MORSE: Multiple Orthogonal Reference Sensitivity Encoding

arXiv:2510.09098 · physics.med-ph · Submitted 2025-10-10 · Read on arXiv

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: "MORSE: Multiple Orthogonal Reference Sensitivity Encoding".

Mira: Here's the gist of the paper. Parallel imaging is ubiquitous in MRI, enabling diverse applications such as ultra-high-resolution functional and quantitative imaging with greater temporal resolution or reduced scan times respectively <ref:2510.09098#pg2>.

Kai: Here we present a computationally efficient method of robustly estimating coil sensitivities, and reconstructing under-sampled images using a data-driven regularised SENSE formalism <ref:2510.09098#pg2>.

Mira: Our “MORSE” scheme estimates multiple sensitivities per voxel to address issues such as rapidly varying sensitivities, chemical shift artefact, or insufficient fields of view and provides a data-driven regularisation term for noise control <ref:2510.09098#pg2>.

Kai: MORSE Scheme Overview The MORSE scheme builds upon the approach of exploiting local correlations to estimate accurate coil sensitivities for subsequent use in a regularised-SENSE general linear model formalism <ref:2510.09098#pg2>.

Title and authors: Kai: So we're looking at the paper titled MORSE: Multiple Orthogonal Reference Sensitivity Encoding. It’s about taking parallel imaging, which is super common in MRI right now, and making it work even better for things like getting really high resolution or speeding up scans.

Mira: Right, it’s about estimating those coil sensitivities more robustly so we can reconstruct the images that were undersampled without needing a ton of time for the calculation. It sounds like they’re tackling the problem of having to estimate these sensitivities every single time you scan something.

Lev: From an error correction standpoint, what I always wonder is how much overhead this estimation adds to the actual reconstruction pipeline when you try to run it on real hardware, you know?

Kai: Well, MORSE proposes a way to do that estimation using a data-driven regularised SENSE formalism. Essentially, they’re not just estimating one sensitivity per voxel; they are estimating multiple sensitivities per voxel.

Mira: That multiple estimation idea is key because it lets them fix problems like rapidly varying sensitivities or chemical shift artifacts which can mess up the standard methods. They want to address those issues directly in the setup.

Lev: But how do you make that multi-sensitivity estimation computationally efficient enough for something that needs to be real-time deployment, not just a batch job? That’s where I get cautious about it.

Kai: The efficiency comes from the way they structure it. They use what they call the MORSE scheme, which estimates multiple orthogonal sensitivities per coil element at each voxel in the field of view. It also includes a data-driven regularisation term for noise control, which helps keep things stable during reconstruction.

Mira: So you’re combining those multiple sensitivity estimates with a specific noise control term to make the whole process more robust against noise and variation across different parts of the scan? That sounds like it could improve image quality significantly.

Lev: If we think about this running on a real system, does this multi-sensitivity approach mean we have to do a lot more matrix multiplications during the unfolding step, or is there some clever trick they use to keep that computational cost manageable?

Kai: They reduce the initial computation by doing a reduced voxel-wise computation first. They start with an initial singular value decomposition on the array coil data to extract the principal modes of variation of those coils across space in k-space.

Mira: Extracting those principal modes sounds like a good way to distill the complexity down into something manageable before you go into the full sensitivity estimation part. What about how this helps with field strength variations, since that’s a big thing in MRI today?

Title and authors: Lev: That’s interesting because they mention this method is particularly useful at ultra-high field strengths, like seven tesla or higher. They say it handles images with rapidly varying coil sensitivities and increased main field inhomogeneities even when the fields of view are restricted.

Kai: Exactly, so the core benefit is robustness across different acquisition conditions. They claim this approach works without needing a body coil reference at all, which simplifies things for many setups.

Mira: And when we look at the math behind it, they establish that fully sampled images are just the Hadamard product of those coil sensitivities and some spatially varying magnetization plus noise. Then they use singular value decomposition on the outer product of the coil-wise image values to get those optimal weights.

Lev: That SVD step sounds like where most of the heavy lifting for estimation happens, so what does that look like in terms of scaling? Is it something that becomes a big problem as you increase the number of coils or voxels?

Kai: They then get this tensor, which is expanded by convolving it with a smoothing kernel in image space. This step helps smooth out the sensitivity estimates. Then they perform a higher order sensitivity estimation by doing an equivalent SVD voxel-wise on that weighted outer product, resulting in k values from one up to Norder.

Mira: So you’re essentially getting multiple "higher-order" sensitivity estimates per voxel, which they say increases robustness against variations within the region of interest. That sounds like a way to improve the unfolding performance later on.

Lev: If we have these higher order estimates, how does that translate to something tangible for someone actually building a reconstruction algorithm? Are we talking about a massive increase in memory usage for storing all those k values?

Kai: They state that as Norder goes from one up to Nref, where one is less than or equal to Nref, this should increase the robustness to intra-ROI sensitivity variation and improve subsequent unfolding performance. It’s about getting a better estimate of the underlying structure.

Mira: From a theoretical standpoint, it seems they are using these higher order estimates in a regularised SENSE framework, which means they're adding that data-driven regularisation term to control the noise level during reconstruction. That’s where I see the most potential for image quality gains.

Lev: The paper does mention that this method consistently produced high-quality, artifact-free images with reconstruction times that are feasible for real-time deployment. Can you give me an idea of how much faster they claim it is compared to other established methods?

Title and authors: Kai: They compare it to vendor native GRAPPA based reconstructions, ESPIRiT, and LORAKS. MORSE delivered improved image quality in terms of both alias and noise suppression when compared to those approaches.

Mira: And specifically for T1 weighted gradient echo images, they achieved a substantially shorter reconstruction time, about eight times faster than ESPIRiT. That is a concrete number that speaks to the speed advantage.

Lev: Eight times faster is significant if we are thinking about clinical throughput or even real-time feedback loops in experimental settings. But what about the limitations they admit? Where does this MORSE method stop working, or what assumptions do they have that might break?

Kai: The paper notes that MORSE is a means of estimating one or more sensitivities per voxel along with a data-driven regularisation term. It was shown to operate well with structural and functional neuroimaging protocols at both three tesla and seven tesla strengths, achieving up to eight times acceleration in those cases.

Mira: The authors also mention that the method is open source, which means it’s been used extensively in their lab, making it accessible for broader use. But they do have to be careful about what they assume when they say it works across different field strengths; the robustness depends on how well those free parameters hold up.

Lev: So, to summarize for the engineering side, we have a method that is computationally efficient because of its structure, robust against field strength and resolution changes due to its design, and it can be fast enough for real-time use. But I’m still watching how stable those higher order sensitivity estimates are when you push the limits of complexity.

Kai: Exactly. The MORSE scheme is a way to get one or more sensitivities per voxel plus that noise control term, put into a regularised SENSE framework to get robust, high-quality reconstructions for many different undersampled data types.

Mira: It seems like they’ve managed to combine the need for better sensitivity estimation with the need for image quality improvement through regularization. It’s a practical approach to tackling the inherent challenges in parallel imaging.

Lev: So, MORSE: Multiple Orthogonal Reference Sensitivity Encoding is a method that provides flexibility in estimating sensitivities and noise control while maintaining reasonable computational efficiency, especially when dealing with challenging acquisitions at higher field strengths like seven tesla.

Kai: That’s it for this piece on MORSE. Next time we'll see what other recent papers are hitting the arXiv.

The paper's summary: Kai: So MORSE is this method that estimates multiple coil sensitivities per voxel and adds a data driven regularization term to make the SENSE reconstruction more robust than standard methods, right?

Mira: That's the gist of it. It’s about taking those multiple sensitivity estimates and using them within a regularized sense framework to get better quality images from undersampled data.

Kai: Exactly. So what this means is that we can handle situations where the coil sensitivities are changing really fast, or there are chemical shift artifacts messing things up, because it estimates several possible sensitivities at each point in space.

Mira: The key assumption here is that these extra sensitivity estimates help us better model the underlying magnetization distribution and control the noise during reconstruction. It’s a data driven term that uses those higher order sensitivity values to keep the final image clean.

Kai: From an experimentalist view, this means we can push our limits at ultra high field strengths, like seven tesla, where things get really tricky with main field inhomogeneities and coil sensitivities changing rapidly.

Mira: And from a theorist standpoint, they’re using this approach to find a way to improve the unfolding process itself by using those higher order sensitivity estimates, which should make it more stable when you’re trying to reconstruct the actual image structure.

Kai: They claim this works well across different imaging protocols, so whether you're doing structural or functional neuroimaging at three tesla or seven tesla, the method is designed to be flexible.

Mira: The paper shows that they can achieve up to eight times acceleration in some of those cases compared to other methods like ESPIRiT, and they claim the results are consistent across different imaging types.

Kai: That acceleration number is pretty substantial if it holds up in practice, but we have to remember this method is designed for certain scenarios where those extra estimates help manage the complexity.

Mira: That's true. The real test will be seeing how robust those higher order sensitivity estimates remain when you push the field strength or resolution to the absolute edge of what’s possible.

Kai: So it sounds like MORSE gives us a powerful tool for getting high quality, artifact free images from undersampled data across a wide range of MRI conditions, but we need to see how well those extra sensitivity values hold up under extreme settings.

The paper's improvements: Kai: So what are the specific improvements they suggest in this MORSE scheme to make it even better than what we have now?

Mira: They focus on how those higher order sensitivity estimates actually help when you're doing the final reconstruction step, which is where we add that data driven regularization term.

Kai: It sounds like they're saying that by using those multiple sensitivity values, you can construct a much tighter regularisation term for noise control during the SENSE unfolding.

Mira: Exactly. They are implying that this specific way of using the higher order estimates augments the standard sense formalism to create a more regularised solution for the actual magnetization distribution in the image.

Kai: From an experimental standpoint, this means we can expect cleaner images, with less noise and fewer artifacts across different contrasts, which is what they claim when they compare it against methods like GRAPPA.

Mira: That's good because if the noise control is tighter, you get a better signal-to-noise ratio in the final image without having to oversample as much k space data initially.

Kai: They also suggest that these free parameters within MORSE are quite robust, meaning this approach should still work well even if you change the resolution or the field strength of your MRI machine.

Mira: The assumption underneath that is that those parameters hold up regardless of those changes, which is a big deal for general applicability across different clinical sites or experimental setups.

Kai: If we think about what this means for someone just listening, it’s that you can get high quality images from undersampled data without needing to know the exact coil sensitivities beforehand, as long as you use this specific multi-sensitivity estimation technique.

Mira: It’s about shifting the burden of complex sensitivity estimation onto a data driven term that learns how to control the noise based on those multiple estimates.

Kai: So the implication is that we can potentially get faster reconstruction times, maybe even real time, because they're using this reduced dimensionality trick early on.

Mira: But we still need to watch how stable those higher order sensitivity estimates are when you push the complexity of your acquisition way beyond what they’ve tested so far.

Conclusion: Tom: So to wrap up, MORSE is this method that uses multiple sensitivity estimates and data driven regularization to get robust reconstructions of undersampled MRI data, right?

Kai: That's right. It’s about using those extra sensitivity values and a noise control term within the regularised sense framework to get better results from undersampled data.

Mira: Essentially, it shows how you can combine those higher order estimates with a specific regularization term to improve the quality of your reconstruction, especially when dealing with things like chemical shift artifacts.

Kai: The big implication here is that we can get faster reconstruction times, maybe even real time, while maintaining high image quality across various field strengths and resolutions.

Mira: I think that’s the thing for practical applications—it moves us closer to systems where you don't need massive amounts of data just to get a good picture.

Lev: For real hardware, the idea of using reduced dimensionality first before doing the most computationally intensive steps sounds like it could keep things manageable on our systems, provided that initial singular value decomposition is efficient.

Kai: Right, so it’s not just about the final step being fast; it’s about how you structure the whole process to be computationally efficient in real time.

Mira: And if we look at what this changes for us, it means we have a more flexible tool for handling complex MRI acquisitions without needing perfect knowledge of every single coil sensitivity beforehand.

Lev: I just think that robustness across different field strengths is key; if it works consistently from three tesla up to seven tesla, that’s a lot less calibration work for the end-user, which is what we always try to build for in our error correction work.

Kai: So we're looking at this MORSE method as a solid way to get high quality images from complex undersampled data, and it’s been shown to handle those challenging conditions well.

Mira: It’s definitely a practical advancement because it addresses the noise control directly in the reconstruction process itself, which is where most traditional methods fall short.

Lev: I'm just hoping that as we build out real hardware with these techniques, this approach doesn't introduce new overhead that slows down our error correction pipeline too much.

Kai: Well, for now, it looks like MORSE is a really solid way to get robust results from complex MRI acquisitions without needing perfect knowledge of every single coil sensitivity upfront.

Oliver Josephs, Barbara Dymerska, Nadine N. Graedel, Yael Balbastre, Nadège Corbin, Martina F. Callaghan

University College London

physics.med-ph

Submitted: 2025-10-10

Updated: 2025-10-10

Comments: 10 pages, 6 figures, This work has been submitted to the IEEE for possible publication

Journal ref: Josephs, O., Dymerska, B., Graedel, N.N. et al. Online image reconstruction via Multiple Orthogonal Reference Sensitivity Encoding (MORSE). Magn Reson Mater Phy (2026)

DOI: 10.1007/s10334-026-01387-y

Code: https://github.com/fil-physics/gadgetron-matlab

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

Importance score: 86/100

The gist: Parallel imaging is ubiquitous in MRI, enabling diverse applications such as ultra-high-resolution functional and quantitative imaging with greater temporal resolution or reduced scan times…

Key concepts

MORSE Scheme
The MORSE scheme estimates several orthogonal coil sensitivities at each voxel. This allows the system to account for complex factors like rapidly changing sensitivities or chemical shift artifacts, making it robust for challenging MRI scenarios, especially at ultra-high field strengths.
Regularized SENSE Formalism
This is a mathematical framework that combines SENSE with a data-driven regularization term. This term uses singular values from higher-order sensitivity estimates to control noise effectively during image reconstruction, leading to better image quality compared to standard methods.
Higher Order Sensitivity Estimation
Instead of just one sensitivity per voxel, MORSE estimates multiple sensitivities (k = 1 to Norder). Using these higher-order estimates increases robustness against variations in sensitivity within a region, improving the final unfolding performance.
Data-Driven Regularization Term
This term is constructed using singular values derived from the higher-order sensitivity matrices. It acts as a noise control mechanism within the regularized SENSE framework, ensuring that the reconstructed images are stable and free from artifacts across different imaging protocols.

Terminology

Summary

Parallel imaging is ubiquitous in MRI, enabling diverse applications such as ultra-high-resolution functional and quantitative imaging with greater temporal resolution or reduced scan times respectively <ref:2510. p. 0

Where the paper says it

Graedel, Yael Balbastre, Nadège Corbin, Martina F. Callaghan Abstract— Parallel imaging is ubiquitous in MRI, enabling diverse applications such as ultra-high-resolution functional and quantitative imaging with greater temporal resolution or reduced scan times respectively.

09098#pg2>. Here we present a computationally efficient method of robustly estimating coil sensitivities, and reconstructing under-sampled images using a data-driven regularised SENSE formalism <ref:2510. p. 0
Where the paper says it

Successful unfolding is contingent on robust and accurate estimation of the relative coil sensitivities, which often involves computation times that preclude real-time deployment. Here we present a computationally-efficient method of robustly estimating coil sensitivities, and reconstructing under-sampled images using a data-driven regularised SENSE formalism.

09098#pg2>. Our “MORSE” scheme estimates multiple sensitivities per voxel to address issues such as rapidly varying sensitivities, chemical shift artefact, or insufficient fields of view and provides a data-driven regularisation term for noise control <ref:2510. p. 0
Where the paper says it

Here we present a computationally-efficient method of robustly estimating coil sensitivities, and reconstructing under-sampled images using a data-driven regularised SENSE formalism. Our “MORSE” scheme estimates multiple sensitivities per voxel to address issues such as rapidly varying sensitivities, chemical shift artefact, or insufficient fields of view and provides a data-driven regularisation term for noise control.

09098#pg2>.

MORSE Scheme Overview

The MORSE scheme builds upon the approach of exploiting local correlations to estimate accurate coil sensitivities for subsequent use in a regularised-SENSE general linear model formalism <ref:2510.09098#pg2>. It estimates multiple orthogonal sensitivities per coil element at each voxel in the FoV to account for chemical shifts, partial volume effects, or insufficient FoV <ref:2510. p. 0

Where the paper says it

Multiple Orthogonal Reference Sensitivity Encoding (MORSE) builds upon the approach of exploiting local correlations [6] to estimate accurate coil sensitivities for subsequent use in a regularised-SENSE general linear model formalism. MORSE estimates multiple orthogonal sensitivities per coil element at each voxel in the FoV to account for chemical shifts, partial volume effects, or insufficient FoV.

09098#pg2>. This method addresses issues particularly problematic at ultra-high field (UHF) strengths (≥ 7T) producing high SNR images despite rapidly varying coil sensitivities and/or increased main field inhomogeneities, even with restricted FoVs, and without requiring a body coil reference <ref:2510. p. 0
Where the paper says it

MORSE estimates multiple orthogonal sensitivities per coil element at each voxel in the FoV to account for chemical shifts, partial volume effects, or insufficient FoV. In doing so, it addresses a number of issues that are particularly problematic at ultra-high field (UHF) strengths (≥ 7T) producing high SNR images despite rapidly varying coil sensitivities and/or increased main field inhomogeneities, even with restricted FoVs, and without requiring a body coil reference, which typically isn’t available at UHF.

09098#pg2>.

Theoretical Foundation

The theory establishes that fully sampled images correspond to the Hadamard product of the coil-wise sensitivities, B ∈ CNcoil × Nvoxel, and the spatially-varying magnetisation, rho ∈ CNvoxel (repeated across Ncoil rows) plus some additive measurement noise <ref:2510.09098#pg3>. The background method estimates optimal weights for combining such coil-wise images via a matched filter approach by performing singular value decomposition (SVD) on the outer product, Er ∈ CNcoil × CNcoil, of the coil-wise image values <ref:2510. p. 1

Where the paper says it

The well-established method of Walsh et al. [6] estimates optimal weights for combining such coil-wise images via a matched filter approach.

09098#pg3>. The tensor Ê ∈ CNvoxel×Ncoil×Ncoil can be obtained by convolving the tensor E with a smoothing kernel of size NROI in image space <ref:2510. p. 1
Where the paper says it

𝑁coil: 𝐸̂ 𝑟,𝑗,𝑘 = ∑ 𝐶𝑗,𝑝𝐶𝑘,𝑝 ∗ 𝑝∈𝒩(𝑟) (2) In this case, 𝑬̂𝑟 is no longer rank 1 due to noise, though the first singular vector will continue to explain the majority of the outer product variance, and to be a scaled version of the relative coil sensitivities. In practice, the tensor 𝐄̂ ∈ ℂ 𝑁voxel×𝑁coil×𝑁coil can be obtained by convolving the tensor 𝐄 with a smoothing kernel of size 𝑁ROI in image space.

09098#pg4>.

Key MORSE Components

The MORSE method involves several key steps for computational efficiency and robustness <ref:2510.09098#pg4>. These include:

  1. Reduced voxel-wise computation, where an initial SVD on the array coil data extracts the principal modes of variation of the coils across space in k-space <ref:2510. p. 1

    Where the paper says it

    1) Reduced voxel-wise computation MORSE operates in a reduced dimensionality (“virtual coil”) space. An initial SVD on the array coil data [19-22] extracts the principal modes of variation of the coils across space.

    09098#pg4>.

  2. Flexible spatial weighting, where uniform weights are replaced by arbitrary weights g ∈ RNROI to facilitate a degree of extrapolation of the sensitivity estimates that affords robustness to motion <ref:2510. p. 1

    Where the paper says it

    are replaced by arbitrary weights, 𝒈 ∈ ℝ 𝑁ROI: 𝐄̂v = 𝐄 v ∗ 𝒈 (3) For example, a 3D Gaussian kernel, centred on location 𝑟, up-weights the sensitivity estimates near this voxel and downweights those further away. This is particularly beneficial at the periphery of the imaged anatomy because it facilitates a degree of extrapolation of the sensitivity estimates that affords robustness to motion between the fully sampled reference data and the under-sampled images, if acquired separately.

    09098#pg4>.

  3. Higher order sensitivity estimation, which involves performing an equivalent SVD voxel-wise on the weighted outer product, Êr v ∈ CNcoil × Nref <ref:2510. p. 1

    Where the paper says it

    For computational efficiency, this convolution in image space is implemented as a multiplication in k-space. 3) Higher order sensitivity estimation The equivalent SVD to (1) is performed voxel-wise on the weighted outer product, 𝑬̂𝑟 v ∈ ℂ 𝑁coil × 𝑁ref: 𝑬̂𝑟 v = 𝑼𝑟𝑺𝑟𝑽𝑟 𝐻 (4) The left singular vectors, i.

    09098#pg4>. This allows for retaining multiple “higher-order” sensitivity estimates per voxel, i. p. 1
    Where the paper says it

    Although the higherorder sensitivities will be orthogonal at a given location, this may not be the case across aliased voxel locations. Analogous to the approach used in ESPIRiT [3], retaining multiple “higher-order” sensitivity estimates per voxel, i.

    e. k = 1.. Norder where 1 < Norder ≤ Nref should increase the robustness to intra-ROI sensitivity variation thereby improving subsequent unfolding performance <ref:2510. p. 1
    Where the paper says it

    Analogous to the approach used in ESPIRiT [3], retaining multiple “higher-order” sensitivity estimates per voxel, i. 𝑁order where 1 < 𝑁order ≤ 𝑁ref should increase the robustness to intra-ROI sensitivity variation thereby improving subsequent unfolding performance.

    09098#pg4>.

Performance and Comparison

MORSE consistently produced high-quality, artefact-free images with reconstruction times feasible for real-time deployment <ref:2510.09098#pg2>. In comparison to vendor-native GRAPPA-based reconstructions, ESPIRiT, and LORAKS, MORSE delivered improved image quality in terms of both alias and noise suppression <ref:2510. p. 7

Where the paper says it

MORSE has been successfully deployed in multiple neuroimaging studies at both 3T and 7T, including functional studies of autobiographical memory processing, visual and auditory perception, and quantitative MRI studies of neurodegenerative diseases such as Huntington’s, Alzheimer’s and Parkinson’s. MORSE delivered improved image quality relative to GRAPPA, ESPIRiT, and LORAKS in terms of both alias and noise suppression, with intermediate reconstruction times that were feasible for real-time deployment (Figs.

09098#pg7>. The MORSE reconstruction had the most homogenous tSNR distribution and the highest mean values for all three participants <ref:2510. p. 5
Where the paper says it

No foldover artefact was visible in the mean EPI (not shown) or the tSNR map reconstructed using MORSE. The MORSE reconstruction also had the most homogenous tSNR distribution and the highest mean values for all three participants: tSNRGRAPPA = {12.

09098#pg7>. Furthermore, MORSE achieved a substantially (approximately 8-fold) shorter reconstruction time than ESPIRiT for T1-weighted gradient echo images <ref:2510. p. 5
Where the paper says it

Higher-order sensitivities reduce the artefact for the MORSE and ESPIRiT reconstructions, but with less noise amplification and a significantly (> 25-fold) shorter reconstruction time for MORSE. 4 illustrates the relative performance of the GRAPPA, ESPIRiT and MORSE reconstructions for a PD-weighted gradient echo image at 7T.

09098#pg8>.

Conclusion

MORSE is a means of estimating one or more sensitivities per voxel, together with a data-driven regularisation term, that can be combined in a regularised SENSE framework to deliver robust, high-quality reconstructions of a wide variety of undersampled data <ref:2510.09098#pg2>. MORSE was shown to operate well with a range of structural and functional neuroimaging protocols at 3T and 7T, with up to 8-fold acceleration <ref:2510. p. 9

Where the paper says it

In all of the evaluated contexts, MORSE outperformed LORAKS, ESPIRiT and GRAPPA. MORSE was shown to operate well with a range of structural and functional neuroimaging protocols at 3T and 7T, with up to 8-fold acceleration.

09098#pg7>. Given the feasibility of real-time reconstruction via Gadgetron, this open-source approach has been used extensively in our lab <ref:2510. p. 9
Where the paper says it

MORSE was shown to operate well with a range of structural and functional neuroimaging protocols at 3T and 7T, with up to 8-fold acceleration. Given the feasibility of real-time reconstruction via Gadgetron, this open-source approach has been used extensively in our lab, including with diverse patient cohorts.

09098#pg2>.

Data Availability Statement

The open-source code, as well as scripts necessary to compile a Gadgetron image and create a Docker container encapsulating the MORSE reconstruction, are available at https://github.com/fil-physics/gadgetron-matlab <ref:2510. p. 9

Where the paper says it

DATA AVAILABILITY STATEMENT The open-source code, as well as scripts necessary to compile a Gadgetron image and create a Docker container encapsulating the MORSE reconstruction, are available at https://github. com/fil-physics/gadgetron-matlab.

09098#pg2>. Example kspace datasets at 3T and 7T are available here: https://zenodo. p. 9
Where the paper says it

com/fil-physics/gadgetron-matlab. Example kspace datasets at 3T and 7T are available here: https://zenodo.

org/records/13737462 <ref:2510. p. 9
Where the paper says it

Example kspace datasets at 3T and 7T are available here: https://zenodo. org/records/13737462.

09098#pg8>.

Acknowledgment

This research was funded in whole, or in part, by the Wellcome Trust [2510.09098 pg1]. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission <ref:2510. p. 9

Where the paper says it

ACKNOWLEDGMENT This research was funded in whole, or in part, by the Wellcome Trust [203147/Z/16/Z and 226793/Z/22/Z]. For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission.

09098#pg2>. We thank Christian Lambert and Charlotte Dore for sharing structural imaging data from the qMAP-PD study, as well as Eleanor A. p. 9
Where the paper says it

For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. We thank Christian Lambert and Charlotte Dore for sharing structural imaging data from the qMAP-PD study, as well as Eleanor A.

Maguire and Yan Wu for sharing 7T fMRI data <ref:2510. p. 9
Where the paper says it

We thank Christian Lambert and Charlotte Dore for sharing structural imaging data from the qMAP-PD study, as well as Eleanor A. Maguire and Yan Wu for sharing 7T fMRI data.

09098#pg2>.

References

[1] K. P. Pruessmann, M. p. 9

Where the paper says it

REFERENCES [1] K. Pruessmann, M.

Weiger, M. B. Scheidegger, and P. p. 9
Where the paper says it

Pruessmann, M. Scheidegger, and P.

Boesiger, SENSE: Sensitivity encoding for fast MRI, Magnetic Resonance in Medicine, vol. 42, no. 5, pp. 952-962, 1999 <ref:2510. p. 9
Where the paper says it

Boesiger, "SENSE: Sensitivity encoding for fast MRI," Magnetic Resonance in Medicine, vol. 952-962, 1999, doi: 10.

09098#pg2>.

[2] M. A. Griswold et al. p. 9

Where the paper says it

1002/(sici)1522-2594(199911)42:5<952::Aidmrm16>3. Griswold et al.

, Generalized autocalibrating partially parallel acquisitions (GRAPPA), Magn Reson Med, vol. 47, no. 6, pp. 1202-10 <ref:2510. p. 9
Where the paper says it

, "Generalized autocalibrating partially parallel acquisitions (GRAPPA)," Magn Reson Med, vol. 1202-10, Jun 2002, doi: 10.

09098#pg2>.

[3] M. Uecker et al. p. 9

Where the paper says it

1202-10, Jun 2002, doi: 10. Uecker et al.

, ESPIRiT--an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA, Magn Reson Med, vol. 71, no. 3, pp. 990-1001 <ref:2510. p. 9
Where the paper says it

, "ESPIRiT--an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA," Magn Reson Med, vol. 990-1001, Mar 2014, doi: 10.

09098#pg7>.

[4] S. Aja-Fernandez, G. p. 9

Where the paper says it

990-1001, Mar 2014, doi: 10. Aja-Fernandez, G.

Vegas-Sanchez-Ferrero, and A. p. 9
Where the paper says it

Aja-Fernandez, G. Vegas-Sanchez-Ferrero, and A.

Tristan-Vega, Noise estimation in parallel MRI: GRAPPA and SENSE, Magn Reson Imaging, vol. 32, no. 3, pp. 281-90 <ref:2510. p. 9
Where the paper says it

Tristan-Vega, "Noise estimation in parallel MRI: GRAPPA and SENSE," Magn Reson Imaging, vol. 281-90, Apr 2014, doi: 10.

09098#pg5>.

[5] D. J. Larkman and R. p. 9

Where the paper says it

281-90, Apr 2014, doi: 10. Larkman and R.

G. Nunes, Parallel magnetic resonance imaging, (in eng), Phys Med Biol, vol. 52, no. 7, pp. R15-55 <ref:2510. p. 9
Where the paper says it

Nunes, "Parallel magnetic resonance imaging," (in eng), Phys Med Biol, vol. R15-55, Apr 7 2007, doi: 10.

09098#pg3>.

[6] D. O. Walsh, A. F. Gmitro, and M. p. 9

Where the paper says it

Phys Med Biol. Gmitro, and M.

W. Marcellin, Adaptive reconstruction of phased array MR imagery, Magnetic Resonance in Medicine, vol. 43, no. 5, pp. 682-690 <ref:2510. p. 9
Where the paper says it

Marcellin, "Adaptive reconstruction of phased array MR imagery," Magnetic Resonance in Medicine, vol. 682-690, 2000, doi: 10.

09098#pg2>.

[7] D. Gol Gungor and L. p. 9

Where the paper says it

1002/(sici)1522-2594(200005)43:5<682::Aidmrm10>3. Gol Gungor and L.

C. Potter, A subspace‐based coil combination method for phased‐array magnetic resonance imaging, Magnetic Resonance in Medicine, vol. 75, no. 2, pp. 762-774 <ref:2510. p. 9
Where the paper says it

Potter, "A subspace‐based coil combination method for phased‐array magnetic resonance imaging," Magnetic Resonance in Medicine, vol. 762- 774, 2016-02-01 2016, doi: 10.

09098#pg4>.

[8] K. L. Wright, J. I. Hamilton, M. p. 9

Where the paper says it

762- 774, 2016-02-01 2016, doi: 10. Hamilton, M.

A Griswold et al. p. 9
Where the paper says it

1002/(sici)1522-2594(199911)42:5<952::Aidmrm16>3. Griswold et al.

, Non-Cartesian parallel imaging reconstruction, (in eng), J Magn Reson Imaging, vol. 40, no. 5, pp. 1022-40 <ref:2510. p. 9
Where the paper says it

Seiberlich, "Non-Cartesian parallel imaging reconstruction," (in eng), J Magn Reson Imaging, vol. 1022-40, Nov 2014, doi: 10.

09098#pg4>.

[9] J. Hamilton, D Franson, and N Seiberlich, Recent advances in parallel imaging for MRI, (in eng), Prog Nucl Magn Reson Spectrosc, vol. 101, pp. 71-95 <ref:2510.09098#pg4>.

[10] B. A Poser and K Setsompop, Pulse sequences and parallel imaging for high spatiotemporal resolution MRI at ultra-high field, Neuroimage, vol. 168, pp. 101-118 <ref:2510. p. 9

Where the paper says it

Setsompop, "Pulse sequences and parallel imaging for high spatiotemporal resolution MRI at ultra-high field," Neuroimage, vol. 101-118, Mar 2018, doi: 10.

09098#pg4>.

[11] B Shafieizargar et al. p. 9

Where the paper says it

neuroimage. Shafieizargar, R.

, Systematic review of reconstruction techniques for accelerated quantitative MRI, Magn Reson Med, vol. 90, no. 3, pp. 1172-1208 <ref:2510. p. 9
Where the paper says it

Poot, "Systematic review of reconstruction techniques for accelerated quantitative MRI," Magn Reson Med, vol. 1172-1208, Sep 2023, doi: 10.

09098#pg4>.

[12] O Maier et al. p. 9

Where the paper says it

1172-1208, Sep 2023, doi: 10. Maier et al.

, CG-SENSE revisited: Results from the first ISMRM reproducibility challenge, (in eng), Magn Reson Med, vol. 85, no. 4, pp. 1821-1839 <ref:2510. p. 9
Where the paper says it

, "CG-SENSE revisited: Results from the first ISMRM reproducibility challenge," (in eng), Magn Reson Med, vol. 1821-1839, Apr 2021, doi: 10.

09098#pg4>.

[13] F A Breuer et al. p. 9

Where the paper says it

Magn Reson Med. Breuer et al.

, Controlled aliasing in volumetric parallel imaging (2D CAIPIRINHA), Magnetic Resonance in Medicine, vol. 55, no. 3, pp. 549-556 <ref:2510. p. 9
Where the paper says it

, "Controlled aliasing in volumetric parallel imaging (2D CAIPIRINHA)," Magnetic Resonance in Medicine, vol. 549-556, 2006-03-01 2006, doi: 10.

09098#pg4>.

[14] J R Polimeni et al. p. 9

Where the paper says it

549-556, 2006-03-01 2006, doi: 10. Polimeni et al.

, Reducing sensitivity losses due to respiration and motion in accelerated echo planar imaging by reordering the autocalibration data acquisition, (in eng), Magn Reson Med, vol. 75, no. 2, pp. 665-79 <ref:2510. p. 9
Where the paper says it

, "Reducing sensitivity losses due to respiration and motion in accelerated echo planar imaging by reordering the autocalibration data acquisition," (in eng), Magn Reson Med, vol. 665-79, Feb 2016, doi: 10.

09098#pg4>.

[15] M Blaimer et al. p. 9

Where the paper says it

Magn Reson Med. Blaimer et al.

, 2D-GRAPPA-operator for faster 3D parallel MRI, (in eng), Magn Reson Med, vol. 56, no. 6, pp. 1359-64 <ref:2510. p. 9
Where the paper says it

, "2D-GRAPPA-operator for faster 3D parallel MRI," (in eng), Magn Reson Med, vol. 1359-64, Dec 2006, doi: 10.

09098#pg8>.

[16] B Bilgic et al. p. 9

Where the paper says it

Magn Reson Med. Bilgic et al.

, Improving parallel imaging by jointly reconstructing multi-contrast data, Magn Reson Med, vol. 80, no. 2, pp. 619-632 <ref:2510. p. 9
Where the paper says it

, "Improving parallel imaging by jointly reconstructing multi-contrast data," Magn Reson Med, vol. 619-632, Aug 2018, doi: 10.

09098#pg8>.

[17] J P Haldar and J Zhuo, P-LORAKS: Low-rank modeling of local k-space neighborhoods with parallel imaging data, Magn Reson Med, vol. 75, no. 4, pp. 1499-514 <ref:2510. p. 9

Where the paper says it

Zhuo, "P-LORAKS: Low-rank modeling of local k-space neighborhoods with parallel imaging data," Magn Reson Med, vol. 1499-514, Apr 2016, doi: 10.

09098#pg8>.

[18] J P Haldar, Low-rank modeling of local k-space neighborhoods (LORAKS) for constrained MRI, IEEE Trans Med Imaging, vol. 33, no. 3, pp. 668-81 <ref:2510. p. 9

Where the paper says it

Haldar, "Low-rank modeling of local k-space neighborhoods (LORAKS) for constrained MRI," IEEE Trans Med Imaging, vol. 668-81, Mar 2014, doi: 10.

09098#pg8>.

[19] M Buehrer et al. p. 9

Where the paper says it

668-81, Mar 2014, doi: 10. Buehrer, K.

, Array compression for MRI with large coil arrays, Magnetic Resonance in Medicine, vol. 57, no. 6, pp. 1131-1139 <ref:2510. p. 9
Where the paper says it

Kozerke, "Array compression for MRI with large coil arrays," Magnetic Resonance in Medicine, vol. 1131-1139, 2007-06-01 2007, doi: 10.

09098#pg8>.

[20] F Huang et al., A software channel compression technique for faster reconstruction with many channels, (in eng), Magn Reson Imaging, vol. 26, no. 1, pp. 133-41 <ref:2510. p. 9

Where the paper says it

Duensing, "A software channel compression technique for faster reconstruction with many channels," (in eng), Magn Reson Imaging, vol. 133-41, Jan 2008, doi: 10.

09098#pg8>.

[21] S B King et al., Optimum SNR data compression in hardware using an Eigencoil array, Magnetic Resonance in Medicine, vol. 63, no. 5, pp. 1346-1356 <ref:2510. p. 9

Where the paper says it

Duensing, "Optimum SNR data compression in hardware using an Eigencoil array," Magnetic Resonance in Medicine, vol. 1346-1356, 2010-05-01 2010, doi: 10.

09098#pg8>.

[22] T Zhang et al., Coil compression for accelerated imaging with Cartesian sampling, Magn Reson Med, vol. 69, no. 2, pp. 571-82 <ref:2510. p. 9

Where the paper says it

Lustig, "Coil compression for accelerated imaging with Cartesian sampling," Magn Reson Med, vol. 571-82, Feb 2013, doi: 10.

09098#pg8>.

[23] F H Lin et al., Parallel imaging reconstruction using automatic regularization, Magn Reson Med, vol. 51, no. 3, pp. 559-67 <ref:2510. p. 9

Where the paper says it

Wald, "Parallel imaging reconstruction using automatic regularization," Magn Reson Med, vol. 559-67, Mar 2004, doi: 10.

09098#pg8>.

[24] M S Hansen and T S Sørensen, Gadgetron: an open source framework for medical image reconstruction, (in eng), Magn Reson Med, vol. 69, no. 6, pp. 1768-76 <ref:2510. p. 9

Where the paper says it

Sørensen, "Gadgetron: an open source framework for medical image reconstruction," (in eng), Magn Reson Med, vol. 1768-76, Jun 2013, doi: 10.

09098#pg8>.

[25] V Malekian et al. p. 9

Where the paper says it

Magn Reson Med. Malekian et al.

, Mitigating susceptibility-induced distortions in high-resolution 3DEPI fMRI at 7T, Neuroimage, vol. 279, p. 120294 <ref:2510. p. 9
Where the paper says it

, "Mitigating susceptibility-induced distortions in high-resolution 3DEPI fMRI at 7T," Neuroimage, vol. 120294, Oct 1 2023, doi: 10.

09098#pg8>.

[26] S J Inati et al. p. 9

Where the paper says it

neuroimage. Inati et al.

, ISMRM Raw data format: A proposed standard for MRI raw datasets, Magnetic Resonance in Medicine, vol. 77, no. 1, pp. 411-421 <ref:2510. p. 9
Where the paper says it

, "ISMRM Raw data format: A proposed standard for MRI raw datasets," Magnetic Resonance in Medicine, vol. 411-421, 2017-01-01 2017, doi: 10.

09098#pg8>.

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. Data-driven regularization term integration: The improved system can incorporate a data-driven regularisation term for noise control by using singular values from the higher-order sensitivity matrices to construct specific regularisation terms, which augment[s] the SENSE formalism to estimate the regularised solution for the magnetisation distribution.

  2. Multi-sensitivity estimation flexibility: The system can estimate multiple orthogonal sensitivities per voxel to address issues like rapidly varying sensitivities, chemical shift artefact, or insufficient fields of view by retaining "higher order sensitivity estimates per voxel, i.e., k = 1.. Norder where 1 < Norder ≤ Nref."

  3. Computational efficiency for real-time deployment: The system can achieve real-time online reconstruction with moderate computing power because it operates in a reduced dimensionality (virtual coil) space and performs the most computationally intensive steps, like the SVD, on these reduced matrices.

  4. Robustness to field strength and resolution variations: The improved system will be robust to variations in field strength and resolution because the free parameters of MORSE proved to be robust to the choice of resolution, image contrast and field strength.

  5. Artifact suppression across contrasts: The system can produce images free of aliasing artefacts across multiple contrasts, readouts, sampling trajectories (CAIPIRINHA sampling in 3T 3D-EPI also tested), as demonstrated by its performance comparisons against GRAPPA and ESPIRiT.

Abstract

Parallel imaging is ubiquitous in MRI, enabling diverse applications such as ultra-high-resolution functional and quantitative imaging with greater temporal resolution or reduced scan times respectively. Successful unfolding is contingent on robust and accurate estimation of the relative coil sensitivities, which often involves computation times that preclude real-time deployment. Here we present a computationally-efficient method of robustly estimating coil sensitivities, and reconstructing under-sampled images using a data-driven regularised SENSE formalism. Our MORSE scheme estimates multiple sensitivities per voxel to address issues such as rapidly varying sensitivities, chemical shift artefact, or insufficient fields of view and provides a data-driven regularisation term for noise control. Exemplar structural and functional image reconstructions at 3T and 7T are presented and compared with a vendor-provided GRAPPA reconstruction as well as state-of-the-art ESPIRiT and LORAKS algorithms. MORSE consistently produced high-quality, artefact-free images with reconstruction times feasible for real-time deployment. It is flexible and robust, and made available to the community in open-source as a library of functions within the vendor-agnostic Gadgetron image reconstruction framework.

Related papers