Deep-learning supported pointwise structure-function correlation from healthy eyes to intermediate and late non-exudative age-related macular degeneration

Abstract

Purpose : Therapeutic advances in geographic atrophy (GA) increase the unmet need to link morphological retinal biomarkers with functional endpoints beyond visual acuity testing. The aim of this work is to establish spatial correspondence between deep-learning (DL) quantified Spectral-Domain Optical-Coherence-Tomography (SD-OCT) biomarkers and retinal function using Microperimetry (MP).

Methods : 60 eyes of 60 patients, 20 with intermediate age-related macular degeneration (iAMD), 20 with GA and 20 age-matched healthy controls were prospectively examined by SD-OCT (Spectralis, 97 B-scans) and MP3 and MAIA MP in two consecutive test runs. Validated DL-algorithms quantified photoreceptor thickness (PRT), PR integrity loss (PRL), outer nuclear layer thickness (ONLT), hyperreflective foci volume (HRFV), drusen volume (DV) and retinal pigment epithelium loss (RPEL). Subretinal drusenoid deposits (SDD) were quantified by human experts via pixel-wise annotation. Pointwise co-registration was established between all stimuli and the corresponding OCT B-scans. Multivariate mixed-effect models were calculated to access pointwise retinal sensitivity (PWS) changes for each biomarker, accounting for age and retinal eccentricity.

Results : 10.800 PWS values of 60 eyes were analyzed. PWS was significantly lower in stimuli with PRL without RPEL (-2.81dB,0°) and in areas with both, PRL and RPEL (-10.03dB,0°) compared to areas without any PRL and RPEL (p<0.0001) in both devices in GA. In iAMD, decrease in PRT (0.15dB/μm,0°) and ONLT (0.06dB/μm) and increase in HRFV (-8.56dB/nl) and DV (-0.69dB/nl) had a significant negative effect on retinal sensitivity in MP3 and MAIA (p<0.0001). There was a significant difference between MP3 and MAIA in both disease stages, -3.55dB and -3.36dB/device in iAMD and GA, respectively (both p<0.0001). Overall sensitivity differed significantly between GA, iAMD and controls (p<0.001).

Conclusions : There is a pointwise spatial correlation between DL-quantified biomarkers on SD-OCT and retinal function assessed with MP throughout different disease stages of non-exudative AMD measured in two MP devices. Combination of DL-based SD-OCT assessment and functional testing fills the knowledge gap between imaging morphology and functional testing for the evaluation of individualized disease progression and novel interventional targets.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.