VS-Net: Variable splitting network for accelerated parallel MRI reconstruction

Ben Statton; Carlo Biffi; Cheng Ouyang; Chen Qin; Daniel Rueckert; Declan P O'Regan; Ghalib Bello; Jinming Duan; Jo Schlemper; Wenjia Bai

arxiv: 1907.10033 · v1 · pith:ACY42ZTUnew · submitted 2019-07-19 · 📡 eess.IV · cs.CV

VS-Net: Variable splitting network for accelerated parallel MRI reconstruction

Jinming Duan , Jo Schlemper , Chen Qin , Cheng Ouyang , Wenjia Bai , Carlo Biffi , Ghalib Bello , Ben Statton

show 2 more authors

Declan P O'Regan Daniel Rueckert

This is my paper

Pith reviewed 2026-05-24 18:41 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords parallel MRIvariable splittingcompressed sensingdeep learningimage reconstructionundersampled datamulti-coil MRIenergy minimization

0 comments

The pith

VS-Net reconstructs accelerated parallel MRI data more accurately than prior deep learning methods by unrolling variable splitting iterations into a trainable network.

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

The paper introduces VS-Net as a way to reconstruct images from undersampled multi-coil MR data. It begins with an energy minimization formulation of the parallel compressed sensing problem and derives a variable splitting optimization to solve it. The authors unroll the resulting iterative steps into an end-to-end trainable neural network architecture. A reader would care because this structure could enable shorter scan times while preserving or improving diagnostic image quality. Evaluation on complex-valued knee images at four-fold and six-fold acceleration shows gains over existing deep learning approaches in both accuracy and perceptual quality.

Core claim

The paper claims that the generalized parallel compressed sensing reconstruction problem can be formulated as an energy minimization task, for which a variable splitting optimization method is derived, and that unrolling the iterative process of this scheme produces an end-to-end trainable network, VS-Net, that achieves superior reconstruction accuracy and perceptual quality compared with state-of-the-art deep learning algorithms on complex-valued multi-coil knee images for acceleration factors of four and six.

What carries the argument

VS-Net, the end-to-end trainable deep neural network obtained by unrolling the iterations of the variable splitting optimization derived from the energy minimization formulation of parallel MRI reconstruction.

If this is right

The unrolled network can be trained directly on complex-valued multi-coil data.
Reconstruction accuracy exceeds that of prior deep learning methods at both four-fold and six-fold acceleration.
Perceptual quality of the output images improves relative to existing networks.
The approach applies to undersampled parallel MRI without requiring hand-crafted regularization parameters at inference time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same unrolling strategy might transfer to other iterative solvers used in medical image reconstruction tasks.
The learned parameters could allow the network to adapt automatically to variations in coil sensitivity maps or noise characteristics across scanners.
Increasing the number of unrolled stages or incorporating additional splitting variables might further reduce residual artifacts on difficult cases.

Load-bearing premise

The benefits of the original iterative variable splitting scheme are preserved when its steps are unrolled into a finite-depth network whose parameters are learned from data and then applied to new multi-coil acquisitions.

What would settle it

A head-to-head test on held-out multi-coil knee MRI data in which VS-Net produces lower quantitative accuracy or visibly worse perceptual quality than the strongest competing deep learning reconstruction method would falsify the central claim.

Figures

Figures reproduced from arXiv: 1907.10033 by Ben Statton, Carlo Biffi, Cheng Ouyang, Chen Qin, Daniel Rueckert, Declan P O'Regan, Ghalib Bello, Jinming Duan, Jo Schlemper, Wenjia Bai.

**Figure 2.** Figure 2: Detailed structure of each block in VS-net. DB, DCB and WAB stand for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Quantitative measures versus number of epochs at training (first two [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Visual comparison of results obtained by different methods for Cartesian [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

In this work, we propose a deep learning approach for parallel magnetic resonance imaging (MRI) reconstruction, termed a variable splitting network (VS-Net), for an efficient, high-quality reconstruction of undersampled multi-coil MR data. We formulate the generalized parallel compressed sensing reconstruction as an energy minimization problem, for which a variable splitting optimization method is derived. Based on this formulation we propose a novel, end-to-end trainable deep neural network architecture by unrolling the resulting iterative process of such variable splitting scheme. VS-Net is evaluated on complex valued multi-coil knee images for 4-fold and 6-fold acceleration factors. We show that VS-Net outperforms state-of-the-art deep learning reconstruction algorithms, in terms of reconstruction accuracy and perceptual quality. Our code is publicly available at https://github.com/j-duan/VS-Net.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

VS-Net unrolls variable splitting for parallel MRI but the abstract gives no metrics or stats to support the outperformance claim.

read the letter

VS-Net takes the variable splitting approach to solving the energy minimization problem for multi-coil parallel MRI and unrolls the iterations into an end-to-end network. That is the central move in the paper. The derivation starts from the generalized compressed sensing formulation, applies variable splitting, and produces a finite-depth architecture that can be trained directly on undersampled data. Releasing the code is a clear positive, as it lets others check the implementation and try the method on their own scans. For work in this area around 2019, grounding the network in an explicit optimization scheme rather than an arbitrary stack of layers is a reasonable step and distinguishes it from purely empirical designs. The claim is that it beats other deep learning methods on knee data at 4x and 6x acceleration in both accuracy and perceptual quality. The soft spot is the evaluation. The abstract states the superiority but supplies none of the numbers, no table of PSNR or SSIM values, no list of exact baselines, no dataset size, and no mention of statistical testing. Without those details it is impossible to judge the size of any gain or whether it truly comes from the splitting formulation. The stress-test worry about whether the unrolled layers actually retain the fixed-point properties of the original iterative solver is reasonable here, because the provided text does not include convergence checks or ablations that would address it. This paper is aimed at researchers already working on unrolled networks for MRI reconstruction who want to see one more variant derived from variable splitting. A reader looking for a new paradigm or broad clinical impact will not find it. It is solid enough on the formulation side to deserve peer review so that the experiments can be examined in full, though the current abstract alone would not be enough to accept the central claim.

Referee Report

2 major / 0 minor

Summary. The paper proposes VS-Net, a deep neural network for accelerated parallel MRI reconstruction obtained by unrolling a variable splitting optimization scheme derived from an energy minimization formulation of generalized parallel compressed sensing. It evaluates the approach on complex-valued multi-coil knee images at 4-fold and 6-fold acceleration and claims that VS-Net outperforms state-of-the-art deep learning reconstruction algorithms in reconstruction accuracy and perceptual quality. The code is released publicly.

Significance. If the outperformance claim is substantiated with quantitative metrics and the unrolling is shown to preserve optimization benefits, the work would add to the literature on model-based deep learning for MRI by providing an explicit derivation from variable splitting. The public code release supports reproducibility and is a clear strength.

major comments (2)

[Abstract] Abstract: the central claim that VS-Net 'outperforms state-of-the-art deep learning reconstruction algorithms, in terms of reconstruction accuracy and perceptual quality' is stated without any numerical metrics, baseline comparisons, dataset sizes, statistical tests, or error bars. This absence directly undermines assessment of the primary empirical result.
[Unrolling and network architecture description] The variable splitting derivation and unrolling (following the energy minimization formulation): no analysis is provided showing that the finite-depth network converges to the same fixed point as the iterative solver, and no ablation isolates whether gains arise from the splitting scheme versus network capacity or training choices. This assumption is load-bearing for the claim that the derived network retains advantages on unseen multi-coil data at 4x/6x acceleration.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and the unrolling analysis. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that VS-Net 'outperforms state-of-the-art deep learning reconstruction algorithms, in terms of reconstruction accuracy and perceptual quality' is stated without any numerical metrics, baseline comparisons, dataset sizes, statistical tests, or error bars. This absence directly undermines assessment of the primary empirical result.

Authors: We agree that the abstract would benefit from quantitative support for the performance claim. In the revised manuscript we will update the abstract to include specific metrics (e.g., average PSNR and SSIM values at 4× and 6× acceleration on the knee dataset), the number of test volumes, and a brief reference to the compared methods. This change directly addresses the concern while remaining within abstract length limits. revision: yes
Referee: [Unrolling and network architecture description] The variable splitting derivation and unrolling (following the energy minimization formulation): no analysis is provided showing that the finite-depth network converges to the same fixed point as the iterative solver, and no ablation isolates whether gains arise from the splitting scheme versus network capacity or training choices. This assumption is load-bearing for the claim that the derived network retains advantages on unseen multi-coil data at 4x/6x acceleration.

Authors: We clarify that VS-Net performs a fixed number of unrolled iterations that are trained end-to-end; it is not designed or claimed to reach the fixed point of the original iterative solver. We will add a short paragraph in the methods section explaining this distinction and why a formal fixed-point convergence analysis is not applicable. On the ablation, we acknowledge the value of isolating the contribution of the splitting scheme; we will add a controlled comparison against a generic CNN of matched parameter count in the experiments section of the revision. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation chain is self-contained

full rationale

The paper formulates parallel CS MRI as an energy minimization problem, derives a variable splitting scheme, and unrolls the iterations into VS-Net for end-to-end training. Performance claims rest on empirical evaluation against SOTA methods on held-out multi-coil knee data at 4x/6x acceleration, not on any fitted parameter being renamed as a prediction or on self-citation chains. No equations reduce the reported accuracy or perceptual quality to quantities defined by construction from the inputs; the unrolling is a standard architectural choice whose benefits are validated externally rather than assumed by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields limited visibility; the claim rests on the domain assumption that MRI reconstruction admits an energy-minimization formulation amenable to variable splitting and that finite unrolling yields a trainable network with retained optimization properties.

axioms (1)

domain assumption Parallel MRI reconstruction can be cast as an energy minimization problem whose solution is obtained via variable splitting
Explicitly stated as the starting point for deriving the iterative process that is unrolled.

pith-pipeline@v0.9.0 · 5698 in / 1224 out tokens · 29365 ms · 2026-05-24T18:41:28.937007+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 2 internal anchors

[1]

IEEE Trans

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

work page 2018
[2]

In: NIPS

Yan, Y., et al.: Deep ADMM-Net for compressive sensing MRI. In: NIPS. (2016) 10–18

work page 2016
[3]

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

work page 2018
[4]

IEEE Trans

Aggarwal, H.K., et al.: MoDL: Model-based deep learning architecture for inverse problems. IEEE Trans. Med. Imag. 38(2) (2019) 394–405

work page 2019
[5]

k-Space Deep Learning for Accelerated MRI

Han, Y., et al.: k-space deep learning for accelerated MRI. arXiv:1805.03779 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[6]

Akcakaya, M., et al.: Scan-speciﬁc robust artiﬁcial-neural-networks for k-space interpolation (RAKI) reconstruction: Database-free deep learning for fast imaging. Magn. Reson. Med. 81(1) (2018) 439–453

work page 2018
[7]

Self-Supervised Deep Active Accelerated MRI

Jin, K., et al.: Self-supervised deep active accelerated MRI. arXiv:1901.04547 (2019)

work page internal anchor Pith review Pith/arXiv arXiv 1901
[8]

IEEE Trans

Mardani, M., et al.: Deep generative adversarial neural networks for compressive sensing MRI. IEEE Trans. Med. Imag. 38(1) (2019) 167–179

work page 2019
[9]

IEEE Trans

Tezcan, K., et al.: MR image reconstruction using deep density priors. IEEE Trans. Med. Imag. (2018)

work page 2018
[10]

In: MICCAI

Zhang, P., et al.: Multi-channel generative adversarial network for parallel magnetic resonance image reconstruction in k-space. In: MICCAI. (2018) 180–188

work page 2018
[11]

Uecker, M., et al.: Espirit an eigenvalue approach to autocalibrating parallel MRI: where sense meets grappa. Magn. Reson. Med. 71(3) (2014) 990–1001

work page 2014
[12]

Lu, W., et al.: Implementation of high-order variational models made easy for image processing. Math. Method Appl. Sci. 39(14) (2016) 4208–4233

work page 2016
[13]

Lu, W., et al.: Graph-and ﬁnite element-based total variation models for the inverse problem in diﬀuse optical tomography. Biomed. Opt. Express 10(6) (2019) 2684–2707

work page 2019
[14]

Duan, J., et al.: Denoising optical coherence tomography using second order total generalized variation decomposition. Biomed. Signal Process. Control 24 (2016) 120–127 10 J Duan, et al

work page 2016
[15]

Multimed

Liu, R.W., et al.: Undersampled cs image reconstruction using nonconvex nons- mooth mixed constraints. Multimed. Tools Appl. 78(10) (2019) 12749–12782

work page 2019
[16]

Uecker, M., et al.: Software toolbox and programming library for compressed sensing and parallel imaging, Citeseer

work page
[17]

IEEE Trans

Murphy, M., et al.: Fast l1-spirit compressed sensing parallel imaging MRI: Scalable parallel implementation and clinically feasible runtime. IEEE Trans. Med. Imag. 31(6) (2012) 1250–1262

work page 2012

[1] [1]

IEEE Trans

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

work page 2018

[2] [2]

In: NIPS

Yan, Y., et al.: Deep ADMM-Net for compressive sensing MRI. In: NIPS. (2016) 10–18

work page 2016

[3] [3]

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

work page 2018

[4] [4]

IEEE Trans

Aggarwal, H.K., et al.: MoDL: Model-based deep learning architecture for inverse problems. IEEE Trans. Med. Imag. 38(2) (2019) 394–405

work page 2019

[5] [5]

k-Space Deep Learning for Accelerated MRI

Han, Y., et al.: k-space deep learning for accelerated MRI. arXiv:1805.03779 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[6] [6]

Akcakaya, M., et al.: Scan-speciﬁc robust artiﬁcial-neural-networks for k-space interpolation (RAKI) reconstruction: Database-free deep learning for fast imaging. Magn. Reson. Med. 81(1) (2018) 439–453

work page 2018

[7] [7]

Self-Supervised Deep Active Accelerated MRI

Jin, K., et al.: Self-supervised deep active accelerated MRI. arXiv:1901.04547 (2019)

work page internal anchor Pith review Pith/arXiv arXiv 1901

[8] [8]

IEEE Trans

Mardani, M., et al.: Deep generative adversarial neural networks for compressive sensing MRI. IEEE Trans. Med. Imag. 38(1) (2019) 167–179

work page 2019

[9] [9]

IEEE Trans

Tezcan, K., et al.: MR image reconstruction using deep density priors. IEEE Trans. Med. Imag. (2018)

work page 2018

[10] [10]

In: MICCAI

Zhang, P., et al.: Multi-channel generative adversarial network for parallel magnetic resonance image reconstruction in k-space. In: MICCAI. (2018) 180–188

work page 2018

[11] [11]

Uecker, M., et al.: Espirit an eigenvalue approach to autocalibrating parallel MRI: where sense meets grappa. Magn. Reson. Med. 71(3) (2014) 990–1001

work page 2014

[12] [12]

Lu, W., et al.: Implementation of high-order variational models made easy for image processing. Math. Method Appl. Sci. 39(14) (2016) 4208–4233

work page 2016

[13] [13]

Lu, W., et al.: Graph-and ﬁnite element-based total variation models for the inverse problem in diﬀuse optical tomography. Biomed. Opt. Express 10(6) (2019) 2684–2707

work page 2019

[14] [14]

Duan, J., et al.: Denoising optical coherence tomography using second order total generalized variation decomposition. Biomed. Signal Process. Control 24 (2016) 120–127 10 J Duan, et al

work page 2016

[15] [15]

Multimed

Liu, R.W., et al.: Undersampled cs image reconstruction using nonconvex nons- mooth mixed constraints. Multimed. Tools Appl. 78(10) (2019) 12749–12782

work page 2019

[16] [16]

Uecker, M., et al.: Software toolbox and programming library for compressed sensing and parallel imaging, Citeseer

work page

[17] [17]

IEEE Trans

Murphy, M., et al.: Fast l1-spirit compressed sensing parallel imaging MRI: Scalable parallel implementation and clinically feasible runtime. IEEE Trans. Med. Imag. 31(6) (2012) 1250–1262

work page 2012