POST-TREATMENT PREDICTION OF OPTICAL COHERENCE TOMOGRAPHY USING A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK IN AGE-RELATED MACULAR DEGENERATION

摘要

Purpose: To develop a deep learning model to generate posttreatment optical coherence tomography (OCT) images of neovascular age-related macular degeneration. Methods: Two hundred ninety-eight patients with neovascular age-related macular degeneration were included. The conditional generative adversarial network was trained using 15,183 augmented paired OCT B-scan images obtained from 723 scans of 241 patients at baseline and 1 month after 3 loading doses of an anti-vascular endothelial growth factor treatment. The network was also trained using baseline fluorescein angiography (FA) or indocyanine green angiography (ICGA) images together with baseline OCT images. A test set of 150 images of 50 eyes was used to evaluate its ability to predict the presence of intraretinal fluid, subretinal fluid, PED, and subretinal hyperreflective material. Posttreatment OCT images were compared with images generated from baseline OCT with or without FA and indocyanine green angiography images. Results: The predicted images inferred from baseline OCT images achieved an acceptable accuracy, specificity, and negative predictive value for four lesions (range: 77.0-91.9, 94.1-95.1, and 54.7-96.5%, respectively). The addition of both FA and indocyanine green angiography images improved the accuracy, specificity, and negative predictive value (range: 80.7-96.3, 97.3-99.0, and 59.0-98.3%, respectively). Conclusion: A conditional generative adversarial network is able to generate posttreatment OCT images from baseline OCT, FA, and indocyanine green angiography images.