Generative Adversarial Network "Steerability" for Brain PET Image Generation

An investigation was carried out in which the latent space distribution of a Wasserstein generative adversarial network was varied to observe whether outputs of a particular image classification could be specified. The WGAN was trained using 18F FDG Brain PET image scans of 96 patients with varying degrees / classes of neurodegeneration before being initialized with a chosen probability distribution. At first, the mean of a normal Gaussian distribution with a standard deviation of 1 was varied between -1 and +1 before a beta distribution was utilized with varying parameters. When α = β = 0.5, it was found that the generator was more likely to produce brain images with less neurodegeneracy compared to the normal Gaussian. Initiating the generator with a beta distribution of parameters α = 1, β = 5 gave a higher average EMCI SSIM score over five generative iterations.