Glaucoma rose plots: redesigning circumpapillary progression analysis

Abstract

Listen

Translate article

AbstractPurposeTo evaluate glaucoma rose plot analysis (RPA), a novel display of circumpapillary optical coherence tomography (OCT) progression data.MethodsA novel RPA protocol was developed to automatically analyse and display progression analysis of circumpapillary retinal nerve fibre layer (cRNFL) OCT data. A clustering technique was employed to allow circumferential statistical determination of progressing regions of varying widths, without the use of predefined sectors. Results were presented in the form of an angular histogram for each eye. RPA was evaluated using data from primary open‐angle glaucoma (POAG) or POAG suspect eyes and compared to a gold‐standard determined by three clinicians’ assessment of OCT series and linear regression plots. Each eye was graded as ‘progressing’ or ‘stable/unaffected’. Rose plots were assessed using objective and subjective methods to confirm their relevance and practical application. Objectively, the area of red rose petals was compared. Subjectively, rose plots were assessed by three masked assessors as suspicious of progression or not.ResultsA total of 743 scans making up registered series from 98 eyes were analysed. The mean ± SD number of visits was 8.5 ± 3.8 μm in the progressing eyes, and 7.1 ± 2.8 μm in the stable or unaffected eyes (p = 0.06). Rose plots were able to distinguish progressing eyes from stable or unaffected eyes with area under receiver‐operating characteristic curve (AUROC) of 0.968 (95% CI: 0.92–1.00) compared to 0.706 (95% CI: 0.585–0.826) using global cRNFL thickness. Furthermore, agreement on progression status between clinician graders using RPA was greater than when assessing OCT scans and linear regression plots (Fleiss’ kappa = 0.86, 95% CI: 0.81–0.91 compared with 0.66, 95% CI: 0.54–0.77).ConclusionsGlaucoma RPA is a representative and intuitive progression analysis tool that can improve the reproducibility of clinical decision making. As RPA is a statistical (deterministic) technique not dependent on deep learning, this should facilitate rapid clinical translation.