Meeting Banner
Abstract #2227

Learning cardiac morphology from MR images using a generative adversarial network: a proof of concept study

Davide Piccini1,2,3, Aurélien Maillot1,2, John Heerfordt1,2, Dimitri Van De Ville4,5, Juerg Schwitter6, Matthias Stuber2,7, Jonas Richiardi1,2, and Tobias Kober1,2,3
1Advanced Clinical Imaging Technology, Siemens Healthcare, Lausanne, Switzerland, 2Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, 3LTS5, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, 4Institute of Bioengineering/Center for Neuroprosthetics, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, 5Department of Radiology and Medical Informatics, University Hospital of Geneva (HUG), Geneva, Switzerland, 6Division of Cardiology and Cardiac MR Center, University Hospital of Lausanne (CHUV), Lausanne, Switzerland, 7Center for Biomedical Imaging (CIBM), Lausanne, Switzerland

Learning anatomical characteristics from large databases of radiological data could be leveraged to create realistic representations of a specific subject’s anatomy and to provide a personalized clinical assessment by comparison to the acquired data. Here, we extracted 2D patches containing the descending aorta from 297 3D whole-heart MRI acquisitions and trained a Wasserstein generative adversarial network with a gradient penalty term (WGAN-GP). We used the same network to generate realistic versions of the aortic region on masked real images using a loss function that combines a contextual and a perceptual term. Results were qualitatively assessed by an expert reader.

This abstract and the presentation materials are available to 2020 meeting attendees and eLibrary customers only; a login is required.

Join Here