Determinants of Cone and Rod Functions in Geographic Atrophy: AI-Based Structure-Function Correlation.

To investigate the association between retinal microstructure and cone and rod function in geographic atrophy (GA) secondary to age-related macular degeneration (AMD) by using artificial intelligence (AI) algorithms. Prospective, observational case series. A total of 41 eyes of 41 patients (75.8 ± 8.4 years old; 22 females) from a tertiary referral hospital were included. Mesopic, dark-adapted (DA) cyan and red sensitivities were assessed by using fundus-controlled perimetry ("microperimetry"); and retinal microstructure was assessed by using spectral-domain optical-coherence-tomography (SD-OCT), fundus autofluorescence (FAF), and near-infrared-reflectance (IR) imaging. Layer thicknesses and intensities and FAF and IR intensities were extracted for each test point. The cross-validated mean absolute error (MAE) was evaluated for random forest-based predictions of retinal sensitivity with and without patient-specific training data and percentage of increased mean-squared error (%IncMSE) as measurement of feature importance. Retinal sensitivity was predicted with a MAE of 4.64 dB for mesopic, 4.89 dB for DA cyan, and 4.40 dB for DA red testing in the absence of patient-specific data. Partial addition of patient-specific sensitivity data to the training sets decreased the MAE to 2.89 dB, 2.86 dB, and 2.77 dB. For all 3 types of testing, the outer nuclear layer thickness constituted the most important predictive feature (35.0, 42.22, and 53.74 %IncMSE). Spatially resolved mapping of "inferred sensitivity" revealed regions with differential degrees of mesopic and DA cyan sensitivity loss outside of the GA lesions. "Inferred sensitivity" accurately reflected retinal function in patients with GA. Mapping of "inferred sensitivity" could facilitate monitoring of disease progression and serve as "quasi functional" surrogate outcome in clinical trials, especially in consideration of retinal regions beyond areas of GA.

Informations de copyright

Copyright © 2020 Elsevier Inc. All rights reserved.