Low-Field Interventional Cardiac Magnetic Resonance Imaging Through 3D Radial “Beating and Breathing ” Roadmap Imaging Using Multi-Dynamic Low-Rank Deep Image Priors (ML-DIP)

Presenting Author(s)

• ST

  Samuel T Ting, PhD

  Assistant Professor
  The Ohio State University
  Columbus, Ohio, United States

Primary Author(s)

• JW

  Jianli Wei, PhD

  Research Scientist
  The Ohio State University, United States

Co-Author(s)

• XS

  Xavier Sieber, MSc

  PhD student
  University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland
• 

  Jérôme Yerly, PhD

  Senior scientist
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland
• YL

  Yingmin Liu, PhD

  Research Engineer
  The Ohio State University
  Columbus, Ohio, United States
• YD

  Yu Ding, PhD

  Research Scientist
  The Ohio State University
  Columbus, Ohio, United States
• KB

  Katherine Binzel, PhD

  Research Scientist
  The Ohio State University
  Columbus, Ohio, United States
• JV

  Juliet Varghese, PhD

  Assistant Professor
  The Ohio State University
  Columbus, Ohio, United States
• 

  Ruud B. van Heeswijk, PhD, FSCMR

  Senior lecturer
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland
• MS

  Matthias Stuber, PhD

  Professor/Director
  CIBM/CHUV/UNIL
  Lausanne, Switzerland
• 

  Orlando P. Simonetti, PhD, FSCMR

  Professor
  The Ohio State University
  Columbus, Ohio, United States
• 

  Rizwan Ahmad, PhD

  Professor
  The Ohio State University
  Columbus, Ohio, United States

Background: Interventional Cardiac Magnetic Resonance (iCMR) uses MRI to provide noninvasive guidance during cardiovascular procedures. Low-field MRI systems (0.55 T) allow the safe use of conventional catheters and guidewires and provide a large bore size (80 cm). However, current implementations primarily rely on 2D imaging with a limited number of predefined planes, which reduces robustness in device tracking during interventions. To address this limitation, we propose a precomputed 3D CardioRespiratory-Motion-resolved (CRM) roadmap for iCMR, followed by robust device tracking. In this abstract, we focus on the roadmap component.

Methods: A precomputed 3D cine roadmap provides soft-tissue contrast and free-breathing motion information, while the subsequent real-time scan focuses on robust device tracking. To achieve a high-quality roadmap, we adopted ML-DIP [1], a 3D cine reconstruction framework originally developed for Cartesian sampling. However, because radial sampling provides denser coverage of central k-space—where cardiac and respiratory motion information is concentrated—we replaced Cartesian sampling with a 3D phyllotaxis radial trajectory. This adaptation allowed us to incorporate a temporal flow consistency penalty specifically optimized for radial data, thereby enhancing temporal fidelity in the reconstructed cine roadmap.

On a 0.55T scanner (MAGNETOM FreeMax, Siemens Healthcare, Forchheim, Germany), a healthy volunteer is scanned under free-breathing conditions post infusion of 4 mg/kg ferumoxytol to enhance blood-pool contrast. Data acquisition includes 17 radial spokes/frame with one superior–inferior (SI) self-gating line repeatedly which yields TR of 4.6 ms per line, a temporal resolution of ~83 ms per frame. The imaging field of view (FOV) was isotropic at 220 mm with a matrix size of 224 × 224 ×224, corresponding to a spatial resolution of ~0.98 mm³. It was trained for 14,000 iterations, requiring ~11.4 hours on a single GPU (NVIDIA A100, 40 GB).

Results: Figure 1 shows three orthogonal views (xy, xz, and yz) across time t, where soft tissue structures with both cardiac and respiratory motion are clearly and smoothly visualized. Figure 2 presents reformatted long-axis view, and short axis view over time t. Figure 3 illustrates the temporal evolution along the red cut-off line, where both cardiac and respiratory waveforms are distinctly delineated.

Conclusion: Unlike iCMR real-time imaging, this roadmap can be generated separately—potentially one day prior to the intervention—thus decoupling the computationally intensive reconstruction process from the procedure itself. Real-time imaging during iCMR can focus on device tracking, supported by image registration as well as object detection and tracking algorithms.

Figure 1. The three rows display orthogonal slices (xy, xz, yz) of the reconstructed cine frames over one cardiac cycle. Each frame was acquired with 17 radial spokes (4.6 ms per spoke), yielding a temporal resolution of 12 frames per second. Only every other frame is shown here for visualization. Red arrows on the images highlight excellent blood-myocardium contrast and sharp myocardium wall boundaries.
Figure 2. Reformatted long-axis views (first two rows) and short-axis view (third row) over one cardiac cycle, shown at the same temporal resolution as in Figure 1 (only 6 of the 12 reconstructed frames are shown). The reconstructed cine frames clearly demonstrate soft tissue contrast, including visualization of the papillary muscles and myocardium, while capturing free-breathing cardiac and respiratory motion.
Figure 3. Spatiotemporal profile along the red cut line across 250 frames. Both cardiac motion and respiratory motion are clearly visible in healthy volunteer. The data were acquired under free-breathing conditions, resulting in smooth and regular cardiac and respiratory waveforms.