Abstract
Purpose :
Retinal nerve fiber layer (RNFL) thickness as assessed with OCT may be overestimated due to a logarithmic transformation of the Fourier-transformed spectrometer output, a preprocessing step done to make low reflective layers more visible. This overestimation decreases with decreasing image quality (SNR). Indeed, it has been shown that cataract, or a neutral density (ND) filter that mimics cataract, causes a decrease in the observed RNFL thickness. Theoretically, the logarithmic transformation induced overestimation could be avoided by segmenting the raw OCT intensity signal instead of the logarithmic intensity image. The aim of this study was to compare OCT derived RNFL thickness as a function of SNR for various analysis methods.
Methods :
OCT scanning was performed in 20 healthy subjects with and without ND filters; optical density was 0, 0.3, 0.6, and 0.9. The macular region was scanned using a Canon HS-100. We reconstructed the raw OCT signal from the log-transformed image by assuming that the 0-255 gray scale corresponded to a dynamic range of 30 dB. FWHM measurements from (1) the log-transformed data (the original image) and (2) the (reconstructed) raw signal A-scans were compared to each other, to (3) the built-in software, and to (4) a generic segmentation algorithm, the IOWA reference algorithm. The thickness was averaged over an area corresponding to the ETDRS grid.
Results :
Mean RNFL thickness ± SD from the FWHM of a Gaussian curve fitted to the original logarithmic intensity image was 58.5±4.1, 54.4±3.1, 43.3±3.2, and 34.2±3.5 μm for optical density 0, 0.3, 0.6, and 0.9, respectively. For the reconstructed raw intensity signal this was 35.2±3.5, 35.2±3.3, 33.3±3.9, and 28.8±4.9 μm. Canon software reported values were: 40.2±3.8, 41.0±3.6, 37.8±3.6 and 34.5±3.4 μm. IOWA algorithm values were: 32.5±4.4, 27.5±3.7, 17.5±2.5 14.1±4.5 μm. For optical density 0, the mean RNFL thickness was significantly different for each pairwise comparison of the four algorithms (two-way ANOVA P<0.001; post-hoc tests all P<0.02). The effect of the filters was different for the different algorithms (repeated-measures ANOVA; significant interaction between filter and algorithm; P<0.001).
Conclusions :
Different analysis methods differ in the RNFL thickness they report and in the corresponding SNR dependence. The most accurate and robust technique still has to be defined.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.