A Novel Conditional Medical Image Generation GAN with Latent Space Optimal Transport

Generative adversarial network (GAN) is widely used to augment training set in medical imaging. However, training GANs can be challenging, as they are susceptible to mode collapse. These challenges stem from the fact that deep neural networks can only represent continuous mappings, while in GAN, the generator aims to compute discontinuous transportation maps between the noise distribution and the data distribution. We propose a novel approach that can overcome mode collapse, and in turn, lead to improved capability by GAN to generate specific classes of medical images. Our method starts by mapping the data manifold to the latent space using an autoencoder (AE). Subsequently, we encode the sample labels and integrate them with the latent representations. Next, we employ extended semi-discrete optimal transport (SDOT) mapping, which maps a Gaussian distribution to the empirical latent distribution, thereby generating new latent representations with known labels. Finally, we employ GAN to establish the mapping from the continuous latent distribution induced by the extended SDOT mapping to the real data distribution, generating high-quality images. We conducted extensive experiments on the DermaMNIST and BloodMNIST datasets. The experimental results highlight the exceptional performance of our model in generating images belonging to specified classes.