Machine Learning With Optical Coherence Tomography for Glaucoma Diagnosis

Machine learning classifiers combining OCT-derived minimum rim width and retinal nerve fiber layer thickness achieved high diagnostic accuracy for mild to moderate glaucoma, with combined parameters outperforming RNFL alone.

To investigate the diagnostic performance of different machine learning classifiers (MLC) trained with retinal nerve fiber layer thickness (RNFLT) and minimum rim width (MRW) measurements obtained from spectral-domain optical coherence tomography (OCT).

113 eyes with mild to moderate glaucoma and 154 healthy eyes were included. The global average of the MRW and RNFLT measurements were obtained from OCT scans of the optic nerve and peripapillary region. Ten MLCs algorithms were compared when trained with MRW data only, RNFLT data only, and both MRW and RNFLT parameters. Receiver operating characteristic (ROC) curves were built for each MLC in training set. The highest AUC in each training group were compared using the De Long method.

The Random Forest MLC presented the highest diagnostic performance when trained with both RNFLT and MRW data (AUC=0.979, 95% confidence interval [CI]: 0.959-0.995), with a sensitivity of 98% at 80% specificity and 96% at 90% specificity. When using the RNFLT data only, the Multilayer Perceptron presented the highest performance among MLCs (AUC=0.942, 95%CI: 0.887-0.974, and a sensitivity of 93% at 90% specificity), and the Random Forest had the highest AUC value when trained with the MRW data only (AUC=0.967, 95%CI: 0.937-0.985, sensitivity of 94% at 90% specificity). The AUC from MLC trained with both RNFLT and MRW parameters was significantly higher than that obtained with RNFLT alone (P=0.029).

MLC combining RNFLT and MRW data can provide high diagnostic accuracy for glaucoma.