Spatial optimisation of retinal thickness analyses for improved accuracy of diagnosing age-related macular degeneration (AMD)

Abstract

Purpose : Diagnosing AMD remains a challenge, partly due to the abundance of clinical data available for analysis. To mitigate this and harness the potential of clinical data for developing automated diagnostic algorithms; the study aims to determine the optimal spatial extent of retinal layer thicknesses which maximises diagnostic accuracy of early and intermediate AMD versus normal eyes.

Methods : A total of 192 eyes (48 intermediate AMD [iAMD], 48 early AMD [eAMD], and 96 normal eyes) were propensity-score matched by age, sex, and refraction. Optical coherence tomography macular cube scans (24×24° or ~6912×6912µm) were manually-segmented for retinal nerve fibre (RNFL), ganglion cell (GCL), inner plexiform (IPL), inner nuclear (INL), outer plexiform/outer nuclear (OPL/ONL), inner and outer segment (IS/OS), and retinal pigment epithelium to Bruch’s membrane (RPE-BM) layers. Each layer was divided into 60×60 (0.01mm²) grids, averaged for thickness across concentric areas of varying diameters from the fovea, and regressed against diagnoses of iAMD, eAMD, and normal eyes to determine diagnostic accuracies (area under the curve [AUC]).

Results : Diagnostic accuracy (AUC, [95% CI]) for iAMD vs normal (0.9 [0.84, 0.95], P < 0.0001) (Fig. 1A) was maximised using retinal layer thicknesses within an area of 300µm radius, modelled with RPE-BM, RNFL, INL, and IPL in descending order of absolute magnitude (|Z|, 4.2 to 1.9) (Fig. 1B). Diagnostic accuracies for eAMD vs normal (0.86 [0.79, 0.92], P < 0.0001) (Fig. 1C) and iAMD vs eAMD (0.84 [0.76, 0.92], P < 0.0001) (Fig. 1E) were maximised using retinal layer thicknesses within a 100µm radius, with the former modelled from RNFL, INL, IS/OS, and OPL/ONL (|Z|, 4.71 to 2.68) (Fig. 1D), and the latter from RPE-BM, INL, and IPL (|Z|, 2.1 to 1.3) (Fig. 1F).

Conclusions : Diagnostic models of eAMD and iAMD can be efficiently developed by evaluating retinal layer thickness within 300µm eccentricity from the fovea; optimising input for future automated AMD diagnostic algorithms. The inclusion of inner retinal layers into diagnostic models warrants further research into investigating inner retinal changes in AMD.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

Diagnostic accuracies for iAMD vs normal (A), eAMD vs normal (C) and iAMD vs eAMD (E). Absolute magnitude of standardised effect size of each retinal layer on regression analysis for iAMD vs normal (B), eAMD vs normal (D) and iAMD vs eAMD (F).