This study investigates an approach to generate a realistic, heterogeneous database of simulated cardiac MR images to aid the development of fully automated and generalizable deep learning based segmentation algorithms, less sensitive to variability in CMR image appearance. XCAT phantoms were used to create the virtual population by altering the heart position and geometry and MRXCAT approach was improved to simulate more organs. Images simulated in this study were quantitatively and qualitatively comparable to real CMR images acquired by two different sites and vendors. Initial experiments using such a heterogeneous image dataset show a positive impact on the segmentation performance.
This abstract and the presentation materials are available to members only; a login is required.