Early Detection of Optic Nerve Changes on Optical Coherence Tomography Using Deep Learning for Risk-Stratification of Papilledema and Glaucoma

Background: The use of artificial intelligence is becoming more prevalence in medicine with numerous successful examples in ophthalmology. However, much of the work has been focused on replicating the works of ophthalmologists. Given the analytical potentials of artificial intelligence, it is plausible that artificial intelligence can detect microfeatures not readily distinguished by humans. In this study, we tested the potential for artificial intelligence to detect early optic coherence tomography changes to predict progression toward papilledema or glaucoma when no significant changes are detected on optical coherence tomography by clinicians. Methods: Prediagnostic optical coherence tomography of patients who developed papilledema (n = 93, eyes = 166) and glaucoma (n = 187, eyes = 327) were collected. Given discrepancy in average cup-to-disc ratios of the experimental groups, control groups for papilledema (n = 254, eyes = 379) and glaucoma (n = 441, eyes = 739) are matched by cup-to-disc ratio. Publicly available Visual Geometry Group-19 model is retrained using each experimental group and its respective control group to predict progression to papilledema or glaucoma. Images used for training include retinal nerve fiber layer thickness map, extracted vertical tomogram, ganglion cell thickness map, and ILM-RPE thickness map. Results: Trained model was able to predict progression to papilledema with a precision of 0.714 and a recall of 0.769 when trained with retinal nerve fiber layer thickness map, but not other image types. However, trained model was able to predict progression to glaucoma with a precision of 0.682 and recall of 0.857 when trained with extracted vertical tomogram, but not other image types. Area under precision-recall curve of 0.826 and 0.785 were achieved for papilledema and glaucoma models, respectively. Conclusions: Computational and analytical power of computers have become an invaluable part of our lives and research endeavors. Our proof-of-concept study showed that artificial intelligence (AI) algorithms have the potential to detect early changes on optical coherence tomography for prediction of progression that is not readily observed by clinicians. Further research may help establish possible AI models that can assist with early diagnosis or risk stratification in ophthalmology.