A Fast Generative Adversarial Network for High-Fidelity Optical Coherence Tomography Image Synthesis

(1) Background: We present a fast generative adversarial network (GAN) for generating high-fidelity optical coherence tomography (OCT) images. (2) Methods: We propose a novel Fourier-FastGAN (FOF-GAN) to produce OCT images. To improve the image quality of the synthetic images, a new discriminator with a Fourier attention block (FAB) and a new generator with fast Fourier transform (FFT) processes were redesigned. (3) Results: We synthesized normal, diabetic macular edema (DME), and drusen images from the Kermany dataset. When training with 2800 images with 50,000 epochs, our model used only 5 h on a single RTX 2080Ti GPU. Our synthetic images are realistic to recognize the retinal layers and pathological features. The synthetic images were evaluated by a VGG16 classifier and the Frechet inception distance (FID). The reliability of our model was also demonstrated in the few-shot learning with only 100 pictures. (4) Conclusions: Using a small computing budget and limited training data, our model exhibited good performance for generating OCT images with a 512 x 512 resolution in a few hours. Fast retinal OCT image synthesis is an aid for data augmentation medical applications of deep learning.

Automated summary

This study proposes a fast generative adversarial network (GAN) for generating high-fidelity optical coherence tomography (OCT) images. By redesigning the network architecture, the image quality of the synthetic images is improved, and high-resolution images can be generated in a short time with limited computing resources.