Software updates of OCT segmentation algorithms influence longitudinal assessment of retinal atrophy

To investigate whether there is a systematic difference in peripapillary retinal nerve fiber layer (pRNFL) thickness measurements between subsequent updates of pRNFL segmentation software provided by Heidelberg Spectralis optical coherence tomography (OCT).

In total, 838 pRNFL scans from 213 multiple sclerosis (MS) patients and 61 healthy controls were analyzed. All scans were performed on the same OCT device followed by automated segmentation (HRA 5.6.4.0) and data extraction. Subsequently, all scans were re-segmented with an updated software version (HRA 6.0.7.0). To assess level of agreement between the two algorithms, Bland-Altman Plots were constructed. Paired samples t-test and linear regression analyses were used to investigate for differences in mean thickness and proportional bias respectively.

Overall, the updated version showed an overestimation of 0.16μm [95%CI 0.097-0.23, p<0.001] for the global pRNFL thickness compared to the earlier version. The largest differences were found for the nasal inferior (mean ∆ 0.29μm, p<0.001) and temporal inferior (mean ∆ 0.43μm, p<0.001) sectors. Inspection of the Bland-Altman Plot revealed that the difference between the two versions could be up to 6μm for the global mean. There was no proportional bias for the global mean (β=0.003, p=0.245) nor for any of the separate sectors.

The data show a significant difference in pRNFL thickness measurements between two subsequent versions of the same segmentation software. Although the mean difference was relatively small, the differences within the individual subject could be considerably higher than the known atrophy rate of 1μm/year in MS.