GAN-Based Realistic Gastrointestinal Polyp Image Synthesis

Polyps in the gastrointestinal (GI) tract in the human body are one of the most significant symptoms of gastric and colorectal cancer and some other diseases. This paper proposes Generative Adversarial Networks (GANs) based methods that first use a StyleGAN2-ada to generate random polyp masks, which are used to create composite images with healthy GI images. Then a conditional GAN is used to translate these composite images into synthetic polyp images. The proposed approach can produce a high amount of realistic GI polyp images and can increase F1-score and IoU in polyp detection by around 4% when used in the training phase of the YOLOv4 object detector.