Currently there is strong interest in data-driven approaches to medical image classification. However, medical imaging data is scarce, expensive, and fraught with legal concerns regarding patient privacy. Typical patient data consent forms only allow images to be used in medical journals or for education purposes, meaning the majority of medical data is not available for general public research. Synthetic medical images promise a solution to these problems. We propose a novel, two-stage pipeline for generating synthetic medical images from a pair of generative adversarial networks (GANs), which we test in practice on retinal fundi images. The first stage generates synthetic segmentation masks and the second stage converts the masks to photorealistic images. The images from Stage-II, along with their corresponding segmentations from Stage-I, are used to train a u-net segmentation network. On the u-net, our synthetic data pipeline received an F1 score of 0.8877, in comparison to a score of 0.8988 when tested with ground truth data, showing a negligible difference between synthetic and real patient datasets.

该研究提出了一种基于生成对抗网络的新型双阶段流水线，用于从医学图像数据中生成合成医学图像，并开发了SynthMed在线数据库以促进医学图像任务的发展。

双对抗生成网络生成的合成医学图像