Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back

Rapid Fire Abstracts

Rapid Fire 09- Image Reconstruction (including machine learning) II

Self-supervised motion-corrected reconstruction for single heartbeat cardiac cine MRI using Neural Fields and Deep Image Prior (RF_SA_449)

Saturday, February 1, 2025
9:50 AM – 10:00 AM East Coast USA Time
Location: Blue Room PreFunction

    Presenting Author(s)

  • TC

    Tabita A. Catalán, MSc

    Predoctoral Researcher
    Millennium Institute for Intelligent Healthcare Engineering iHEALTH, Chile

    • Primary Author(s)

  • TC

    Tabita A. Catalán, MSc

    Predoctoral Researcher
    Millennium Institute for Intelligent Healthcare Engineering iHEALTH, Chile

    • Co-Author(s)

  • Rene Michael M Botnar, PhD

    Director and Professor
    Institute for Biological and Medical Engineering
    UC Chile, Chile

  • FS

    Franscisco Sahli Costabal

    Pontifical Catholic University of Chile, Chile

  • Claudia Prieto, PhD

    Professor and Director for Research and Innovation
    School of Engineering, Pontificia Universidad Católica de Chile, Chile

Background: 2D cardiac cine is the gold standard for cardiac functional assessment. Recent works have successfully incorporated cardiac motion correction (MC) in the reconstruction, demonstrating the feasibility of single heartbeat imaging for both classical (MC-CINE1) and supervised deep learning2. MC-CINE requires long reconstruction times, while supervised approaches require large databases, which are not always available. Self-supervised approaches use untrained neural networks architectures as implicit regularizers, and train on the undersampled data of each subject (e.g. Time Depend Deep Image Prior (DIP)3 and neural fields (NF)-cMRI4). Physics informed neural networks with NF have been proposed for non-rigid image registration (WarpPINNs5). We propose MC-DIP, a self-supervised reconstruction approach that combines motion estimation with WarpPINNs and motion corrected TD-DIP undersampled reconstruction.

Methods: Reconstruction is done in two stages (Fig.1). First, cardiac displacement fields between cardiac phases are modelled as NF and calculated from a preliminary GRASP6 reconstruction using an extension of WarpPINNs to perform groupwise registration for all pairs of cardiac phases simultaneously. The estimated motion is incorporated in a TD-DIP reconstruction, using MapNet with a circular encoding. The estimated motion is incorporated in the forward model using the MC matrix formulism7. MC-DIP was evaluated in single heartbeat data (465 spokes) from a mid-ventricular short axis slice for a healthy subject at 1.5 T (Ingenia, Philips, Best, The Netherlands) using a 28-channel cardiac coil; field of view (FOV) = 256 x 256 mm2; 8 mm slice thickness; resolution = 2x2mm2; TE/TR = 1.16/2.3 ms; b-SSFP readout; radial tiny golden angle of 23.6, flip angle 60, 31 cardiac phases. MC-DIP was compared against GRASP and MC-CINE for 26-fold undersampling. Self-supervised methods are trained for 7k iterations, with learning rate 2e-3 and decay 0.8 after 1k iterations (7min).

 

Results:

MC-DIP is compared against GRASP and MC-DIP in diastole and for vertical and horizontal temporal profiles in Fig.2. Fig.3 shows MC-DIP reconstruction for different cardiac phases. MC-DIP exhibits less blurring than GRASP and comparable image quality to MC-CINE in a fraction of the time (16 min against >4 hours).

Conclusion: The proposed MC-DIP enables self7-supervised motion estimation with neural fields and motion corrected reconstruction with DIP, providing results comparable to MC-CINE with a reduced reconstruction time. Further work will incorporate motion estimation and motion corrected reconstruction in an end-to-end unrolled approach.