A GAN-based with expert-validated data augmentation method for wireless capsule endoscopy images of small intestine polyp

Wireless capsule endoscopy (WCE) is a noninvasive method for examining the entire small intestine. An automatic polyp segmentation system can assist physicians in diagnosing polyps and accurately accessing lesions, improve clinical performance, and reduce expert time. However, collecting enough polyp images with WCE to train a deep learning model is challenging. Many data augmentation techniques were commonly used to address the problem of insufficient data. However, these techniques may not introduce enough diversity and generality to the training dataset. We introduce an expert-validated seamless cloning algorithm and a GAN-based refinement method to generate synthetic WCE polyp images. We then built an efficient small intestine polyp segmentation model using these synthetic data and the transfer learning with the pretrained weights from a colon polyp dataset. Our synthetic data closely resemble real polyps; experts have difficulty distinguishing between the real and synthetic images. The proposed small intestine polyp segmentation model in WCE images achieved a Dice coefficient of 0.89 for pixel level, precision of 0.9, and recall of 0.88 for polyp level. In this paper, we introduced a feasible method to expand a small dataset by generating synthetic data, which boosts the data quantity and diversity, thus improving the polyp segmentation model’s performance and enhancing generalization.