Artificial Intelligence Characterization of Spatial Patterns of Microvascular Dropout in Glaucoma

Abstract

Purpose : To characterize the spatial patterns of microvascular dropout in glaucoma using an unsupervised artificial intelligence approach.

Methods : This retrospective cohort study included a total of 1899 Angiovue optic nerve head optical coherence tomography angiography (OCTA) scans of 707 eyes of 397 healthy, glaucoma suspect and glaucoma patients from the Diagnostic Innovations in Glaucoma Study. An unsupervised artificial intelligence technique “Archetypal Analysis” was applied to pre-processed en-face OCTA images obtained at the level of the retinal peripapillary capillary density network. Correlations of OCTA derived spatial patterns with 24-2 visual field (VF) mean deviation (MD), 24-2 VF pattern standard deviation (PSD), 10-2 VF-MD, and 10-2 VF-PSD were evaluated. Diagnostic accuracy of spatial patterns to detect past glaucoma progression was calculated and compared with those of circumpapillary capillary density (cpCD) and retinal nerve fiber layer thickness (cpRNFLT) using 310 eyes (of 195 glaucoma suspect and glaucoma patients) with a minimum of 3 years of follow-up and 5 reliable 24-2 VF tests before their last OCTA image. Progression was defined based on an event-based glaucoma progression analysis (GPA) criterion. Generalized mixed-effects model was used to adjust for correlations of different metrics at the patient and eye levels.

Results : Eleven distinct spatial patterns were identified across the spectrum of disease severity (Figure 1). Notably, 10 of the 11 patterns of microvascular loss preserved the less vulnerable papillomacular area. Eight (1,5,6,7,8,9,10,11), 5 (1,7,8,10,11), 5 (1,8,9,10,11) and 5 (1,3,8,10,11) spatial patterns were significantly associated with 24-2 MD, 24-2 PSD, 10-2 MD and 10-2 PSD, respectively (P<0.05 for all). The archetypal model (area under the receiver operating characteristic curve [AUC]=0.75) outperformed cpCD (AUC=0.66) and cpRNFLT (AUC=0.65) in detecting past glaucoma progression (P=0.017).

Conclusions : Unsupervised artificial intelligence techniques are capable of identifying patterns of microvascular dropout in glaucoma using OCTA images. These patterns show promise not only in qualitative evaluation but also in the quantitative assessment of glaucomatous microvascular dropout associated with VF damage and VF progression.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1: Spatial patterns (SP) of microvascular dropout (right eye format) in glaucoma identified by archetypal analysis.