An Artificial Intelligence Approach to Assess Spatial Patterns of Retinal Nerve Fiber Layer Thickness Maps in Glaucoma

Purpose: The purpose of this study was to classify the spatial patterns of retinal nerve fiber layer thickness (RNFLT) and assess their associations with visual field (VF) loss in glaucoma. Methods: We used paired reliable 24-2 VFs and optical coherence tomography scans of 691 eyes from 691 patients. The RNFLT maps were used to determine the RNFLT patterns (RPs) by non-negative matrix factorization (NMF). The RPs were correlated with mean deviation (MD), spherical equivalent (SE), and major blood vessel locations. The RPs were further used to predict the 52 total deviation (TD) values by linear regression compared with models using 24 15-degree sectors. Last, we associated the RPs with average TDs of the central upper two locations (C2-TD). Stepwise regression was applied to remove redundant features. Results: NMF highlighted 16 distinct RPs. Twelve RPs had arcuate-like informative zones (iZones): six with superior iZones, five with inferior iZones, and one RP with a bihemifield iZone, and four with non-arcuate-like temporal or nasal iZones. Twelve, nine, nine, and nine RPs were significantly (P < 0.05) correlated to MD, SE, and superior and inferior artery locations, respectively. Using RPs significantly (P < 0.05) improved the prediction of 52 TDs compared with using 24 15-degree sectors. Using RPs significantly (P < 0.001) improved the C2-TD prediction related to thinning in the inferior vulnerability zone compared with using the 24 sectoral RNFLTs. Conclusions: Using RPs improved the VF prediction compared with using sectoral RNFLTs. Translational Relevance: The RPs characterizing both pathological and anatomical variations can potentially assist clinicians better assess RNFLT loss.