Synthetic Data Generation of Microbial Keratitis Slit Lamp Photos Using Limited Data

Abstract

Purpose : Artificial Intelligence (AI) has shown promise for screening eye conditions like microbial keratitis (MK), a top five cause of blindness worldwide, in non-ophthalmology settings. However, prior screening tools suffered from having trained on limited data. Recent advances in generative adversarial networks (GANs) can train on limited data to create synthetic data to supplement limited datasets. Prior literature focused on synthetic retinal fundus and OCT images. Here, we develop a novel synthetic slit lamp photography (SLP) model with a limited dataset.

Methods : Our GAN was trained on 133 SLPs from a limited public dataset of eyes with MK to generate synthetic SLPs of eyes with MK using StyleGAN2-ADA (an adaptive discriminator augmentation GAN shown to perform well with limited data). We also used transfer learning from a model previously trained on the Flickr-Faces-HQ dataset. To assess synthetic image quality, we performed a visual Turing test. Four ophthalmologists tested their ability to distinguish 25 real images and 25 of our non-curated synthetic images. In addition, we used the widely adopted Kernel inception distance (KID) for limited datasets to measure realism and variation of synthetic images.

Results : Expert reviewers had varied abilities to distinguish non-curated synthetic SLPs from real ones with Fleiss’ free-marginal kappa of 0.54 (95% CI 0.39-0.69) showing intermediate agreement. Among four experts, accuracies ranged from 80%-92% with sensitivities 68%-100% and specificities 68%-100%. The KID curve during training is shown in Figure 1. The best KID score achieved was 0.0179.

Conclusions : Recent developments in GAN training with limited data have improved the ability to develop synthetic datasets using limited medical data. In our resultant non-curated synthetic slit lamp photos, experts had varied abilities to distinguish synthetic from real images suggesting the potential utility of using synthetic images to supplement training data

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. A) Kernal inception distance during training (lower KID corresponds to realism). B) Random sample of non-curated synthetic and real microbial keratitis images.