Discussion
In this study, we successfully combined high-resolution MRI scans with an automated segmentation and 3D ellipsoid fitting to describe the shape of 31 eyes. Additionally, the dependency of the resulting parametrisation on the included retinal contour and its reproducibility were evaluated. Finally, a 3D ellipsoid description based on 2D imaging data was compared with the 3D results.
The used MRI acquisition methods resulted in high-resolution 3D images of the eye that provided a base for accurate segmentation of the eye. With mean differences of only 0.1 mm, the segmented contours showed to be in line with laser interferometry results for the central retina. Although different methods can be used to correct these contours for differences in gazing direction during MR-imaging by aligning them to the same axis, these methods resulted in almost identical data with negligible mean differences of 0.03 mm. Furthermore, the reproducibility of the 3D segmented retinal contours was high, with a mean difference of 0.12 mm. Although these results were obtained using a high-resolution 7 Tesla MRI and a custom-made eye coil, it has recently been shown that images with a similar resolution and quality can be obtained using clinical 3 Tesla MRI-scanners and commercially available eye-coils,18 28 enabling clinical application of these methods.
The obtained MRI-derived retinal contours could be accurately described by ellipsoids, with mean differences of 0.03 mm for the central and 0.13 mm for the peripheral retina. Optically, these differences correspond to refractive differences of approximately 0.1 and 0.3 Diopter, which is sufficient to study for example the effect of peripheral vision on the progression of myopia in children.9 However, the parameters for the central retina, such as the horizontal radius, showed a high within-subject variability for small changes in evaluated retinal fraction, although the resulting ellipsoids accurately describe the measured contour. For these evaluations, a 1.6 mm decrease in the central radius can for example be compensated by 1.5 mm posterior shift of the ellipsoid centre (figure 3). The resulting ellipsoids, however, differ less than 0.1 mm over the central 60°, explaining the lower reproducibility and strong variation of the individual ellipsoid parameters for different included retinal fractions (online supplemental appendix B, figure B2). A similar variation was observed when the central retina would be described in terms of vertex radius of curvature and asphericity (online supplemental appendix B, figure B4).7 12 16 When the shape of the central retina is however obtained using 3D fitting with the centre of the ellipsoid fixed to the central axis at half the axial length, stable and reproducible ellipsoidal parameters are obtained without a significant increase in fitting residuals (online supplemental appendix B, figure B5 and B6). This indicates that such a reduction in df is required for meaningful comparison of the central retinal shape between subjects.
For the peripheral retina, a reproducible and stable ellipsoid description was found between 220° and 280° of the retinal contour. At these fractions, the horizontal and vertical ellipsoid centre coordinates as well as the corresponding rotations remain close to zero. As a result, these parameters could be fixed in future studies, resulting in a faster fitting procedure. In 42% of the subjects, the horizontal and vertical radii differed more than 0.5 mm. This asymmetry could be relevant for ocular proton therapy planning, which currently uses a geometric eye model in which the eye is assumed to be rotational symmetrical.29 For such applications, an ellipsoid based on two orthogonal 2D images would already be an improvement, but will still result in larger differences with the measured 3D contours than a full 3D fit. The resulting differences in ellipsoid radii are however small, generally ≤0.2 mm, when the ellipse centre is fixed to the half the axial length when describing the central retina, or when this ellipse centre is included in the fitting or fixed to the centre of the vitreous body when describing the peripheral retina.
The results obtained within this study are in accordance with earlier MRI-based retinal shape studies. For the central retina, the data of the emmetropic population of Pope et al show mean horizonal and central ellipsoid radii of approximately 12 and 11 mm, which are similar to the results of this study. Additionally, their data show a similar large variation between subjects, for example 6 mm for the horizontal radius.7 16 For the peripheral retina, Lim et al present results for 240° of the retinal contour and Pope et al evaluated a slightly larger part, 270° of the retinal contour.15 16 Both studies report similar horizontal and vertical radii, 11–12 mm and 10–11 mm, respectively. The radii reported by Lim et al are slightly smaller than the values reported by Pope et al and the values reported within this study. This could be explained by the difference in the MRI resolution, as a lower resolution can result in an apparent inward shift of the retina due to partial volume effects,30 resulting in smaller radii.
While the presented results are in line with earlier MR-based retinal shape research, they differ from earlier studies using laser interferometry. The mean vertex radii of curvature for the central retina determined using laser interferometry by Verkicharla et al was about 2.0 mm larger than the currently presented vertex radius of curvature (online supplemental appendix B, figure B4).12 Even though this difference could be the result of the unstable central fit, it might also result from the difference in used imaging methods or the limited amount of data points, <20, available with laser interferometry. Due to its much larger amount of data points and the availability of 3D assessments, MRI might be more reliable than laser interferometry to quantify the retinal shape. Additionally, MRI is not influenced by refraction and is not limited to assessments of the central 80° of the retina or less, making it a more favourable method to measure the retinal shape.
Other methods to image the eye, such as CT and B-scan ultrasonography, are also not affected by refraction and could therefore be considered as an alternative to MRI, especially since they are generally faster to acquire. However, CT-scans expose a subject to radiation and have a lower resolution than MRI and B-scan ultrasonography is generally limited to a 2D field of view and has a high interobserver variability for geometrical measurements.31–33 As a result, both are less suitable than MRI to assess the 3D retinal shape.
It should be noted that the current results are based on a relatively small group of volunteers, and the presented parametric description of the retinal shape might therefore not hold for the entire population. However, the same methodology could directly be used to determine these metrics for a specific group of subjects. An additional concern could be that the presented method might be unable to describe pathological retinal shapes, for example on the presence of a staphyloma or intraocular mass. Although different authors have shown that the segmentation of MRI-scans can be adopted to include such ocular pathologies,34 35 these pathologies might result in an retinal shape that is not accurately described by an ellipsoid.36 Depending on the application, an approximate ellipsoidal description might still be sufficient in such instances, but using alternatives such as the complete 3D retinal contour should be considered.
In conclusion, this study provides a method to reproducibly determine and quantify the 3D retinal shape from MRI data. Two models are proposed, one which describes the complete posterior segment of the eye and an additional one for only the more central retina. Due to the high precision and reproducibility of this method, the resulting 3D shapes can be used as input in other research, such as optical ray-tracing simulations or the analysis of myopia progression. Furthermore, they can be used to improve the accuracy of the retinal shapes used in eye model-based treatment planning, which can for instance improve the proton therapy planning for eye tumours. With that, this study provides a complete base for widespread implementation of the 3D retinal shape as a parameter in the evaluation of the eye.