Purchase this article with an account.
Ali Raza, Donald Hood; Maps Relating OCT Retinal Nerve Fiber Layer to Visual Fields Influence Evaluations of Structure-Function Models. Invest. Ophthalmol. Vis. Sci. 2013;54(15):2257.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the effect of using a particular map relating optical coherence tomography (OCT) to visual fields (VFs) when evaluating structure-function models.
Optic disc frequency domain OCT (Topcon, Inc.) volume scans and 24-2 VFs (Zeiss, Inc.) were obtained from 96 eyes of 75 glaucoma patients and suspects (56.4 ± 12.4 yrs, MD -4.0 ± 5.3 dB), 48 eyes of 48 healthy controls (51.4 ± 7.4 yrs, MD -0.2 ± 0.9 dB), and 15 eyes of 11 patients with severe ischemic optic neuropathy (ION; 64.7 ± 10.9 yrs, MD -19.2 ± 7.1 dB). Circumpapillary retinal nerve fiber layer (cpRNFL) measurements were obtained by hand-correcting a segmentation algorithm.1 Using two maps relating OCT to VFs, Garway-Heath et al. (GH, Fig. 1A)2 and Harwerth et al. (H, Fig. 1B),3 structure-function relationships were plotted for both the superior and inferior retina (Fig. 2A). Two models were compared: the Harwerth et al. non-linear model (H-NLM)3 and the Hood and Kardon linear model (HK-LM).4 The 8 H-NLM parameters were obtained from the literature3 and are independent of any particular OCT to VF map. Because the 2 HK-LM parameters depend on the map used, they were derived from the control and ION populations (green and red in Fig 2A). Models were evaluated on the independent glaucoma population (without free parameters) using the coefficient of variation of the root mean squared error (CVRMSE), where 0 indicates a perfect model fit.
The residuals for the H-NLM were considerably larger (Fig 2B,C). The HK-LM did better than the H-NLM using the GH map by about a factor of 5 (superior CVRMSE 0.25 vs 1.22; inferior 0.28 vs 1.55; Fig 2B) and the H map by about a factor of 2 to 3 (superior CVRMSE 0.22 vs 0.67; inferior 0.24 vs 0.56; Fig 2C).
Given the H-NLM converts VF and cpRNFL values into ganglion cell counts, it is difficult to understand the model’s poorer performance when using the anatomically more defensible GH map. Therefore, further work is needed to determine the accuracy of these derived ganglion cell estimates. 1Raza et al. AO 2011, 2Garway-Heath et al. IOVS 2002, 3Harwerth et al. PRER 2010, 4Hood & Kardon PRER 2007
This PDF is available to Subscribers Only