| Description |
- Background: Artificial Intelligence (AI) models, like those used in diabetic retinopathy screening, require large amounts of data to train. Models for diseases with limited public data like microbial keratitis (MK), a top five cause of blindness worldwide, are therefore difficult to train. Generative adversarial networks (GANs) are AI models that train using limited data to create synthetic images. We develop a novel slit lamp photography (SLP) GAN with limited data to improve the performance of an MK screening model.
Methods: We trained StyleGAN2-ADA on 57 healthy and 27 MK SLPs to generate synthetic images. To assess synthetic image quality, we performed a visual Turing test. Three ophthalmologists tested their ability to identify 20 images each of 1) real healthy, 2) real diseased, 3) synthetic healthy, and 4) synthetic diseased. We used the widely adopted Kernel inception distance (KID) for limited datasets to measure realism and variation of synthetic images. We then trained two DenseNet121 models to grade images as healthy or MK with 1) only real images and 2) real supplemented with GAN-generated images. To assess model performance, we report performance on 10 healthy and 9 MK real images not used in either training.
Results: Cornea experts on average rated synthetic images as good quality (83.3% ± 14.6). Synthetic and real images depicted pertinent anatomy and pathology for accurate classification (96.3% ± 2.19). Despite limited data, experts rated some synthetic images as real. The best KID score achieved was 0.03 for healthy images and 0.02 for diseased images. The real and synthetic data model (AUROC: 0.93, sensitivity: 88.9%, specificity: 100%) outperformed the model trained on only limited real data (AUROC: 0.76, sensitivity: 88.9%, specificity: 80%).
Conclusions: MK classification was improved by AI supplementation of limited real training data with synthetic data generated by GANs.
|
|---|
