Model-based Convolutional De-Aliasing Network Learning for Parallel MR Imaging

Cheng Li; Qiegen Liu; Shanshan Wang; Taohui Xiao; Yanxia Chen

REVIEW 4 major objections 5 minor 15 references

Model-based Convolutional De-Aliasing Network Learning for Parallel MR Imaging

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proposes a model-based convolutional de-aliasing network that reconstructs multi-coil undersampled MRI k-space data without explicit coil sensitivity estimation, using a split-Bregman iteration unrolled into learnable network…

desk verdict A sensible calibrationless unrolled network whose abstract overstates its results: the architecture is worth refereeing, but the evaluation needs a fair MoDL baseline, error bars, and a runtime measurement. read the letter →

arxiv 1908.02054 v1 pith:CRV64KQ2 submitted 2019-08-06 eess.IV cs.CV

Yanxia Chen , Taohui Xiao , Cheng Li , Qiegen Liu , Shanshan Wang This is my paper · ORCID

classification eess.IVcs.CV

keywords parallelmagneticresonanceimagingdeeplearningreconstructionsplitBregmaniterationcalibrationlessmulti-coilk-spaceconvolutionalneuralnetworkundersampledMRI

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Parallel MRI reconstructs images from undersampled multi-coil k-space data, but most methods either need accurate coil sensitivity maps or slow iterative optimization. This paper proposes a deep network that folds a model-based de-aliasing objective into learnable split-Bregman iterations, so the network jointly learns spatial filters, coil-correlation filters, and all regularization parameters end to end. The central claim is that this calibrationless design reconstructs accelerated brain MRI from 12-coil data without explicit sensitivity estimation, and does so with reconstruction quality comparable to or better than three established baselines across uniform, random, Poisson, and radial masks at accelerations from 3x to 9x. A sympathetic reader would care because it points toward fast, calibration-free parallel MRI reconstruction where the physics of multi-coil acquisition is built into the network rather than estimated separately.

What carries the argument

The load-bearing object is the split-Bregman unrolled network defined by Eq. (8). The reconstruction starts from a model that combines data fidelity with two learnable filter operators, $\Phi_s$ in the spatial domain and $\Phi_{coils}$ across coils; introducing a consensus variable $V$ and applying split Bregman yields three simple update steps for $X$, $V$, and the Bregman multiplier $b$. Each update is then mapped to a network layer—Recon, Addition, Conv1, Nonlinear, Conv2, Multi—so the iterative algorithm itself becomes the architecture, and the filter operators are realized as CNNs whose weights and control points are learned. This machinery is what lets the method exploit spatial redundancy and multi-coil correlation jointly while keeping the reconstruction calibrationless.

What would settle it

Retrain the comparison model-based deep-learning baseline on the same 80 training images, same undersampling masks, and same validation set, then evaluate on the same 50 test images; if the baseline then matches or exceeds the proposed method in the configurations where the paper reports superiority, the central comparative claim would be falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an unrolled convolutional network derived from a split-Bregman solution of a de-aliasing objective can replace explicit coil-sensitivity estimation in parallel MRI. The network minimizes a data-fidelity term plus two learned regularizers—one acting in the spatial domain and one across the coil dimension—and unrolls the iteration into alternating reconstruction, addition, convolution, nonlinearity, and multiplier-update layers. All parameters, including the penalty parameter, step sizes, filters, biases, and the shape of the piecewise-linear nonlinearity, are updated during training by backpropagation. On 50 in vivo brain test images, the method reports higher average PSNR/SSIM than the SPIRiT and SAKE baselines in nearly all tested configurations, and is comparable to or slightly better than the model-based deep-learning baseline in most cases, with the main advantage being that it avoids explicit sensitivity calculation.

Load-bearing premise

The evaluation assumes the comparison model-based deep-learning baseline is configured and trained under conditions comparable to the proposed method, even though the paper does not state whether it was retrained on the same training images and masks.

Editorial extensions

If this is right

If the central claim holds, parallel MRI reconstruction can be performed without estimating coil sensitivity maps, removing a common source of artifacts.
The unrolled split-Bregman structure means the network inherits the data-fidelity constraint of the model, so reconstructions stay consistent with measured k-space, not just visually plausible.
Because all reconstruction parameters are learned, the method can adapt penalty and step sizes to the undersampling pattern, which the paper links to its stable high-acceleration results such as 6x Poisson and 9x radial.
The speed gain over iterative methods like SAKE and SPIRiT would make the approach practical in scan workflows where reconstruction latency matters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the calibrationless property suggests the network could transfer across coil arrays with different numbers and geometries without retraining on sensitivity maps, though the paper only demonstrates a fixed 12-coil head coil.
Beyond the paper: the split-Bregman unrolling framework could be applied to other inverse problems where a consensus variable separates a linear forward model from a learned regularizer, such as super-resolution or non-Cartesian trajectories.
Beyond the paper: because the paper does not report variance across test splits or whether the comparison baseline was retrained on the same data, the exact PSNR differences should be treated as configuration-specific rather than universal claims about method superiority.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a real architectural contribution — a split-Bregman unrolled network that keeps spatial and coil regularization in separate learned CNN branches and never computes sensitivity maps. The math is standard and the idea is sensible. But the paper's headline claim doesn't match its own numbers, and the evaluation has holes that need fixing before I'd trust the comparisons.

What's new: most unrolled methods like MoDL and ADMM-Net either assume known sensitivities or estimate them as a separate step. This work removes that step and learns coil correlations through a second filter branch. I haven't seen that exact split before, and the derivation from split Bregman to the layers in Eq. (8) is straightforward enough to follow. The piecewise-linear nonlinearity with learned control points is also a nice touch, even if it's a small detail.

Where it's soft: first, Table 1 doesn't support 'superior performance' in the abstract. MoDL wins the 1D Uniform 4x case by about 2 dB (35.60 vs 33.56) and also wins 1D Random 4x, 2D Poisson 6x, 2D Poisson 9x, and 2D Radial 6x. The proposed method wins the rest, including a few near-ties. So it's roughly competitive, not uniformly superior. Second, there's no runtime anywhere, yet the introduction says 'faster reconstruction speed.' That claim is unmeasured. Third, the MoDL comparison is under-documented: no statement about whether MoDL was retrained on the same 80 training images and masks, or run with released weights. That's a big deal because a mismatch in training data could reverse the ranking. Fourth, no error bars on the 50-image averages, so we don't know if small PSNR differences (36.99 vs 36.53, for example) are meaningful. No code or data is released either, which is frustrating but common for pre-acceptance. These are all fixable in revision. The method itself has no fatal mathematical flaw that I can see; the mapping from the regularizer gradient to two nested CNNs is an approximation stated as 'can be regarded as,' which is fine as a modeling choice but not a derivation.

Who this is for: people working on calibrationless parallel MRI or unrolled reconstruction with learned regularizers. It's a useful baseline in a niche that isn't crowded. I'd send it to peer review, but only with a major revision that fixes the baseline protocol, adds variance, measures runtime, and tones down the abstract. A serious editor shouldn't desk-reject it; the architecture deserves referee time.

Referee Report

4 major / 5 minor

Summary. The paper proposes a model-based convolutional de-aliasing network for multi-coil parallel MRI reconstruction. The method unrolls split Bregman iterations into a fixed-length network whose convolutional filters, piecewise-linear nonlinearity control points, penalty parameter rho, step-size parameters, and update-rate parameter are learned end-to-end from undersampled k-space data. The authors state that the method avoids explicit coil sensitivity estimation and claim superior quantitative and qualitative performance, faster reconstruction speed, and more stable reconstruction than SPIRiT, SAKE, and MoDL. Experiments use an in vivo 12-channel brain dataset with 80 training, 20 validation, and 50 test images, under 1D uniform, 1D random, 2D Poisson, and 2D radial sampling masks at acceleration factors from 3 to 9, reporting NMSE, PSNR, and SSIM.

Significance. If the method and the comparisons hold, the paper provides a calibrationless unrolled deep-learning baseline for parallel MRI and a clear demonstration that learned convolutional filter operators can replace hand-designed spatial and coil-correlation regularizers. The split-Bregman derivation is transparent, and the evaluation is genuinely out-of-sample on 50 held-out test images, which are strengths. The promise of releasing source code is also positive. However, the central empirical claim of superiority over MoDL is not supported by the paper's own Table 1, and the comparison protocol is underdocumented; the significance of the contribution is therefore contingent on a corrected and fairly executed evaluation.

major comments (4)

[Section 2, Eq. (8)] The mapping from the regularization gradient sum in Eq. (7) to the two-layer CNN in Eq. (8) is introduced with the phrase 'can be regarded as,' but no justification is given for why this approximation is valid or how it affects the split Bregman interpretation. Since the convolutional filters, the piecewise-linear control points, and the penalty parameters are all learned during training, the network does not provably inherit the convergence properties or the model-based guarantees of split Bregman. Please provide a derivation, a controlled experiment comparing the CNN approximation with the explicit regularization gradient, or at least a clear discussion of this approximation as a heuristic design choice.
[Section 3 and Table 1] The paper claims in the Abstract and Section 1 that the proposed method achieves 'superior performance' and 'more stable reconstruction' than MoDL, but the paper's own Table 1 shows MoDL with higher PSNR in several configurations, including 1D Uniform 4x (35.60 vs. 33.56), 1D Random 4x (34.14 vs. 33.45), 2D Poisson 9x (30.01 vs. 29.94), and 2D Radial 6x (34.79 vs. 34.72). The comparison is also not documented as fair: Section 3 does not state whether MoDL was retrained on the same 80 training images, tuned on the same 20 validation images with the same masks, or used with released pretrained weights. Please describe the MoDL training protocol precisely, retrain all baselines under identical data splits and mask settings if necessary, report error bars or statistical significance over the 50 test images, and revise the abstract and conclusions to match the actual measured outcomes.
[Section 1 and Section 3] The claim of 'faster reconstruction speed' is never measured anywhere in the paper. Section 3 reports training time (45 hours on an Intel Xeon CPU) but no reconstruction/inference time per image for the proposed method or for SPIRiT, SAKE, or MoDL. Similarly, 'more stable reconstruction performance' is asserted qualitatively without a quantitative stability metric. Please report wall-clock reconstruction times on identical hardware and define what is meant by stability (e.g., variance across test images, sensitivity to initialization, or robustness across masks).
[Section 3] The experimental setup is missing important reproducibility details. The sentence describing the hardware is incomplete ('...equipped with online model training took 45 hours...'), there is no statement about whether a GPU was used, and the generation of the undersampling masks (ACS sizes, random seed, Poisson/radial parameters) is not specified. These details matter for reproducing the reported PSNR/SSIM numbers and for assessing whether the baselines were given comparable computational resources.

minor comments (5)

[Section 3] The phrase 'all the images were obtained by the direct square root of the multi-coil MR images reconstructed by different methods' is unclear; please specify how the coil images were combined into the final magnitude image.
[Figure 3] The error maps in Figure 3 are not accompanied by a color scale or a consistent display window, which makes the qualitative comparison of error maps hard to interpret.
[Section 2] The symbol F is used both for the 2D Fourier transform in A=MF and for the gradient of the regularization function in Eq. (7); this notational collision is confusing and should be resolved.
[Section 2] The derivation of the Addition layer update in Eq. (8) would be clearer if the relationship between mu1, mu2, alpha_r, and rho were stated again near the layer definition, since the notation (n,k) is already dense.
[Section 3] The sentence 'The scanning parameters were TR = 2500ms, TE = 149ms and voxel resolution = 0.9 x 0.9 x 0.9mm' should specify the slice thickness and matrix size, as those affect the acceleration factor and the practical reconstruction problem.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the unrolled network is trained on a separate training set and evaluated on held-out test images, and no fitted parameter is presented as an independent prediction.

full rationale

The paper's derivation chain is self-contained rather than circular. The proposed reconstruction model in Eqs. (1)-(5) is an unconstrained optimization problem with data fidelity and learned regularization terms; the split Bregman updates in Eqs. (6)-(8) are derived explicitly from that model, and the network architecture is presented as the unrolled version of those updates. The filter operators, control points, and penalty parameters are trainable, but they are fitted on 80 training images with 20 validation images and evaluated on 50 held-out test images, so the reported PSNR/SSIM/NMSE results are out-of-sample predictions rather than re-statements of the training fit. The claim of not requiring explicit sensitivity estimation is a design choice of the model, not a quantity fitted from data and then renamed as a prediction. The self-citations in the paper (references [8] and [10], which share an author with the present paper, and reference [14] from the wider MoDL literature) are used only as background descriptions of prior parallel MRI and deep-learning methods; they are not invoked to justify the correctness of the proposed method or to forbid alternatives. The concern that the MoDL baseline may not have been trained under identical conditions, and the observation that Table 1 shows MoDL outperforming the proposed method at several undersampling rates, are evaluation-protocol and correctness issues rather than circularity. There is no equation, parameter, or cited result in this paper whose content is equivalent by construction to the claimed output.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a fully learned set of network parameters, a domain-specific forward model without sensitivity maps, and the unverified convergence of split Bregman for a non-convex CNN prior. There are no new physical entities, but the method's generality depends on the fairness of the evaluation protocol.

free parameters (6)

Conv1 weights and biases (w1, b1) = learned, initialized with DCT filters of size 3x3x8
Feature extraction filters in the V update; learned during backpropagation (Section 3, parameter settings).
Conv2 weights and biases (w2, b2) = learned, initialized with TV filters of size 3x3x1
Feature fusion filters in the V update; learned during training.
Piecewise linear control points {p_i, q_i} = learned
Parameters of the nonlinear activation SPLF, updated during backpropagation.
Penalty parameter rho = initialized 0.2, then updated
Controls data consistency weight in the Recon layer; treated as learnable.
Step size alpha_r (and mu1, mu2) = initialized 0.3; mu2 = 0.06, mu1 = 0.94
Gradient descent step for the V update; learned during training.
Update rate eta_tilde = initialized 1.8
Bregman multiplier update rate; learned during training.

assumptions (4)

domain assumption Split Bregman iteration converges for the learned non-convex CNN regularizer and yields a valid solution to Eq. (2).
Split Bregman is guaranteed for convex L1 problems (reference [15]); here the regularizer is a learned CNN with piecewise linear activations, for which convergence is not established. Invoked in Section 2 when Eq. (6) is derived.
domain assumption The forward model A = MF with no coil sensitivity maps is sufficient for multi-coil reconstruction.
The data fidelity term treats each coil's k-space separately, implying that reconstruction of coil images followed by RSS is valid without sensitivity estimation. This is the basis for the 'no explicit sensitivity' contribution, stated in Section 1 and used in Eq. (1).
ad hoc to paper The CNN layers (Conv1, nonlinear, Conv2) can accurately approximate the regularization gradient sum over filters.
Eq. (8) replaces the analytic form with the statement 'can be regarded as' two convolutions, an approximation that is not justified analytically.
domain assumption The 80 training images are representative of the 50 test images and of clinical brain MRI generally.
Generalization of the trained network depends on this; no cross-dataset evaluation is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Model-based Convolutional De-Aliasing Network Learning for Parallel MR Imaging." pith.science (2026). https://pith.science/paper/CRV64KQ2

@misc{pith2026190802054,
  author       = {Pith},
  title        = {Pith review of: Model-based Convolutional De-Aliasing Network Learning for Parallel MR Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRV64KQ2}},
  note         = {Machine review of arXiv:1908.02054}
}

read the original abstract

Parallel imaging has been an essential technique to accelerate MR imaging. Nevertheless, the acceleration rate is still limited due to the ill-condition and challenges associated with the undersampled reconstruction. In this paper, we propose a model-based convolutional de-aliasing network with adaptive parameter learning to achieve accurate reconstruction from multi-coil undersampled k-space data. Three main contributions have been made: a de-aliasing reconstruction model was proposed to accelerate parallel MR imaging with deep learning exploring both spatial redundancy and multi-coil correlations; a split Bregman iteration algorithm was developed to solve the model efficiently; and unlike most existing parallel imaging methods which rely on the accuracy of the estimated multi-coil sensitivity, the proposed method can perform parallel reconstruction from undersampled data without explicit sensitivity calculation. Evaluations were conducted on \emph{in vivo} brain dataset with a variety of undersampling patterns and different acceleration factors. Our results demonstrated that this method could achieve superior performance in both quantitative and qualitative analysis, compared to three state-of-the-art methods.

Figures

Figures reproduced from arXiv: 1908.02054 by the authors.

**Figure 1.** An illustration of the filter operator with convolutional neural networks for both spatial and multi-coil correlations. 2 Method Let A=MF ∈ CM×N denote a measurement matrix, where M is an undersampled matrix and F is the 2D Fourier transform. Our purpose is to reconstruct X from the undersampled k-space data Y ∈ CM×J (M << N), X = {x1|x2| . . . xj |, . . . xJ |} ∈ C N×J , | stacks the vectors as columns and J is t… view at source ↗

**Figure 2.** The proposed convolutional de-aliasing network architecture for pMRI reconstruction. (a) is the flow chart. The orange arrow indicates the process of reconstructing the undersampled k-space data by forward propagation, and the green arrow indicates the parameter updating through back propagation. (b) and (c) are the detailed configurations of Conv1 and Conv2. where L represents the number of filters. Introducing a… view at source ↗

**Figure 3.** Comparison of different methods in reconstruction accuracy with different undersampling patterns and acceleration factors: reconstruction results and error maps are presented with corresponding quantitative measurements in PSNR/SSIM [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

[1]

Pruessmann, K.P., et al.: SENSE: sensitivity encoding for fast MRI. Magn. Reson. Med. 42(5), 952–962 (1999)

work page 1999
[2]

Griswold, M.A., et al.: Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn. Reson. Med. 47(6), 1202–1210 (2002)

work page 2002
[3]

Lustig, M., Pauly, J.M.: SPIRiT: Iterative self-consistent parallel imaging recon- struction from arbitrary k-space. Magn. Reson. Med. 64(2), 457–471 (2010)

work page 2010
[4]

IEEE Signal

Lustig M, Donoho D L, Santos J M, et al.: Compressed sensing MRI. IEEE Signal. Proc. Mag. 25(2), 72–82 (2008) Model-based Convolutional De-Aliasing Network Learning for pMRI 9

work page 2008
[5]

Shin, P.J., et al.: Calibrationless parallel imaging reconstruction based on structured low-rank matrix completion. Mag. Reson. Med. 72(4), 959–970 (2014)

work page 2014
[6]

Chari, L., et al.: A wavelet-based regularized reconstruction algorithm for SENSE parallel MRI with applications to neuroimaging. Med. Image. Anal. 15(2), 185–201 (2011)

work page 2011
[7]

Iterative image reconstruction using a total variation constraint

Block, K.T., et al.: Undersampled radial MRI with multiple coils. Iterative image reconstruction using a total variation constraint. Magn. Reson. Med. 57(6), 1086– 1098 (2007)

work page 2007
[8]

IEEE Trans

Wang, S., et al.: Learning Joint-Sparse Codes for Calibration-Free Parallel MR Imaging. IEEE Trans. Med. Imaging. 37(1), 251–261 (2018)

work page 2018

Show all 15 references

[9]

Hammernik, K., et al.: Learning a variational network for reconstruction of acceler- ated MRI data. Magn. Reson. Med. 79(6), 3055–3071 (2018)

2018
[10]

In: IEEE ISBI, pp 514–517 Prague, Czech Republic (2016)

Wang, S., et al.: Accelerating magnetic resonance imaging via deep learning. In: IEEE ISBI, pp 514–517 Prague, Czech Republic (2016)

2016
[11]

IEEE Trans

Schlemper J, Caballero J, Hajnal J V, et al.: A deep cascade of convolutional neural networks for dynamic MR image reconstruction. IEEE Trans. Med. Imaging. 37(2), 491–503 (2018)

2018
[12]

In: NIPS, pp 10–18

Sun, J., Li, H., Xu, Z.: Deep ADMM-Net for compressive sensing MRI. In: NIPS, pp 10–18. Barcelona SPAIN (2016)

2016
[13]

In: Frangi, A., Schnabel, J., Davatzikos, C., Alberola-Lpez, C., Fichtinger, G

Zhang, P., Wang, F., Xu, W., Li, Y.: Multi-channel Generative Adversarial Network for Parallel Magnetic Resonance Image Reconstruction in K-space. In: Frangi, A., Schnabel, J., Davatzikos, C., Alberola-Lpez, C., Fichtinger, G. (eds.) MICCAI 2018. LNCS, vol. 11070, pp 180-188. ...

2018 doi
[14]

IEEE Trans

Aggarwal H K, Mani M P, Jacob M.: MoDL: Model-Based Deep Learning Archi- tecture for Inverse Problems. IEEE Trans. Med. Imaging. 38(2): 394–405 (2019)

2019
[15]

Goldstein T, Osher S.: The split Bregman method for L1-regularized problems. SIAM. J. Imaging. Sci. 2(2): 323–343 (2009)

2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Pruessmann, K.P., et al.: SENSE: sensitivity encoding for fast MRI. Magn. Reson. Med. 42(5), 952–962 (1999)

work page 1999

[2] [2]

Griswold, M.A., et al.: Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn. Reson. Med. 47(6), 1202–1210 (2002)

work page 2002

[3] [3]

Lustig, M., Pauly, J.M.: SPIRiT: Iterative self-consistent parallel imaging recon- struction from arbitrary k-space. Magn. Reson. Med. 64(2), 457–471 (2010)

work page 2010

[4] [4]

IEEE Signal

Lustig M, Donoho D L, Santos J M, et al.: Compressed sensing MRI. IEEE Signal. Proc. Mag. 25(2), 72–82 (2008) Model-based Convolutional De-Aliasing Network Learning for pMRI 9

work page 2008

[5] [5]

Shin, P.J., et al.: Calibrationless parallel imaging reconstruction based on structured low-rank matrix completion. Mag. Reson. Med. 72(4), 959–970 (2014)

work page 2014

[6] [6]

Chari, L., et al.: A wavelet-based regularized reconstruction algorithm for SENSE parallel MRI with applications to neuroimaging. Med. Image. Anal. 15(2), 185–201 (2011)

work page 2011

[7] [7]

Iterative image reconstruction using a total variation constraint

Block, K.T., et al.: Undersampled radial MRI with multiple coils. Iterative image reconstruction using a total variation constraint. Magn. Reson. Med. 57(6), 1086– 1098 (2007)

work page 2007

[8] [8]

IEEE Trans

Wang, S., et al.: Learning Joint-Sparse Codes for Calibration-Free Parallel MR Imaging. IEEE Trans. Med. Imaging. 37(1), 251–261 (2018)

work page 2018