Fast ungated five-dimensional cardiac MRI on a 1.5 T MR-linac for MRI-guided radiotherapy
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Fast ungated five-dimensional cardiac MRI on a 1.5 T MR-linac for MRI-guided radiotherapy".
Title and authors: Kai: So, we started by looking at the title and authors of this paper, "Fast ungated five dimensional cardiac MRI on a one point five T MR-linac for MRI-guided radiotherapy <ref:2607.16033#pg1,Fast ungated five dimensional cardiac MRI on a 1.5 T MR>." It immediately tells us the focus is on speed and clinical application in a real machine.
Mira: And I think that emphasizes how they’ve moved away from those long acquisition times that previously hampered 5D-MRI methods, positioning this as something truly viable for patient care <ref:2607.16033#pg1>.
Lev: From an engineering standpoint, the title suggests they are tackling the practical bottleneck of getting high-fidelity motion data fast enough to support real therapy decisions.
Kai: It’s about making sure that personalized motion characterization isn't just a theoretical exercise but something that can be delivered quickly enough in a clinical workflow.
Mira: The implication is that we can move closer to personalized treatment plans for cardiac conditions like ventricular tachycardia because the necessary motion data is available almost immediately after the scan.
Lev: It’s about translating complex mathematical modeling into something that runs on a real one point five T MR-linac without requiring hours of processing time between the scan and making a treatment decision <ref:2607.16033#pg1,1.5 T MR-linac>.
Kai: So, we're seeing a convergence here between advanced image reconstruction theory and the hard constraints of clinical timing for medical devices.
Mira: And that’s what this paper is really about: taking high-quality 5D imaging and making it fast enough to be clinically useful for targeted radiotherapy <ref:2607.16033#pg1>.
Lev: It’s about feasibility under real-world constraints, which is always where the biggest hurdles lie in hardware development.
The paper's summary: Kai: Moving into the summary of this paper, what they are essentially proposing is a fast, ungated five dimensional MRI reconstruction method that provides personalized motion characterization for MRI-guided STAR treatments.
Mira: To unpack that simply, they are taking complex cardiac and respiratory motions in a patient's heart scan and reconstructing a model of those motions quickly.
Lev: The method uses the CMR-MOTUS framework which employs a variable density three dimensional Cartesian acquisition with a joint optimization framework to reconstruct both a motion corrected reference image and low rank deformation vector fields or DVFs.
Kai: What’s key here is that they use this joint optimization framework to explicitly disentangle the respiratory and cardiac motion during the reconstruction process, which is what makes it fast.
Mira: They achieve this by exploiting the low rank structure of those DVFs, approximating them as a product of a spatial basis and a temporal basis.
Lev: This low rank approximation allows them to separate the dynamics because they know that respiratory and cardiac motions operate in different frequency ranges, so they can partition the motion components.
Kai: Then they use an acquisition strategy called Ordered Pseudo-RAdial or OPRA sampling trajectory for temporal imaging using balanced steady state free precession sequences.
Mira: And finally, they reconstruct the five dimensional DVFs by multiplying those spatial and temporal basis functions with binned navigator values derived from their acquisition.
Lev: The main implication here is that by framing the motion as low rank, they are making a mathematical assumption that allows them to bypass the need for very long non-Cartesian acquisitions, which is where other methods usually fail.
The paper's improvements: Kai: Now let’s talk about their specific improvements. One major point they highlight is using variable density three dimensional Cartesian acquisition with a joint optimization framework to reconstruct the reference image and low rank deformation vector fields.
Mira: That approach explicitly disentangles respiratory and cardiac motion during optimization, which is what enables this fast five dimensional MRI reconstruction we just talked about.
Lev: It’s significantly more data efficient and compressible than directly modeling motion resolved images, because they exploit that inherent compressibility of DVFs by reconstructing an explicit low rank factorization in space and time.
Kai: That means instead of trying to capture every single pixel movement individually, they are capturing a much smaller set of parameters that describe the deformation.
Mira: They also improve the reconstruction process by enabling retrospective selection of Bresp and Bcardiac with constant image quality, which is different from other methods that sort k-space into cardiorespiratory motion states before reconstructing.
Lev: So they can pick those motion components after the fact while keeping the image quality consistent, which avoids needing to re-sort or interpolate data repeatedly during reconstruction.
Kai: And on a practical note, their system achieves a total end to end latency of approximately seven minutes—one minute for acquisition and six minutes for reconstruction.
Mira: That speed is what makes it so much better than other methods, which often rely on longer non-Cartesian acquisitions and image based reconstruction techniques.
Conclusion: Kai: So to wrap up, the core message of this paper, "Fast ungated five dimensional cardiac MRI on a one point five T MR-linac for MRI-guided radiotherapy," is that they’ve developed a method for rapid high quality personalized motion characterization <ref:2607.16033#pg1,Fast ungated five dimensional cardiac MRI on a 1.5 T MR>.
Mira: They achieved this by combining a three dimensional Cartesian acquisition with low rank deformation vector fields that explicitly separate respiratory and cardiac motion.
Lev: The practical implication is that it’s feasible to obtain accurate five dimensional MRI with a total latency of about seven minutes, which shows strong potential for integration into MRI guided STAR treatments.
Kai: It proves that you can get the necessary personalized motion characterization without waiting hours for reconstruction, which directly supports the clinical goal of maximizing treatment efficiency.
Mira: And they did this through a method that is validated using both digital and physical phantoms, showing a left ventricle DICE coefficient of zero point nine six plus or minus zero point one in the physical phantom test.
Lev: One limitation they mention is that while this method is fast, it’s based on the assumption of the low rank structure of the DVFs holding true.
Kai: That means if the actual motion isn't perfectly low rank, you might see some degradation in accuracy, which is a necessary caveat for any reconstruction method.
Mira: So while they’ve shown great results on their ten volunteers and phantoms, it’s important to remember that this method relies on the assumption that those low rank deformation vector fields accurately describe the patient's movement.
University Medical Center Utrecht
physics.med-ph
Submitted: 2026-07-17
Updated: 2026-11-18
The gist: Fast ungated five-dimensional cardiac MRI on a 1.5 T MR-linac for MRI-guided radiotherapy is a novel method that develops and validates a fast…
Key concepts
- 5D-MRI
- Five-dimensional MRI captures not just spatial information (like 2D images) but also temporal information (motion over time). This is crucial for cardiac imaging because the heart is constantly moving. The goal here is to accurately map how the heart and breathing move simultaneously during a scan.
- Deformation Vector Fields (DVFs)
- DVFs are mathematical representations used to describe how every point in the 3D image changes over time due to motion. Instead of trying to reconstruct a full motion-resolved image, this method uses these fields, which are inherently more compact and easier to process, making the reconstruction much faster and more efficient.
- Low-Rank Motion Model
- The paper assumes that the complex motion (like breathing and heartbeats) can be approximated by a low-rank structure. This means the total motion description can be simplified into two separate, non-overlapping components—one for respiratory motion and one for cardiac motion—which allows the system to disentangle them effectively.
Terminology
Summary
Fast ungated five-dimensional cardiac MRI on a 1. p. 05 T MR-linac for MRI-guided radiotherapy is a novel method that develops and validates a fast, ungated 5D-MRI reconstruction technique to provide personalized motion characterization to support MRI-guided STAR treatments
Background and Motivation
Stereotactic arrhythmia radioablation (STAR) is currently limited by complex cardiorespiratory motion, necessitating patient-specific five-dimensional magnetic resonance imaging (5D-MRI) motion models to overcome this clinical challenge Current 5D-MRI methods are hampered by long acquisition and reconstruction times, which limits their clinical availability for STAR treatments Accurate 5D-MRI must be acquired and reconstructed within 10 minutes to maximize target accuracy, maintain patient comfort, and ensure high treatment efficiency
Methodology of CMR-MOTUS
The proposed method is based on the previously established cardiac MR-MOTUS framework, utilizing a variable-density 3D Cartesian acquisition with a joint optimization framework to reconstruct a motion-corrected reference image and low-rank deformation vector fields (DVFs) This approach exploits the low rank structure of the DVFs to explicitly disentangle respiratory and cardiac motion during optimization The method uses a low-rank motion model where the full DVFs are approximated as D ≈ Φ s × Φ t with spatial basis Φ ∈ R,01 and temporal basis ψ ∈ R%1 such that 3LR + TR ≤ 3TN To separate the motion dynamics, the low-rank structure is partitioned into respiratory and cardiac components by recognizing that their frequency ranges do not overlap
Acquisition and Reconstruction
The study leverages a Cartesian sampling pattern, specifically the Ordered Pseudo-RAdial (OPRA) sampling trajectory, which is a Cartesian strategy optimized for temporal imaging using balanced steady-state free precession (bSSFP) sequences The 5D CMR-MOTUS framework was implemented using the autodifferentiation capabilities of the PyTorch framework for joint stochastic optimization over 15 epochs The reconstruction involves binning the temporal navigators into Bresp and Bcardiac bins, and then constructing the 5D DVFs by multiplying Φ resp and Φ cardiac with the respective binned navigator values
Validation Results
Validation using digital and physical phantoms demonstrated accurate 5D-MRI reconstruction, achieving a left-ventricle DICE coefficient of 0.96 ± 0.01 in the physical phantom In the volunteer cohort, the 5D-MRI scans showed strong motion consistency when compared to standard 2D cine MRI, yielding a cardiac motion magnitude error of 0. p. 01 ± 0.9 mm and a respiratory motion magnitude error of 0. p. 12 ± 2.9 mm Crucially, the data were acquired in one minute and reconstructed in six minutes
Conclusion
The proposed 5D-MRI method enables rapid, high-quality, and personalized motion characterization, demonstrating potential for integration into MRI-guided STAR treatments The method combines a 3D Cartesian acquisition low-rank deformation vector fields that explicitly disentangle the respiratory and cardiac motion It is feasible to obtain accurate 5D-MRI with a total latency of approximately seven minutes (one minute acquisition and sixminute reconstruction) The method was extensively validated using phantoms and demonstrated accurate 5D-MRI reconstructions
Data Availability
The 3D k-space data, 5D reconstructions, 2D cine imaging, and manual segmentations for the ten volunteers are publicly available at https://doi.org/10.5281/zenodo. p. 1 p. 121278894 The reconstructions and k-space are publicly available at https://doi. p. 1org/10.5281/zenodo.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Variable density 3D Cartesian acquisition with a joint optimization framework allows for reconstruction of a motion-corrected reference image and low-rank deformation vector fields (DVFs). This explicitly disentangles respiratory and cardiac motion during optimization, which enables fast 5D-MRI reconstruction.
-
Explicitly modeling motion using deformation vector fields is significantly more data efficient and compressible than directly modeling motion-resolved images. This exploits the
inherent compressibility of DVFs by reconstructing an explicit low-rank factorization in space and time.
-
The method enables retrospective selection of Bresp and Bcardiac with constant image quality, unlike methods that sort k-space into cardiorespiratory motion states prior to reconstruction. This allows for
retrospective selection of Bresp and Bcardiac with constant image quality.
-
The system achieves a total end-to-end latency of approximately seven minutes (one minute acquisition and six minute reconstruction), which is
significantly faster than other methods, which rely on longer non-Cartesian acquisitions and image-based reconstruction.
-
The 5D CMR-MOTUS framework can be combined with, for example, a non-Cartesian acquisition, though Cartesian sampling provides a
good trade-off between reduced computational cost, temporal incoherency, and greater resilience to system imperfections.
Sources
- Feasibility of Continuous Ventricular Volumetric Quantification in Arrhythmias using Real-Time 3D CMR-MOTUS
- Technical Report (v1.0)--Pseudo-random Cartesian Sampling for Dynamic MRI
- Adam: A Method for Stochastic Optimization
- nnInteractive: Redefining 3D Promptable Segmentation
Related papers
- BART Online Open-Source Sequence Toolbox for Computational MRI
- Constrained estimation of rotational invariants of the cumulant expansion (RICE) for rapid tensor-valued diffusion MRI
- Improved Myelin Water Fraction Estimation and Uncertainty Mapping in Multicomponent T2 Relaxometry via NNLS Bootstrap Aggregation
- Radiation damage to normal mammalian tissue in vivo with laser-driven protons at ultra-high instantaneous dose rate
- BEAM3R: Beam's-eye-view architecture with Mamba-3 for implicit dose reconstruction
- MORSE-PI -- Flexible and artefact-free image reconstruction for structural and functional QSM and other phase-critical imaging applications