Datum predavanja: četvrtak, 12. 9. 2024.
Vrijeme predavanja: 11h

Abdullah Abdullah, The Chinese University of Hong Kong
Title: Novel numerical reconstruction and isolation of different nonlinear dynamics in video via latent space disentanglement of an untrained generator network and applications to dynamic MRI

Abstract:
Processing different types of dynamics in video data is a highly relevant problem in video analysis particularly in dynamic medical imaging where contrast enhancement, respiratory motion and patient moments poses a great challenge due to its effects on the image reconstruction as well as for its subsequent interpretation. The analysis and further processing of the dynamics of interest is often complicated by additional unwanted dynamics. This work proposes a novel nonlinear approach for the reconstruction and subsequent separation of different types of nonlinear dynamics in a video data via deep learning. The dynamic images are represented as the forward mapping of a sequence of time dependent latent space variables via an untrained generator neural network with no supervision. Different kinds of dynamics are characterized independently from each other via latent space disentanglement using one dimensional prior information, called triggers. Leveraging the triggers, the method successfully reconstruct a video containing different dynamics from highly underdampled data with parallel imaging. The model also detect the unknown dynamic and subsequently freeze any selection of dynamics and obtain accurate independent representations of the other dynamics of interest at any phase of the frozen dynamic. The method is tested on both synthetic data and real MRI datasets where contrast intensity, breathing, respiratory and body motion are separated.