Fig. 2
From: OCT-based diagnosis of glaucoma and glaucoma stages using explainable machine learning
(a) Illustration of top 30 features selected by feature selection techniques. (b) Human versus Machine (RF) performance comparison for Glaucoma stage diagnosis. (c) Overall glaucoma diagnosis. Two optometrists and one glaucoma specialist were provided with RNFL and GC-IPL thickness data with masking (only keeping the numerical thickness values that were fed to the ML models). All the normal and glaucoma patients were randomly assigned, so the clinicians could not predict the order in which they appeared. Once they completed the grading, sensitivity and specificity were calculated (from TP, FP, TN, FN) and plotted in the ROC plot generated for the RF model (using the same data with one-vs one approach) for comparison. Reported AUC = mean ± standard error, standard error = standard deviation/ √n , n = 5 (for five-fold cross validation). Note: RNFL: Retinal nerve fibre layer, GC-IPL: ganglion cell–inner plexiform layer, ILM-RPE: Inner limiting membrane-retinal pigment epithelium, TSNIT: temporal-superior-nasal-inferior-temporal, std: standard deviation.
