November 23, 2020
Speaker: Gabriel della Maggiora. Professional collaborator in the project «Development of new imaging techniques to study the brain in severe mental health disorders» ACT192064. Director: Cristian Tejos. Talk summary: Generally, inverse problems do not have a unique solution that can be considered correct. The traditional approach to solving inverse problems through supervised learning usually involves some corruption process, which can be stochastic or deterministic, over a clean version of the image. This process means that in a problem where several predictions can be considered correct, only one is given as target. Thus, the supervised model learns only to predict a single image for each input. In this work we adress this by using a GAN model that learns the underlying distribution of the ill-posed problem. We show that by learning this distribution we can solve the undersampled image reconstruction in MRI problem with an undersampling factor of 16x. Furthermore, we show that the learned distribution can be used to solve a Compressed Sensing optimization problem in the latent space of the GAN instead of the pixel space yielding accurate reconstructions.