Discriminating glaucomatous and compressive optic neuropathy on spectral-domain optical coherence tomography with deep learning classifier.

BACKGROUND/AIMS: To assess the performance of a deep learning classifier for differentiation of glaucomatous optic neuropathy (GON) from compressive optic neuropathy (CON) based on ganglion cell-inner plexiform layer (GCIPL) and retinal nerve fibre layer (RNFL) spectral-domain optical coherence tomography (SD-OCT).
METHODS: Eighty SD-OCT image sets from 80 eyes of 80 patients with GON along with 81 SD-OCT image sets from 54 eyes of 54 patients with CON were compiled for the study. The bottleneck features extracted from the GCIPL thickness map, GCIPL deviation map, RNFL thickness map and RNFL deviation map were used as predictors for the deep learning classifier. The area under the receiver operating characteristic curve (AUC) was calculated to validate the diagnostic performance. The AUC with the deep learning classifier was compared with those for conventional diagnostic parameters including temporal raphe sign, SD-OCT thickness profile and standard automated perimetry.
RESULTS: The deep learning system achieved an AUC of 0.990 (95% CI 0.982 to 0.999) with a sensitivity of 97.9% and a specificity of 92.6% in a fivefold cross-validation testing, which was significantly larger than the AUCs with the other parameters: 0.804 (95% CI 0.737 to 0.872) with temporal raphe sign, 0.815 (95% CI 0.734 to 0.896) with superonasal GCIPL and 0.776 (95% CI 0.691 to 0.860) with superior GCIPL thicknesses (all p<0.001).
CONCLUSION: The deep learning classifier can outperform the conventional diagnostic parameters for discrimination of GON and CON on SD-OCT. © Author(s) (or their employer(s)) 2020. No commercial re-use. See rights and permissions. Published by BMJ.