Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI

arXiv:2611.05169 · eess.IV · Submitted 2026-11-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI".

Dev: Here's the gist of the paper. Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI develops an accelerated Zero Echo Time (ZTE) MRI reconstruction framework that jointly estimates the 3D image and rigid-motion parameters, while compensating for missing k-space center data and angular undersampling artifacts within a regularized maximum a posteriori (MAP) formulation<ref:2611.05169#pg2>.

Rosa: The proposed method improves both PSNR and SSIM relative to CG-SENSE and Autofocusing (e.g., ∼10– 15 dB higher PSNR for gap ≥3.5), and reduces rotational motion error in accelerated experiments<ref:2611.05169#pg9>.

Dev: Framework Formulation The method is formulated as a MAP estimation termed Energy-based Motion Correction (EnMoCo), which is solved by alternating between k-space rigid-motion updates (trajectory rotation and phase modulation) and conjugate-gradient data-consistency updates<ref:2611.05169#pg4>.

Rosa: The joint recovery minimizes the cost function L(x; ξ) = 1/2σ2 y − Aξx2 + Rw(x), where Rw: CMx×My×Mz → R is an image domain regularizer, represented by the 3D MuSE<ref:2611.05169#pg6>.

Title and authors: Rosa: We started by looking at the title and authors of this paper, Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI. It’s a pretty mouthful, but it tells you exactly what the paper is aiming to achieve.

Dev: Yeah, it points toward solving two major issues: motion correction and image reconstruction in one go for ZTE MRI that needs to be accelerated acquisition. It’s not just about fixing one part of the problem anymore; it’s about making them happen together.

Taro: So what is the big picture here? Is this paper trying to solve a fundamental limitation in how we do ZTE imaging, or is it more of an incremental improvement on existing techniques?

Rosa: It's aiming for a unified framework, which means they want to jointly estimate the image and the rigid motion parameters simultaneously. They are specifically targeting ZTE MRI because it has those inherent challenges like missing k-space center data and artifacts from angular undersampling.

Dev: So if you break that down, what does that mean for someone running a scan? They don't have to run a separate motion estimation step and then plug the results into another reconstruction algorithm.

Taro: That sounds like it simplifies the workflow significantly, but I wonder how much complexity is hidden inside this unified optimization scheme?

Rosa: The paper describes solving this as a regularized maximum a posteriori estimation problem where they are solving for both the image and those nuisance variables, which include rigid motion parameters and unmeasured k-space samples.

Dev: And they use an energy based model called MuSE for the regularization, which provides an explicit energy defined by a compact convolutional neural network that is trained via multiscale denoising score matching. That’s their way of building the prior.

Taro: So this MuSE model isn't just some plug and play denoiser from a standard SENSE based reconstruction method, right?

Rosa: No, the paper makes it clear that the MuSE prior is algorithmically unrelated to MUSE or AMUSE48 which are SENSE based reconstruction methods for multishot diffusion MRI. The explicit prior is what sets this apart and allows them to use optimization algorithms with convergence properties that are already established under standard assumptions.

Dev: And they use a majorization minimization scheme, or MM scheme, to solve the cost function in a way that alternates between updating the k-space rigid motion and doing conjugate gradient data consistency updates. That’s how they tackle the math.

Taro: Alternating steps sound like a lot of computational work running in sequence. What’s the latency like when you're trying to get those motion updates done?

Rosa: They are solving for those motion parameters and unmeasured k-space samples at the same time during the image reconstruction of that volume. This joint recovery happens simultaneously with the image estimation.

The paper's summary: Dev: We’ve touched on the setup, but now let’s look at what they actually found in this paper called Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI. They summarized their overall approach to solving the ZTE reconstruction problem.

Rosa: Essentially, they introduced this unified optimization scheme as a solution to solve for the image jointly with those nuisance variables, which are the rigid motion parameters and those unmeasured k-space samples, all within a single optimization scheme.

Taro: So what is the practical payoff for someone who is using ZTE MRI that struggles with missing data at the center of their k-space? Does this actually help them see better structures?

Dev: The payoff is that they are compensating for both the central gap and those undersampled locations, which means they are not ignoring those missing samples. This gives you a more complete picture compared to just throwing away that data.

Rosa: Right, and in terms of the regularization, they use this MuSE prior which provides an explicit energy parameterized by a compact convolutional neural network trained via multiscale denoising score matching. That’s their explicit way of guiding the reconstruction toward a good solution.

Taro: So, what's the practical advantage over just using standard denoisers that are already available? Is it something they are introducing new knowledge there?

Dev: The key distinction is that this MuSE prior is algorithmically unrelated to MUSE or AMUSE48 which are SENSE based reconstruction methods for multishot diffusion MRI. It’s a fundamentally different way of defining the energy function.

Rosa: And because it’s an explicit prior, it allows them to use optimization algorithms with convergence properties that are already established under standard assumptions, which makes the entire process more tractable than some other posterior sampling methods.

Taro: So, if we were to try this out, is there a specific scenario where you’d expect this to work exceptionally well?

Dev: They emphasize that it’s typically less computationally demanding than diffusion posterior sampling methods, which makes it more feasible for using in three dimensional settings. This tractability is a big win for practical application.

The paper's improvements: Rosa: Now let’s talk about the specific improvements this paper suggests they achieved. What are the tangible benefits they claim?

Dev: They show significant gains in image quality, claiming that this method improves both PSNR and SSIM relative to CG-SENSE and Autofocusing. Specifically, for a gap of three point five samples or more, they see about ten to fifteen decibels higher PSNR.

Taro: Ten to fifteen decibels is a big number in terms of image quality metrics. Does that translate into something noticeable when you look at actual tissue contrast?

Rosa: It translates to reduced motion blur and inconsistencies in the tissue contrast compared to those baselines when the dead time gaps are between three and four point five samples. They also maintain high fidelity across gap sizes up to four for both acceleration factors, showing superior robustness under high acceleration.

Dev: And they are also tackling artifact suppression consistently for larger dead time gaps of three to four point five samples by combining motion estimation with an energy based denoiser. This is particularly useful when you’ve got those longer gaps where artifacts tend to be more pronounced.

Taro: So if I’m looking at the results, am I seeing less noise and distortion in areas that were affected by movement?

Rosa: Yes, they show reduced motion blur and better tissue contrast in those problematic regions when you compare them to the baselines across larger dead time gaps of three to four point five samples.

Dev: And one specific improvement they highlight is the multi resolution search for rotation parameters, which gives a lower l1 error for those rotation parameters. That means more precise motion tracking during the process.

Conclusion: Rosa: So, to wrap this up on Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI. This paper introduces a framework that jointly estimates the image and the rigid motion parameters along with unmeasured k-space samples using alternating updates between k-space motion updates and conjugate gradient data consistency steps.

Dev: Ultimately, they showed that this approach works by demonstrating superior performance over established methods like CG-SENSE and Autofocusing, especially when dealing with those larger gaps of three point five samples or more where the signal is still good enough for these gains.

Taro: I’m just thinking about how this impacts the field as a whole. It suggests that we can start treating motion correction not as a separate problem but as something inherently tied to the reconstruction itself.

Rosa: Exactly, it means we can build systems that are more robust across different hardware setups and more tolerant of those challenging acquisition conditions in real-world settings.

Dev: And this paper provides a clear roadmap for how to integrate motion estimation into the reconstruction loop efficiently while maintaining good convergence properties under standard assumptions.

Taro: It’s a solid contribution to making sure that when we push for speed in ZTE MRI, we don't sacrifice the image quality we need.

Rosa: That’s what this work on Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI shows us about getting better results in fast three dee imaging <ref:2611.05169#pg1,Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI>.

Yan Chen, Curtis A. Corum, James H. Holmes, Mathews Jacob

Electrical and Computer Engineering, University of Virginia · Champaign Imaging LLC

eess.IV

Submitted: 2026-11-03

Updated: 2026-11-03

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI develops an accelerated Zero Echo Time (ZTE) MRI reconstruction framework that jointly estimates the 3D image and…

Key concepts

Energy-based Motion Correction (EnMoCo)
This is the core framework that estimates both the image and motion parameters simultaneously. It works by repeatedly updating the rigid-motion parameters (like rotation) and then refining the image reconstruction using data consistency steps, guided by an energy minimization cost function.
MuSE Prior
MuSE is a pre-trained model used as a regularization strategy to guide the image reconstruction process. It is implemented in a 2.5D form by combining energies from two orthogonal views, making it data-efficient and helping to suppress artifacts in the reconstructed image.
Alternating Minimization Scheme
The optimization process involves switching between two main tasks: first, updating the motion parameters (using methods like steepest descent), and second, updating the image volume (using Conjugate Gradient). This alternating approach allows for joint recovery of both variables.
Multi-resolution MoCo Prior
This technique gradually increases the resolution of a k-space mask during optimization. It starts with a coarse mask and expands it to include higher spatial frequencies as the process continues, which helps in recovering missing k-space data more effectively.

Terminology

Summary

Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI develops an accelerated Zero Echo Time (ZTE) MRI reconstruction framework that jointly estimates the 3D image and rigid-motion parameters, while compensating for missing k-space center data and angular undersampling artifacts within a regularized maximum a posteriori (MAP) formulation<ref:2611. p. 0

Where the paper says it

Purpose: To develop an accelerated Zero Echo Time (ZTE) MRI reconstruction framework that jointly estimates the 3D image and rigid-motion parameters, while compensating for missing k-space center data and angular undersampling artifacts within a regularized maximum a posteriori (MAP) formulation. Methods: We formulate the joint recovery as a MAP estimation, termed Energy-based Motion Correction (EnMoCo), which is solved by alternating between k-space rigid-motion updates (trajectory rotation and phase modulation) and conjugate-gradient data-consistency updates.

05169#pg2>. The proposed method improves both PSNR and SSIM relative to CG-SENSE and Autofocusing (e. p. 0
Where the paper says it

5–5 samples, rotations up to ±17◦ , translations up to ±10 mm, and a range of scan times (153 to 611 seconds). In simulations, the proposed method improves both PSNR and SSIM relative to CG-SENSE and Autofocusing (e.

g., ∼10– 15 dB higher PSNR for gap ≥3. p. 0
Where the paper says it

In simulations, the proposed method improves both PSNR and SSIM relative to CG-SENSE and Autofocusing (e. , ∼10– 15 dB higher PSNR for gap ≥3.

5), and reduces rotational motion error in accelerated experiments<ref:2611. p. 0
Where the paper says it

, ∼10– 15 dB higher PSNR for gap ≥3. 5), and reduces rotational motion error in accelerated experiments.

05169#pg9>.

Framework Formulation

The method is formulated as a MAP estimation termed Energy-based Motion Correction (EnMoCo), which is solved by alternating between k-space rigid-motion updates (trajectory rotation and phase modulation) and conjugate-gradient data-consistency updates<ref:2611.05169#pg4>. The joint recovery minimizes the cost function L(x; ξ) = 1/2σ2 y − Aξx2 + Rw(x), where Rw: CMx×My×Mz → R is an image domain regularizer, represented by the 3D MuSE<ref:2611. p. 2

Where the paper says it

The above properties allow us to account for motion exclusively in k-space, without having to generate intermediate images. 2 Joint Recovery of Motion & Image We propose to jointly estimate the image x and the 3D motion parameters ξ by minimizing the cost function L(x; ξ): L(x; ξ) = 1 2σ2 ∥y − Aξx∥ 2 2 + Rw(x), (5) where Rw : CMx×My×Mz → R is an image domain regularizer, represented by the 3D MuSE38, and σ 2 is the variance of the measurement noise.

05169#pg6>. The motion parameters ξ are estimated simultaneously during the reconstruction of the image volume x<ref:2611. p. 2
Where the paper says it

5D strategy, where the 3D prior is the sum of the MuSE energies of the 2D slices. The motion parameters ξ are estimated simultaneously during the reconstruction of the image volume x.

05169#pg6>.

Regularization Strategy

The regularization is provided by a pre-trained Multi-Scale Energy-based model (MuSE), implemented in a data-efficient 2.5D form by combining slicewise energies from two orthogonal views to reduce 3D training demands<ref:2611. p. 0

Where the paper says it

Regularization is provided by a pre-trained Multi-Scale Energy-based model (MuSE), implemented in a data-efficient 2. 5D form by combining slicewise energies from two orthogonal views to reduce 3D training demands.

05169#pg4>. The MuSE prior is algorithmically unrelated to MUSE (Multiplexed Sensitivity-Encoding) and AMUSE48, which are SENSE-based reconstruction methods for multishot diffusion MRI<ref:2611. p. 2
Where the paper says it

For regularization, we use the Multi-Scale Energy-based model (MuSE) 38 , which provides an explicit energy (negative log-prior) parameterized by a compact CNN trained via multiscale denoising score matching. The MuSE prior used in this work is algorithmically unrelated to MUSE (Multiplexed Sensitivity-Encoding 47) and AMUSE48, which are SENSE-based reconstruction methods for multishot diffusion MRI.

05169#pg4>. The 2.5D construction approximates the 3D prior as a sum of slice-wise energies computed from two orthogonal views to mitigate view-dependent artifacts<ref:2611. p. 8
Where the paper says it

To reduce 3D training demands, we use a data-efficient 2. 5D construction that aggregates slice-wise energies from two orthogonal views, aiming to mitigate view-dependent artifacts while remaining practical in limited-data settings.

05169#pg5>.

Optimization Scheme

The optimization uses a Majorizationminimization (MM) scheme to solve the cost function in Eq. (5)<ref:2611. p. 0

Where the paper says it

Conclusion: A unified MAP formulation with k-space motion modeling and an explicit energy-based prior can improve reconstruction quality for motion-corrupted, accelerated ZTE relative to standard reconstruction and motion-correction baselines evaluated in this work. KEYWORDS: ZTE, motion correction, dead time gap, energy model WORD COUNT: 5000 arXiv:2611.

05169#pg4>. This MM reformulates the optimization into an alternating minimization scheme, which alternates between motion update using steepest descent and image update using the Conjugate Gradient (CG) algorithm<ref:2611. p. 3
Where the paper says it

{x n+1 , ξ n+1} = arg min x,ξ 1 2σ 2 ∥y − Aξx∥ 2 2 + Rw(x n ) + ∇xRw(x n )(x − x n ) + L 2 ∥x − x n ∥ 2 2 , (11) where L is the Lipschitz constant of the score network, and n is the index of the iteration. MM reformulates the optimization into an alternating minimization scheme, which alternates between motion update using steepest descent: ξ n+1 = ξ n − α∇ξ 1 2σ2 ∥y − Aξ n x n ∥ 2 2 , (12) and image update using the Conjugate Gradient (CG) algorithm: x n+1 = AH ξ n+1Aξ n+1 σ 2 + LI !

05169#pg4>. The algorithm implicitly recovers missing samples in the center k-space through this alternation<ref:2611. p. 3
Where the paper says it

−1 1 σ 2 AH ξ n+1 y + Lxˆ n , (13) where xˆ n is the denoised version of x n: xˆ n = x n − 1 L ∇xRw(x n ) , (14) obtained by subtracting the MuSE score. The algorithm implicitly recovers missing samples in the center k-space.

05169#pg4>. A multi-resolution MoCo prior is introduced, where a k-space mask Ms gradually expands to include higher spatial frequencies as the optimization progresses<ref:2611. p. 3
Where the paper says it

We use a coarse-to-fine (multiresolution) motion-estimation strategy. Specifically, we introduce a k-space mask Ms that initially retains only a central low-resolution region, which gradually expands to include higher spatial frequencies as the optimization progresses.

05169#pg4>.

Performance Evaluation

The method demonstrates effectiveness on in vivo 3D ZTE data and simulated data to replicate more challenging hardware scenarios, including dead time gaps spanning 1.5–5 samples, rotations up to ±17◦, translations up to ±10 mm<ref:2611. p. 0

Where the paper says it

Results: We demonstrate the effectiveness of the proposed method on in vivo 3D ZTE data and simulated data to replicate more challenging hardware scenarios, including dead time gaps spanning 1. 5–5 samples, rotations up to ±17◦ , translations up to ±10 mm, and a range of scan times (153 to 611 seconds).

05169#pg9>. Experimental reconstructions show reduced motion blur and inconsistencies in the tissue contrast compared to the same baselines across larger dead time gaps of 3–4. p. 0
Where the paper says it

5), and reduces rotational motion error in accelerated experiments. Experimental reconstructions show reduced motion blur and inconsistencies in the tissue contrast compared with the same baselines across larger dead time gaps of 3–4.

5 samples<ref:2611. p. 0
Where the paper says it

Conclusion: A unified MAP formulation with k-space motion modeling and an explicit energy-based prior can improve reconstruction quality for motion-corrupted, accelerated ZTE relative to standard reconstruction and motion-correction baselines evaluated in this work. KEYWORDS: ZTE, motion correction, dead time gap, energy model WORD COUNT: 5000 arXiv:2611.

05169#pg9>. EnMoCo maintains high-fidelity reconstructions across gap sizes up to 4 across both acceleration factors, demonstrating superior robustness under high acceleration<ref:2611. p. 7
Where the paper says it

While TV-MoCo provides reasonable artifact suppression at 306 s, its performance degrades substantially on the 153 s dataset: the reduced data redundancy destabilizes its motion parameter estimation, leaving visible residual blurring and localized structural errors. In contrast, EnMoCo maintains highfidelity reconstructions across gap sizes up to 4 across both acceleration factors.

05169#pg9>.

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

  1. Unified Joint Reconstruction Framework for ZTE MRI: The system can jointly estimate image x, rigid-motion parameters xi, and unmeasured k-space samples by alternating between k-space motion updates with conjugate gradient dataconsistency steps and MuSE-based regularization using an explicit energy prior in a data efficient 2.5D form.

  2. Robust Motion Correction Across Acceleration Factors: The system can maintain high fidelity reconstructions across gap sizes up to 4 across both acceleration factors, demonstrating superior robustness under high acceleration compared to baselines like AF and TV-MoCo.

  3. Enhanced Artifact Suppression for Large Dead Time Gaps: The system consistently yields reduced motion and aliasing artifacts, particularly for larger gaps (greater than or equal to 4), by combining motion estimation with an energy-based denoiser, as demonstrated in the comparison of gap sizes 3 to 5.

  4. Improved Motion Estimation Accuracy via Multi-resolution Search: The system can achieve lower l1 error for rotation parameters by employing a coarse-to-fine (multiresolution) motion estimation strategy, which progressively expands a k-space mask from 16 mm resolution up to 1 mm.

  5. Data Efficient Prior Training: The system utilizes a data efficient 2.5D construction that aggregates slice wise energies from two orthogonal views, allowing the prior to be trained with relatively limited data and reducing the demand for fully independent 3D training examples.

Abstract

Zero echo time (ZTE) MRI is near-silent, but dead time gaps, head motion, and undersampling jointly degrade reconstruction. To develop an accelerated ZTE MRI reconstruction framework that jointly estimates the 3D image and rigid-motion parameters, this work introduces Energy-based Motion Correction (EnMoCo), a regularized maximum a posteriori (MAP) formulation that also compensates for missing k-space center data and angular undersampling artifacts. Solved by alternating between k-space rigid-motion updates and conjugate-gradient data-consistency updates, the framework incorporates regularization via a pre-trained Multi-Scale Energy-based model (MuSE) implemented in a data-efficient 2.5D form. Evaluations on simulated and in vivo 3D ZTE data, spanning dead time gaps of 1.5 to 5 samples, rotations up to 17 degrees and translations up to 10 mm, demonstrate that EnMoCo significantly improves PSNR and SSIM relative to CG-SENSE and Autofocusing baselines (achieving roughly 10 to 15 dB higher PSNR for gaps >= 3.5) while reducing motion blur and tissue contrast inconsistencies.

Sources

Related papers