Detection of Structural Glaucoma Progression with Deep Learning on Serial Optic Disc Photographs

PURPOSE: To design a supervised deep learning (DL) model to detect glaucoma progression with serial optic disc photographs (DPs). DESIGN: A retrospective longitudinal cohort study. PARTICIPANTS: One thousand five hundred ten eyes (916 patients) with ≥2 years of follow-up and 2 pairs of DPs per eye were included. METHODS: Longitudinal series of DPs were labeled as having evidence of progression or stable by 2 ophthalmologists, and discrepancies were adjudicated by 2 glaucoma specialists. An automated cropping was applied centered on the optic disc to reduce less relevant information. The dataset was split into training and testing/validation sets with an 80/10/10 ratio. A twin convolutional neural network was designed to assess baseline and final DPs to detect glaucoma progression. MAIN OUTCOME MEASURES: Area under receiver operating characteristic curves (AUCs) for detection of glaucoma progression; sensitivity and specificity for automated classification compared to clinical classification as ground truth. RESULTS: Baseline visual field mean deviation (±standard deviation) was -4.0 (±5.6) dB. Twenty-two percent of eyes deteriorated based on the clinical review of DPs. The final DL model's AUC (95% confidence interval) for detection of glaucoma progression was 0.821 (0.764-0.887) with an overall accuracy for classification of 72% (66%-88%) with a sensitivity and specificity of 87% (60%-95%) and 68% (61%-93%), respectively. CONCLUSIONS: Our twin convolutional neural network model is able to detect glaucoma progression with clinically relevant accuracy. Deep learning is promising as an adjunctive method for clinical decision-making for detection of structural glaucoma progression. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.