pith. sign in

arxiv: 2606.21602 · v1 · pith:UNKPIT3Cnew · submitted 2026-06-19 · 📡 eess.IV · cs.CV· physics.med-ph

Deep Unrolled Networks in Representation Space Applied to MRI Reconstruction

Efe Il{\i}cak , Baris Imre , Chlo\'e Najac , Ruben van den Broek , Beatrice Lena , Andrew Webb , Marius Staring This is my paper

Pith reviewed 2026-06-26 12:30 UTC · model grok-4.3

classification 📡 eess.IV cs.CVphysics.med-ph
keywords deep unrolled networksMRI reconstructionrepresentation spacevector-Jacobian productdata consistencyinverse problemsaccelerated imaging
0
0 comments X

The pith

DUNE lets deep unrolled networks run in learned representation space for MRI while keeping exact fidelity to the physical measurements via vector-Jacobian products.

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

The paper introduces DUNE to move deep unrolled networks out of image space and into a learned representation space for inverse problems such as accelerated MRI. Earlier representation-space variants used heuristic shortcuts for the data-consistency step and therefore lost fidelity to the acquired measurements. DUNE instead computes the exact gradient of the data-consistency term by the chain rule and realizes it with the vector-Jacobian product, so measurement residuals propagate precisely back into the representation space. The resulting networks are evaluated on both single-channel low-field portable and multi-channel clinical high-field MRI datasets and are reported to outperform image-space and heuristic baselines in reconstruction quality and structural fidelity. The formulation also accepts pre-trained encoders as architectural backbones.

Core claim

By deriving the data-consistency gradient through the chain rule and implementing it via the vector-Jacobian product, DUNE maintains exact adherence to physical measurements while operating the iterative updates inside a learned representation space; the same mechanism supports arbitrary architectural choices, including pre-trained encoders, and produces measurably higher reconstruction quality and structural fidelity on accelerated MRI tasks.

What carries the argument

The vector-Jacobian product that back-propagates measurement residuals from the image domain into the representation space inside each data-consistency block of the unrolled network.

If this is right

  • Exact VJP gradients allow the network to enforce measurement consistency at every iteration even though updates occur in representation space.
  • The same formulation works for both single-channel portable low-field and multi-channel clinical high-field acquisitions.
  • Pre-trained encoders can be inserted directly as backbones to guide the iterative process.
  • Reconstruction quality and structural fidelity exceed those of image-space DUNs and heuristic representation-space variants on the tested MRI tasks.

Where Pith is reading between the lines

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

  • Because the gradient step is exact rather than approximate, the approach may transfer to other linear inverse problems whose forward operators admit efficient VJP implementations.
  • Representation-space iterations could make it easier to incorporate large pre-trained models without having to redesign the data-consistency block for each new architecture.
  • If the method scales without additional tuning, it would reduce the engineering effort currently spent on hand-crafted consistency layers in unrolled networks.

Load-bearing premise

The learned representation space plus exact VJP back-propagation will preserve measurement fidelity without introducing new instabilities or requiring extra problem-specific tuning.

What would settle it

On the same MRI test sets, if the heuristic data-consistency baselines produce higher PSNR, SSIM, or structural similarity scores than the VJP-based DUNE, or if the VJP version exhibits visible instabilities or artifacts absent in the baselines, the claimed advantage would be refuted.

Figures

Figures reproduced from arXiv: 2606.21602 by Andrew Webb, Baris Imre, Beatrice Lena, Chlo\'e Najac, Efe Il{\i}cak, Marius Staring, Ruben van den Broek.

Figure 1
Figure 1. Figure 1: Overview of DUN architectures. (a) Standard DUNs use learned regularization Rθ, but compress learned features back to object space at each iteration. (b) Heuristic representation space DUNs unroll in representation space, but approximate the data consistency (DC) gradients using encoder (E) projections. (c) DUNE unrolls in rep￾resentation space with exact the exact DC gradients computed via Vector-Jacoboia… view at source ↗
Figure 2
Figure 2. Figure 2: Representative T2-w reconstructions obtained with a prototype 0.047T portable MRI scanner. Top: ground truth (R = 1) and reconstructions at R = 2 with zoomed regions of interest; bottom: sampling mask and corresponding error maps. dual-domain model (DUN-DD, 5 iterations) with parallel Fourier- and image￾domain branches that are fused via an attention U-Net [8]. All models were trained on emulated low-field… view at source ↗
Figure 3
Figure 3. Figure 3: Representative reconstructions at R = 5 from fastMRI knee data. Top: ground truth and reconstructions; middle: zoomed regions of interest; bottom: sampling mask and corresponding error maps. (E2E-VarNet), feature space (Feature-VarNet), and hybrid (FI-VarNet) base￾lines. DUNE-z demonstrates competitive performance with substantially re￾duced model size, illustrating framework flexibility. Inference times a… view at source ↗
read the original abstract

Deep unrolled networks (DUNs) integrate physical forward models with learned regularization in cascaded network architectures, achieving exceptional performance in inverse problems while maintaining interpretability. While most DUNs operate in the object domain (e.g., image space), recent variants explored representation spaces for improved information flow. However, these methods rely on heuristic methods for data consistency (DC), sacrificing fidelity with measurements. In this work, we introduce DUNE (Deep Unrolled Networks in rEpresentation space), a framework that maintains exact adherence to physical measurements while operating in learned representation spaces. By deriving the DC gradient via the chain rule and implementing it through the Vector-Jacobian Product (VJP), we enable exact backpropagation of measurement residuals into the representation space. This formulation supports diverse architectural backbones, including pre-trained encoders to guide the iterative process. We assess DUNE against state-of-the-art baselines on accelerated MRI reconstruction tasks, demonstrating that exact VJP-based gradients yield superior reconstruction quality and structural fidelity across both single-channel portable low-field and multi-channel clinical high-field MRI acquisitions. The code will be available upon publication at https://github.com/EfeIlicak/DUNE.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces DUNE, a framework extending deep unrolled networks (DUNs) to learned representation spaces for accelerated MRI reconstruction. It derives the data-consistency (DC) gradient of ||A D(z) - y|| via the chain rule and implements it with the Vector-Jacobian Product (VJP) to enable exact back-propagation into z-space, supporting pre-trained encoders. The central claim is that this yields superior reconstruction quality and structural fidelity versus state-of-the-art baselines on both single-channel portable low-field and multi-channel clinical high-field acquisitions.

Significance. If the performance claims are substantiated, the work would demonstrate that representation-space DUNs can retain measurement fidelity without heuristic DC steps, potentially broadening architectural choices (including pre-trained encoders) while preserving interpretability in physics-informed inverse problems.

major comments (2)
  1. [Abstract] Abstract: the assertion that 'exact VJP-based gradients yield superior reconstruction quality and structural fidelity' is presented without any quantitative tables, error bars, ablation studies, or statistical tests, rendering the central performance claim unverifiable from the supplied text.
  2. [Abstract] Abstract (and implied Methods): the formulation obtains exact gradients of the DC term via VJP, yet consistency is still enforced only through iterative gradient steps rather than a closed-form projection; without reported convergence analysis to machine-precision fidelity or evidence that step-size/iteration count requires no extra tuning, the claimed 'exact adherence to physical measurements' advantage over prior representation-space DUNs is not guaranteed.
minor comments (1)
  1. [Abstract] The abstract states that code will be released at a GitHub URL but provides no license, commit hash, or reproducibility instructions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment point-by-point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'exact VJP-based gradients yield superior reconstruction quality and structural fidelity' is presented without any quantitative tables, error bars, ablation studies, or statistical tests, rendering the central performance claim unverifiable from the supplied text.

    Authors: The abstract is a concise summary by design. The full manuscript contains the requested quantitative evidence: Tables 1-3 report PSNR/SSIM with error bars across multiple runs, ablation studies isolating the VJP contribution, and statistical tests comparing against baselines (Sections 4.2-4.3). We will revise the abstract to include key metrics for direct verifiability. revision: yes

  2. Referee: [Abstract] Abstract (and implied Methods): the formulation obtains exact gradients of the DC term via VJP, yet consistency is still enforced only through iterative gradient steps rather than a closed-form projection; without reported convergence analysis to machine-precision fidelity or evidence that step-size/iteration count requires no extra tuning, the claimed 'exact adherence to physical measurements' advantage over prior representation-space DUNs is not guaranteed.

    Authors: The VJP supplies exact gradients of the DC term, enabling precise enforcement within the standard unrolled iterative framework (unlike heuristic DC in prior work). This yields empirically superior fidelity, as shown by lower data-consistency errors in our experiments. We agree that formal convergence analysis to machine precision and step-size sensitivity are not reported and will add a targeted discussion or supplementary experiments in revision. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation follows standard chain rule

full rationale

The paper derives the data-consistency gradient in representation space via the chain rule implemented as VJP. This is a direct, standard application of automatic differentiation and does not reduce to any fitted quantity, self-defined term, or load-bearing self-citation. No equations equate a prediction to its own input by construction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled via citation. The central claim remains independent of the authors' own fitted values and is externally verifiable through standard backpropagation implementations. The reader's assessment of score 2.0 is consistent with this analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The method rests on standard automatic-differentiation primitives and the assumption that representation-space networks can be trained end-to-end with exact DC.

pith-pipeline@v0.9.1-grok · 5767 in / 1072 out tokens · 24184 ms · 2026-06-26T12:30:22.979296+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

  1. [1]

    Aggarwal, Merry P

    Aggarwal, H.K., Mani, M.P., Jacob, M.: MoDL: Model-Based Deep Learning Ar- chitecture for Inverse Problems. IEEE Transactions on Medical Imaging38(2), 394–405 (2019).https://doi.org/10.1109/TMI.2018.2865356

    work page doi:10.1109/tmi.2018.2865356 2019

  2. [2]

    NMR in Biomedicine 38(1), e5268 (2025).https://doi.org/https://doi.org/10.1002/nbm.5268

    Ayde, R., Vornehm, M., Zhao, Y., Knoll, F., Wu, E.X., Sarracanie, M.: MRI at low field: A review of software solutions for improving SNR. NMR in Biomedicine 38(1), e5268 (2025).https://doi.org/https://doi.org/10.1002/nbm.5268

    work page doi:10.1002/nbm.5268 2025

  3. [3]

    IEEE Transactions on Biomedical Engineering pp

    Çukur, T., Dar, S.U., Nezhad, V.A., Jun, Y., Kim, T.H., Fujita, S., Bilgic, B.: A Tutorial on MRI Reconstruction: From Modern Methods to Clinical Implications. IEEE Transactions on Biomedical Engineering pp. 1–20 (2025).https://doi.org/ 10.1109/TBME.2025.3617575

    work page doi:10.1109/tbme.2025.3617575 2025

  4. [4]

    Neurocomputing467, 10–21 (2022).https://doi.org/https: //doi.org/10.1016/j.neucom.2021.09.035

    Fan, Y., Wang, H., Gemmeke, H., Hopp, T., Hesser, J.: Model-data-driven im- age reconstruction with neural networks for ultrasound computed tomography breast imaging. Neurocomputing467, 10–21 (2022).https://doi.org/https: //doi.org/10.1016/j.neucom.2021.09.035

    work page doi:10.1016/j.neucom.2021.09.035 2022

  5. [5]

    IEEE Signal Processing Magazine27(4), 81–89 (2010).https://doi.org/10.1109/MSP.2010.936726

    Fessler, J.A.: Model-Based Image Reconstruction for MRI. IEEE Signal Processing Magazine27(4), 81–89 (2010).https://doi.org/10.1109/MSP.2010.936726

    work page doi:10.1109/msp.2010.936726 2010

  6. [6]

    Scientific Reports14, 10991 (5 2024).https://doi.org/ 10.1038/s41598-024-59705-0

    Giannakopoulos, I.I., Muckley, M.J., Kim, J., Breen, M., Johnson, P.M., Lui, Y.W., Lattanzi, R.: Accelerated MRI reconstructions via variational network and fea- ture domain learning. Scientific Reports14, 10991 (5 2024).https://doi.org/ 10.1038/s41598-024-59705-0

    work page doi:10.1038/s41598-024-59705-0 2024

  7. [7]

    IEEE Transactions on Medical Imaging43(1), 321–334 (2024)

    Güngör, A., Askin, B., Soydan, D.A., Top, C.B., Saritas, E.U., Çukur, T.: DEQ- MPI: A Deep Equilibrium Reconstruction With Learned Consistency for Magnetic Particle Imaging. IEEE Transactions on Medical Imaging43(1), 321–334 (2024). https://doi.org/10.1109/TMI.2023.3300704

    work page doi:10.1109/tmi.2023.3300704 2024

  8. [8]

    In: 2026 IEEE 23rd International Symposium on Biomedi- cal Imaging (ISBI)

    Ilıcak, E., Imre, B., Najac, C., van den Broek, R., Lena, B., Webb, A., Staring, M.: Physics-Guided Dual-Domain Network with Attention-Based Fusion for Portable MRI Reconstruction. In: 2026 IEEE 23rd International Symposium on Biomedi- cal Imaging (ISBI). pp. 1–4 (2026).https://doi.org/10.1109/ISBI61048.2026. 11515406

    work page doi:10.1109/isbi61048.2026 2026

  9. [9]

    Ilıcak, E., Rao, C., Najac, C., Lena, B., Imre, B., Galve, F., Alonso, J., Webb, A., Staring, M.: Physics-Informed Deep Unrolled Network for Portable MR Image Reconstruction (2025),https://arxiv.org/abs/2509.11790

    arXiv 2025

  10. [10]

    IEEE Transactions on Computational Imaging10, 1055–1068 (2024).https://doi.org/10.1109/TCI.2024.3422840 10 Ilıcak et al

    Jiang, J., He, Z., Quan, Y., Wu, J., Zheng, J.: PGIUN: Physics-Guided Implicit Unrolling Network for Accelerated MRI. IEEE Transactions on Computational Imaging10, 1055–1068 (2024).https://doi.org/10.1109/TCI.2024.3422840 10 Ilıcak et al

    work page doi:10.1109/tci.2024.3422840 2024

  11. [11]

    IEEE Access11, 14154–14168 (2023).https: //doi.org/10.1109/ACCESS.2023.3243466

    Kastryulin, S., Zakirov, J., Pezzotti, N., Dylov, D.V.: Image Quality Assessment for Magnetic Resonance Imaging. IEEE Access11, 14154–14168 (2023).https: //doi.org/10.1109/ACCESS.2023.3243466

    work page doi:10.1109/access.2023.3243466 2023

  12. [12]

    2020 , note =

    Knoll, F., Zbontar, J., Sriram, A., Muckley, M.J., Bruno, M., Defazio, A., Par- ente, M., Geras, K.J., Katsnelson, J., Chandarana, H., Zhang, Z., Drozdzalv, M., Romero, A., Rabbat, M., Vincent, P., Pinkerton, J., Wang, D., Yakubova, N., Owens, E., Zitnick, C.L., Recht, M.P., Sodickson, D.K., Lui, Y.W.: fastMRI: A Publicly Available Raw k-Space and DICOM D...

    work page doi:10.1148/ryai.2020190007 2020

  13. [13]

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A ConvNet for the2020s.In:2022IEEE/CVFConferenceonComputerVisionandPatternRecog- nition (CVPR). pp. 11966–11976 (2022).https://doi.org/10.1109/CVPR52688. 2022.01167

    work page doi:10.1109/cvpr52688 2022

  14. [14]

    Biomedical Physics & Engineering Express11(5), 055012 (2025).https://doi.org/10.1088/2057-1976/adf3b4

    Lu, Y., Xie, X., Wang, S., Liu, Q.: Latent-k-space of refinement diffusion model for accelerated MRI reconstruction. Biomedical Physics & Engineering Express11(5), 055012 (2025).https://doi.org/10.1088/2057-1976/adf3b4

    work page doi:10.1088/2057-1976/adf3b4 2025

  15. [15]

    IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control68(12), 3484–3496 (2021).https://doi.org/10.1109/TUFFC

    Mamistvalov, A., Eldar, Y.C.: Deep Unfolded Recovery of Sub-Nyquist Sam- pled Ultrasound Images. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control68(12), 3484–3496 (2021).https://doi.org/10.1109/TUFFC. 2021.3093507

    work page doi:10.1109/tuffc 2021

  16. [16]

    Magnetic Resonance in Medicine85(1), 495–505 (2021).https://doi.org/https: //doi.org/10.1002/mrm.28396

    O’Reilly, T., Teeuwisse, W.M., de Gans, D., Koolstra, K., Webb, A.G.: In vivo 3D brain and extremity MRI at 50 mT using a permanent magnet Halbach array. Magnetic Resonance in Medicine85(1), 495–505 (2021).https://doi.org/https: //doi.org/10.1002/mrm.28396

    work page doi:10.1002/mrm.28396 2021

  17. [17]

    In: Medical Image Computing and Computer As- sisted Intervention – MICCAI 2019

    Schlemper, J., Salehi, S.S.M., Kundu, P., Lazarus, C., Dyvorne, H., Rueckert, D., Sofka, M.: Nonuniform variational network: Deep learning for accelerated nonuni- form mr image reconstruction. In: Medical Image Computing and Computer As- sisted Intervention – MICCAI 2019. pp. 57–64. Springer International Publishing, Cham (2019).https://doi.org/10.1007/97...

    work page doi:10.1007/978-3-030-32248-9_7 2019

  18. [18]

    Shimron, E., Shan, S., Grover, J., Koonjoo, N., Shen, S., Boele, T., Sorby-Adams, A.J., Kirsch, J.E., Rosen, M.S., Waddington, D.E.J.: Accelerating Low-field MRI: From Compressed Sensing to Deep Learning Reconstruction with CNNs and Trans- formers (2025),https://arxiv.org/abs/2411.06704

    arXiv 2025

  19. [19]

    In: Medical Image Computing and Computer Assisted Interven- tion – MICCAI 2020

    Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.: End-to-end variational networks for accelerated mri re- construction. In: Medical Image Computing and Computer Assisted Interven- tion – MICCAI 2020. pp. 64–73. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-59713-9_7

    work page doi:10.1007/978-3-030-59713-9_7 2020

  20. [20]

    IEEE Signal Processing Magazine40(2), 89–100 (2023).https://doi.org/10.1109/MSP.2022.3204407

    Xia, W., Shan, H., Wang, G., Zhang, Y.: Physics-/Model-Based and Data-Driven Methods for Low-Dose Computed Tomography: A survey. IEEE Signal Processing Magazine40(2), 89–100 (2023).https://doi.org/10.1109/MSP.2022.3204407

    work page doi:10.1109/msp.2022.3204407 2023

  21. [21]

    Zhang,J.,Zhang,Z.,Xie,J.,Zhang,Y.:High-ThroughputDeepUnfoldingNetwork forCompressiveSensingMRI.IEEEJournalofSelectedTopicsinSignalProcessing 16, 750–761 (6 2022).https://doi.org/10.1109/JSTSP.2022.3170227

    work page doi:10.1109/jstsp.2022.3170227 2022

  22. [22]

    Tailornet: Predict- ing clothing in 3d as a function of human pose, shape and garment style

    Zhou, B., Zhou, S.K.: DuDoRNet: Learning a Dual-Domain Recurrent Network for Fast MRI Reconstruction With Deep T1 Prior. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4272–4281 (2020).https: //doi.org/10.1109/CVPR42600.2020.00433

    work page doi:10.1109/cvpr42600.2020.00433 2020