margin seems to be the projection of a number of different
recognizable anatomic landmarks, introducing greater variability between experts, depending on the landmarks they use to define “their” rim, and thus great variability for the quantification of glaucoma progression. Landmarks of the NCO will remain the same and, therefore, are expected to be relatively stable throughout the course of the glaucoma. An ideal reference plane based on a stable structure is critical in longitudinal imaging, glaucomatous analysis, and neuropathy analysis of the ONH. The NCO-based reference plane has the potential to more sensitively detect specific glaucomatous ONH changes, such as alterations in the anterior laminar surface and prelaminar neural tissue internal limiting membrane.12,13
Although NCO-based metrics cannot replace the clinically appreciated optic disc margin, because the NCO is expected to be stable, it has the potential to provide a basis for other 2D or 3D ONH parameter quantification, and this would aid clinicians to more easily and better interpret the progression of glaucoma.
Thus, one of the major advantages of our present approach over our previous voxel classification approach9
is that the present approach was able to segment natural ONH anatomic structures of NCO and optic cup at the RPE/BM complex to enable all the advantages such structures may provide (such as the ability to compute 3D parameters based on a reference plane). In the previous approach, the RS from fundus photographs was used as truth in the training phase for the voxel classification and this resulted in mimicking the subjective assessment of the clinical defined optic disc margin and optic cup seen on photographs. As shown in the examples of and 8, although the segmentation of the previous approach closely corresponds to that of the RS, it does not overlap with a single constant structure in SD-OCT volumes. However, our core hypothesis—that segmentation of NCO will allow better estimation of glaucoma progression than the voxel classification-based approach—must be tested. This is possible only in a prospective study of sufficient duration.
There are several limitations to this preliminary study. One is that we used close to isotropic SD-OCT volumes. Potentially, a fully isotropic SD-OCT can lead to more accurate segmentation and corresponding parameter measurements. The 2D measurements of this work were not substantially influenced because they were computed on an isotropic X-Y plane. However, for future volumetric measurements, if applicable, it may be desirable to compute the volumetric parameters in the isotropic OCT space. Another was that the flattening of the raw SD-OCT greatly improved the motion artifacts and enabled us to compare the NCO-based 2D metrics with the 2D metrics of the clinical optic disc margin. However, this was not perfect, as shown in . For the 2D measurements on the projection image, we corrected the nonoptimal flattening problem by extrapolating the average radial positions outside the estimated NCO with those inside it (see NCO and Optic Cup Segmentation at the RPE/BM Plane). For the volumetric measurements, it might have been necessary to transform the NCO-based reference plane back to the original raw SD-OCT space and to compute the volumetric parameters in the nonflattened isotropic space. Yet another is that, as reported,21
with the glaucomatous damage of the lamina cribrosa and the peripheral scleral connective tissue, the cup enlarged and the NCO position might have changed relative to the peripheral sclera. Strouthidis et al.12
suggested an alternative reference plane that was further from the center of the NCO boundary. This alternative reference plane could be obtained in a fixed distance from the segmented NCO and would have been less likely to deform posteriorly. Such change in reference plane position can be readily implemented in our algorithm, if desired.
In summary, we developed a novel automated graph-search-based machine algorithm to segment the NCO and optic cup at the level of RPE/BM complex in 3D OCT volumes of the ONH. In most eyes, the algorithm parameters correlated well with the RS parameters. However, a small discrepancy existed between the NCO and the clinical disc margin in some eyes. In addition, because of the relative stability of the NCO reference plane and the objective nature of the algorithm, we predict that the measurements of the NCO-based 2D or 3D glaucomatous parameters in volumetric OCT would be more reproducible than those of the RS parameters based on fundus photographs or even on the OCT parameters of the previous generation time-domain OCT. Additional work will be necessary to test our core hypothesis and to explore novel, objective, reproducible NCO-based parameters that correlate well with disease stage and progression.