Realistic AS-OCT images can be generated via machine learning

A study found that a generative adversarial network (GAN) could produce realistic high-resolution anterior segment optical coherence tomography images (AS-OCT) that could be suitable for machine learning and image analysis tasks.

The study included nearly 143,000 ASC-OCT scans from the American University of Beirut Medical Center. Researchers trained the Style and WAvelet based GAN to generate realistic AS-OCT images that were evaluated via the Fréchet Inception Distance (FID) Score and by 3 blinded refractive surgeons who were tasked with discerning between real and generated AS-OCT images.

The GAN-generated images demonstrated visual and quantitative similarity to real AS-OCT images, with an FID score of 6.32. Only half of the time (51.7%) the surgeons could tell the difference between the real versus generated images, which the authors noted was “not significantly better than chance” (P>.30).

A convolutional neural network (CNN) was trained to assess the suitability of the generated images for machine learning tasks, and the CNN accuracy improved from 78% to 100% when synthetic images were added to the dataset.

The GAN-generated images were “upsampled” using enhanced super-resolution GAN (ESRGAN) to achieve high resolution. These versions were “objectively more realistic and accurate compared with the traditional upsampling techniques by scoring a lower Learned Perceptual Image Patch Similarity of 0.0905 compared with 0.4244 of bicubic interpolation,” the authors reported.

Reference

Assaf JF, Mrad AA, Reinstein DZ, et al. Creating realistic anterior segment optical coherence tomography images using generative adversarial networks. Br J Ophthalmol. 2024:bjo-2023-324633. doi:10.1136/bjo-2023-324633