Assessment of synthetic post-therapeutic OCT images using generative adversarial network in patients with macular edema secondary to retinal vein occlusion

Shi Feng¹Jingyuan Yang¹Xinyu Zhao¹Jianchun Zhao²Yunfeng Du²Weihong Yu¹Dayong Ding²Xirong Li³Youxin Chen^1*

• ¹Department of Ophthalmology, Peking Union Medical College Hospital (CAMS), Beijing, China
• ²Visionary Intelligence Ltd., Beijing, China
• ³Renmin University of China, Beijing, Beijing Municipality, China

The aim of this study is to generate post-therapeutic optical coherence tomography (OCT) images based on pre-therapeutic OCT by generative adversarial networks (GANs). The synthetic images enable us to predict the short-term therapeutic efficacy of retinal vein occlusion (RVO) patients receiving intravitreal injection of anti-vascular endothelial growth factor (VEGF).The study involved patients with RVO who received intravitreal anti-VEGF injection from November 1, 2018 to November 30, 2019. The OCT images taken before and shortly after treatment, with an interval of 4-8 weeks, were collected and randomly divided into the training set and test set at the ratio of approximately 3:1. The model is constructed based on the pix2pixHD algorithm and synthetic OCT images are evaluated in terms of the picture quality, authenticity, the central retinal thickness (CRT), the maximal retinal thickness, the area of intraretinal cystoid fluid (IRC), the area of subretinal fluid (SRF).Three supporting models, including macular detection model, retinal stratification model and lesion detection model, were constructed. Segmentation of macular location, retinal structure and typical lesions were added to the input information. After verifying their accuracy, supporting models were used to detect CRT, the maximal retinal thickness, IRC area and SRF area of the synthetic OCT images. The output predictive values are compared with real data according to annotation on the real post-therapeutic OCT images.Results: 1140 pairs of pre-and post-therapeutic OCT images from 95 RVO eyes were included in the study and 374 images were annotated. 88% of the synthetic images were considered to be qualified. The accuracy of discrimination of real versus synthetic OCT images was 0.56 and 0.44 for two retinal specialists. The accuracy to predict treatment efficacy of CRT, the maximal retinal thickness, IRC area and SRF area were 0.70, 0.70, 0.92 and 0.78, respectively.Our study proves that GANs is a reliable tool to predict the therapeutic efficacy of anti-VEGF injections in RVO patients. Evaluations conducted both qualitatively and quantitatively showed that our model can generate high-quality posttherapeutic OCT images. Consequently, it has great potential in predicting treatment efficacy, providing guidance to clinical decision making.