Abstract
Purpose :
We were interested in whether differences exist in manual segmentation (MS) vs automatic segmentation (AS) of optical coherence tomography (OCT) in the Bruch’s Membrane Opening–Minimum Rim Width (BMO-MRW) and Retinal Nerve Fiber Layer Thickness (RNFLT) in glaucoma eyes, and whether differences occur more frequently in the same clock-hour sectors.
Methods :
We used spectral-domain OCT (Heidelberg Engineering, Heidelberg, Germany) to obtain BMO-MRW measurements in 24 radial Optic Nerve Head (ONH) B-scans, and RNFLT measurements in peripapillary circle scans (12 degrees) from 2 different timepoints in glaucoma eyes.
We obtained AS with the Heidelberg Eye Explorer Software. The software marked the position of the internal limiting membrane (ILM) and the BMO to calculate the MRW, and the ILM and posterior boundary of the RNFL. We then used custom software to perform manual corrections of the same parameters.
‘Machine error’ was the MRW or RNFLT AS value subtracted from the MS value. We calculated the proportional error by dividing the mean absolute error (AE) by the median MS value. To test whether errors occurred more often in any sector, we fit the AE per clock hour using a mixed effects model(R).
Results :
We included the MRW and RNFLT scans from 162 glaucoma eyes (162 subjects) at 2 timepoints 4.28 years (SD 0.48, range 2.5-5.1) apart. Average age was 70.24 (10.21, 42-91), and 94 (58.02%) were female.
Mean MRW (SD, range) with AS and MS was 253.0µm (89.2, 22.6-523.4) and 249.8µm (90.1, 11.5-534.5), respectively; and RNFLT was 85.37µm (32.8, 8.6-208.2) with AS and 86.88µm (34.1, 8.6-208.2) with MS.
The mean AE for MRW was 11.36µm (16.4, 0-166.6), and for RNFLT was 3.35µm (4.6, 0-83.2). MRW showed a larger proportional AE than RNFLT (4.49% versus 4.13%, respectively, p=.004).
The AE in the superior sector of the MRW tended to be larger, although not significant (p=.075). RNFLT did not show sectors with larger errors (p>.1 for all). We found no evidence for errors happening in the same sectors between techniques (p=.79).
Conclusions :
The error with AS is small and proportionally larger in MRW than RNFLT. The error in MRW tends to be larger in the superior sector. There is no evidence of sectorial correlation between the errors of MRW and RNFLT, suggesting no common source of error. This higher variability may decrease the ability to detect progression with MRW as compared to RNFLT.
This is a 2021 ARVO Annual Meeting abstract.