Evaluation of Generative Adversarial Networks for High-Resolution Synthetic Image Generation of Circumpapillary Optical Coherence Tomography Images for Healthy Eyes

Abstract

Purpose : To evaluate the use of generative adversarial networks for the synthesis of high-resolution circumpapillary optical coherence tomography (OCT) images.

Methods : Cross-sectional circumpapillary OCT images of 1590 Chinese, 1040 Malay, and 1327 Indian healthy eyes from the Singapore Epidemiology of Eye Diseases (SEED) study were obtained from a commercial high resolution OCT system (Cirrus HD-OCT, Carl Zeiss Meditec) using an optic nerve head centered imaging protocol. These images were used to train ethnicity-specific Progressively Growing Generative Adversarial Network (PGGAN) models for the generation of synthetic images. In addition, another PGGAN model was trained using the images from all the three ethnicity datasets (a total of 3957 images). An evaluation dataset was constructed for each model using 25 generated synthetic images and an equal number of actual images from the corresponding training data. Two clinicians were asked to review the authenticity of the images in each dataset and were not provided with prior knowledge of the distributions.

Results : Examples of the generated images are shown in Figure 1. Each clinician reviewed the datasets independently. Between the two clinicians, the average accuracy was 49% for the Chinese evaluation dataset (Precision=49%, Recall=48%), 37% for the Malay evaluation dataset (Precision=31%, Recall=28%), 43% for the Indian evaluation dataset (Precision=28%, Recall =33%), and 37% for the combined evaluation dataset of all three ethnicities (Precision=27%, Recall=26%). There were differing opinions on 52% of the synthetic images generated from the Chinese, Malay, and Indian PGGAN models and 40% of the synthetic images from the combined ethnicity-based PGGAN model. Only 24%, 20%, 32%, and 28% of the images generated from the Chinese, Malay, Indian, and combined PGGAN models respectively, were correctly identified as synthetic images by both the clinicians.

Conclusions : Synthetic circumpapillary OCT images generated using PGGAN approach were difficult to discern from actual images. The results suggest a potential use of deep learning-based generative models for data generation and augmentation.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure1. Synthetic image samples generated for a) Chinese, b) Malay, c) Indian and d) combined ethnicities using PGGAN models are as shown.