A deep-learning approach to identifying quantitative OCT biomarkers predictive of visual deficit measured by microperimetry at 24 months in the OAKS trial

Abstract

Purpose : To assess predictive value of structural geographic atrophy (GA) features for macular sensitivity

Methods :
Post hoc analysis of fundus autofluorescence (FAF) and spectral domain optical coherence tomography (SDOCT) data for 339 patients in the OAKS phase 3 study assessing pegceptacoplan on GA (NCT03525613). GA in FAF images was identified by expert human graders; SDOCT-based GA features were segmented by a fully automated, validated deep-learning model. FAF and SDOCT images were topographically aligned via registration of microperimetry points. The predictive value of GA features for the retinal sensitivity (in decibels; dB) of each microperimetric point was assessed.

Results : Retinal sensitivity was inversely correlated with GA features detected with FAF for intra-lesional (Pearson correlation coefficient R -0.49), juxta-lesional (within 200µm of lesion perimeter; R -0.44), and extra-lesional (R-0.40) microperimetric regions. A stronger correlation was observed for SDOCT-based GA features (intra-lesional R -0.56; juxta-lesional R -0.52; extra-lesional R -0.49 for photoreceptor degeneration [PRD]). When considering scotomatous (<1dB) microperimetric points, 3782 did not overlap with GA features on FAF imaging; yet 53% (2042/3782) of these points overlapped with - and thus ascribable to - PRD. In regions without detectable GA on FAF, PRD at baseline was predictive of future visual loss. Of 3703 microperimetric points with overlapping PRD in isolation at baseline (i.e., without FAF grading of GA), 44% (1645/3703) became scotomatous at 24 months, compared to 22% in regions without any GA features on SDOCT or FAF (P<0.0001) (Figure).

Conclusions : Structural SDOCT biomarkers, as segmented by our deep-learning platform, can account for cross-sectional visual function, and predict future visual deficit to a greater extent than FAF.
This permits structure-function correlation and tracking while adjusting for patient-specific patterns of GA lesions, and may prove useful in clinical trial design and clinical decision-making in GA.

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