Purchase this article with an account.
Sieun Lee, Morgan Heisler, Paul Mackenzie, Marinko V Sarunic, Mirza Faisal Beg; Quantitative longitudinal and cross-sectional shape analysis of retina by registration. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):5293.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Quantitative shape analysis of multiple retinas is often limited by the lack of spatial correspondence among the retinas. We present two frameworks that register longitudinal and cross-sectional optical coherence tomography (OCT) data and enable detailed morphological comparison.
Eyes were imaged with a custom 1060-nm swept-source OCT system. The 3D images were corrected for axial motion and smoothed. Retinal nerve fiber layer (RNFL) was segmented using a 3D graph-cut algorithm.<br /> Longitudinal data were registered by voxel-wise retinal surface matching between the baseline (template) and each follow-up image (target). The surfaces were first registered approximately by mathematical current representation and reproducing kernel Hilbert space norms. Spherical demons registration followed to yield voxel-wise, exact matching.<br /> For cross-sectional data, a common atlas was generated as a mean template using the f-shape framework such that each surface could be approximated as a deformation of the atlas. The mean template RNFL surface geometry and thickness were concurrently optimized.<br /> To characterize the repeatability of the registration pipeline, 3 healthy subjects were imaged 9 times over 3 weeks, and the RNFLs of the follow-up images were registered to that of the baseline. RNFL thickness was compared at each voxel across the time points to quantify the variability. For cross-sectional analysis, mean templates were generated from 20 normal eyes of 10 subjects and 26 glaucomatous eyes of 16 subjects, and the RNFL thickness was compared at each voxel between the two groups.
Fig. 1 shows an example of longitudinal RNFL thickness maps acquired over 3 weeks, with the follow-up images registered to the baseline image. Table 1 demonstrates the high voxel-wise repeatability of the RNFL thickness measurement by registration. Fig. 2 shows the average RNFL thickness maps of the normal and glaucomatous eyes, generated by the f-shape framework, which show the spatial pattern of glaucomatous RNFL thinning. Fig. 3 shows the regional statistical significance in differentiating the normal and glaucomatous groups. Fig. 2 and 3 suggest the RNFL thinning in glaucoma may occur preferentially along the thickest part of the RNFL.
We demonstrated computational frameworks capable of spatially-detailed study of longitudinal or cross-sectional retinal morphology and voxel-wise shape variability analysis.
This PDF is available to Subscribers Only