Cross-Platform Comparison of Retinal Layers in Multiple Sclerosis Utilizing a Novel Open-Source Optical Coherence Tomography Automated Segmentation Algorithm

Background:  Retinal pathology in multiple sclerosis (MS) is well described. Changes within different layers of the retina in MS may correlate with different aspects of the disease. However, widespread use of discrete retinal layer measures is hampered by a paucity of available automated segmentation algorithms operative across different optical coherence tomography (OCT) platforms.

Objectives:  To determine cross-sectional and longitudinal agreement of retinal layer thickness measures derived from an open-source, fully-automated, OCT segmentation algorithm, across two spectral-domain devices (Spectralis and Cirrus HD-OCT), and to compare these measures between MS patients and healthy controls.

Methods:  Fully automated segmentation of 172 concomitantly acquired Cirrus and Spectralis macular scans, from 68 MS patients and 22 healthy controls, was performed. A longitudinal cohort with 275 scans from 51 subjects with a mean follow up of 1.4(±0.9) years also underwent segmentation. This recently described and validated segmentation algorithm utilizes a random forest classifier to produce probability maps of boundaries between layers.

Results:  In the cross-sectional cohort, Bland-Altman analyses revealed low mean differences and narrow limits of agreement (LOA), across both devices, for ganglion cell layer + inner plexiform layer (GCIP) 0.26 µm (LOA:-2.65, 3.17), inner nuclear layer + outer plexiform layer (INL+OPL) -1.09 µm (LOA:-4.21, 2.02) and outer nuclear layer + photoreceptor segments (ONL+PR) 0.20 µm (LOA:-3.20, 3.61) thicknesses. Larger mean differences with narrow LOA were noted for macular-retinal nerve fiber layer (mRNFL) 5.11µm (LOA:1.86, 8.34) thickness measures. The mean differences and LOA for all layers were similar between MS patients and healthy controls. Compared to controls, MS patients had reduced mRNFL, GCIP and ONL  thicknesses across both platforms,  adjusting for age and gender (p<0.05 for all). Using similar analyses in the longitudinal cohort, we compared the changes in layer thicknesses between scanners and found small mean differences for changes in all layers, with moderate LOA, mRNFL -0.19 µm (LOA:-3.92, 3.52), GCIP 0.06 µm (LOA:-2.33, 2.45), INL+OPL 0.015 µm (LOA:-3.51, 3.54) and ONL+PR 0.21 µm (LOA:-3.0, 3.42).

Conclusions: Cross-sectional GCIP, INL+OPL and ONL+PR thickness measures agree well at the cohort and individual levels, as evidenced by low mean differences and narrow LOA respectively. In longitudinal comparisons good agreement at the cohort level for changes in all retinal layers, with greater variability at the individual level was noted. Utilizing this open-source segmentation algorithm, it would be possible to compare data acquired using different OCT platforms, thereby facilitating broader utilization of OCT as an outcome measure in multi-center clinical trials of neuroprotective and remyelinating therapies.