acceptodds
Withdrawn from ICLR 2027

RELAY: Cross-Step Data-Consistency Reuse for Few-Step Medical Image Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse-view computed tomography (CT) and accelerated magnetic resonance imaging (MRI) reduce radiation dose and scan time but yield severely ill-posed inverse problems. Plug-and-play reconstruction avoids protocol-specific retraining by combining a pretrained image prior with operator-specific data consistency (DC), while consistency models reduce prior evaluation to only a few neural function evaluations (NFEs). For sparse-view Radon and multi-coil encoding, however, DC typically requires iterative conjugate gradients (CG), making its truncation a critical source of error under tight computational budgets. We show that this inner solve becomes the primary bottleneck as operator conditioning worsens, while noise injection and momentum provide little benefit. To address this challenge, we introduce RELAY, a training-free solver that transfers the DC solution across outer steps and restarts standard CG using the exact residual of the current system. RELAY improves DC accuracy under unchanged inner-iteration and NFE budgets while preserving the exact DC solution. We further establish non-asymptotic guarantees that characterize its cross-step error propagation and conditioning-dependent iteration savings. Across sparse-view CT and accelerated multi-coil low-field MRI, RELAY achieves the best PSNR among learned baselines across all tested settings, uses substantially fewer CG iterations than a cold start, and on CT surpasses compressed sensing run to convergence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical image reconstruction, Inverse problems, Sparse-view CT, Accelerated MRI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2aW5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies few-step plug-and-play reconstruction with consistency-model priors for sparse-view CT and multi-coil MRI. In this setting, the data-consistency (DC) step has no closed form and is solved approximately with k conjugate-gradient (CG) iterations. The authors argue that, for poorly conditioned operators, truncation error in this inner solve limits reconstruction quality. They propose RELAY, which initializes each DC solve from the previous outer step's DC solution.

Critical strengths

Simple and practical. RELAY needs no training, no extra NFEs, and only one additional image-sized vector of memory.

Thorough DC diagnostics. The paper sweeps the CG budget, NFE, noise level and prior source domain, and reports cost in operator applications rather than only NFE.

Real acquired data. The evaluation includes real multi-coil k-space data (M4Raw).

Critical weaknesses

Limited novelty: the method is warm-started CG. The paper states that RELAY "is algebraically equivalent to applying CG directly to Eq. 4 with initialization z0 = zprev” (Sec. 3, citing Saad, 2003). It also notes in Sec. 2 that “numerical linear algebra has long used warm starts.” The residual-correction formulation, the “RELAY mechanism” panel in Fig. 2, and the “CG on the correction system” entry in Table 1 therefore describe an implementation of a standard warm start, not a new solver. Initializing an iterative x-update from the previous outer iterate is common practice in ADMM implementations. Given this, the contribution amounts to a choice of initial guess for the inner solve. That is useful, but in my view it is below the level expected for ICLR. Could the authors clarify what is technically new beyond warm-starting CG from the previous solution?

Figure 2 is hard to follow. Figure 2 is not self-contained: it relies on bunch of notations that is defined only on page. 4–5, and its “RELAY mechanism” panel suggests a new solver rather than an initialization choice.

The cold-start baseline is weak, and standard alternatives are missing. The main comparisons attribute RELAY's gains to cross-step reuse relative to a zero-initialized CG solve. However, Table 6 shows that simply initializing CG at the DC anchor v already recovers most of the gain (e.g., 37.34 vs. 38.02 dB at 60 views). The standard remedies for an ill-conditioned DC system are also not compared: preconditioned CG, or simply a better initial guess within existing methods. Could the authors report PnP-CM with anchor-initialized CG, and with preconditioned CG, at a matched operator budget, as main-table baselines?

Baseline configuration and the size of the improvements. Section 4 states that the learned baselines retain “the published sampling and guidance settings.” Table 14, however, shows DPS, DDS, DiffPIR and GDM all running at 6 NFE, far below their published regimes. Several baselines appear not to function. DPS (15.3 dB) and CM4IR (15.6 dB) on OrganAMNIST barely change between 18 and 90 views, so they effectively ignore the measurements. DiffPIR and CM produce results within 0.02 dB of each other in every CT setting, with the same operator count. Could the authors tune the diffusion baselines at their intended NFE (or justify the 6-NFE setting), explain the near-identical DiffPIR and CM results, and report variability across test slices?

Is the gain really from more accurate DC? The paper argues that RELAY helps by making the truncated DC solve more accurate while “preserving the exact DC solution.” If so, cold start and RELAY should converge to the same reconstruction as k grows. However, in Table 10 the cold start at k = 30 is still 4 dB below RELAY (33.85 vs. 37.93 dB), and RELAY itself degrades slightly from k = 20 to k = 30 (38.02 vs. 37.93 dB). Could the authors extend the k sweep until the cold start saturates If RELAY exceeds it, how do the authors explain the gain?

My major concern is whether proposed novelty is enough for the conference like ICLR. Although I highly appreciated for simple and neat algorithms, proposed algorithm’s underlying design is already well-studied and is hard to say significant findings.

Questions for authors

Please refer to the weaknesses section.

Reviewer_RTKoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RELAY, a data-consistency reuse framework for few-step medical image reconstruction, to address the limited accuracy of data-consistency solves under restricted conjugate gradient (CG) iteration budgets in sparse-view CT and accelerated multi-coil MRI. Specifically, the method builds on PnP-ADMM and a pretrained consistency model, reusing the data-consistency solution from the preceding outer iteration, computing the exact residual of the current linear system, and applying CG to solve for a correction. This improves reconstruction accuracy under limited computational budgets. Experiments on multiple CT and MRI datasets demonstrate promising reconstruction performance although several issues remain.

Critical strengths

1.The paper addresses insufficiently accurate data-consistency solves in few-step medical image reconstruction and analyzes how operator conditioning affects the allocation of computational budgets between the inner and outer loops, providing valuable insights into reconstruction optimization under limited budgets. 2.The proposed cross-step data-consistency reuse strategy improves solver accuracy while keeping the number of inner iterations and prior evaluations unchanged. Its non-asymptotic theoretical analysis characterizes cross-step error propagation and iteration savings, providing theoretical support for the effectiveness of the method.

Critical weaknesses
  1. RELAY initializes the current CG solve by reusing the data consistency solution from the previous outer iteration, making its core update similar to existing warm-start CG strategies. The authors are encouraged to clarify its substantive distinctions from conventional warm-start methods and its additional contributions.
  2. Some key diagnostic experiments rely on a fixed penalty parameter and artificially prescribed anchor drift. The authors are encouraged to verify whether the error contraction conditions hold during actual CT and MRI reconstruction and to examine their relationship to solver efficiency and reconstruction quality, thereby strengthening the connection between the theoretical analysis and practical reconstruction.
  3. The appendix states only that the penalty parameter schedule follows an existing implementation with task-specific endpoints, leaving the actual settings insufficiently specified. The authors should provide the specific values of ρ_n , its scheduling rule, and the rationale for these choices.
  4. The authors are encouraged to check the correspondence between the dataset names and standardize author name spelling, year suffixes for publications by the same authors in the same year, and the capitalization of proper names in the references to improve the manuscript’s consistency and readability.
Questions for authors
  1. How does RELAY differ substantively from conventional warm-start CG, and what additional contributions does it offer?
  2. Do the error contraction conditions hold in practical CT and MRI reconstruction, and how do they relate to solver efficiency and reconstruction quality?
  3. What are the specific values and scheduling rules for the penalty parameter ρn\rho_n, and what is the rationale for these choices?
Reviewer_JkFUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RELAY, a training-free data-consistency solver for few-step medical image reconstruction using pretrained consistency-model priors. The authors identify truncated conjugate-gradient (CG) solves as a potential bottleneck in plug-and-play reconstruction, particularly for ill-conditioned acquisition operators such as sparse-view Radon transforms and multi-coil MRI encoding.

To address this issue, RELAY reuses the data-consistency solution from the previous outer iteration as the initialization for the current CG solve. It computes the exact residual of the current linear system and solves a correction system using standard CG. The authors also provide non-asymptotic error bounds describing cross-step error propagation and conditioning-dependent iteration savings.

Experiments on OrganAMNIST, LoDoPaB-CT, and M4Raw demonstrate substantial improvements over cold-started reconstruction, particularly under limited CG iteration budgets. The method achieves competitive reconstruction quality across sparse-view CT and accelerated MRI without requiring acquisition-specific retraining.

Critical strengths

1. A practically important bottleneck is identified and addressed effectively.

The paper focuses on an important issue in few-step reconstruction: reducing the number of prior evaluations does not necessarily yield efficient or accurate reconstruction when the inner data-consistency solves remain insufficiently converged.

The authors provide empirical evidence that the initialization and accuracy of these truncated solves strongly affect reconstruction quality.

The proposed strategy is simple, training-free at reconstruction time, and does not modify the underlying data-consistency objective. This makes it potentially useful as a practical improvement to existing plug-and-play reconstruction pipelines.

2. Consistent empirical improvements with controlled solver analyses.

The evaluation covers both sparse-view CT and accelerated multi-coil MRI, demonstrating that the approach can be applied to different acquisition operators.

The authors also provide useful ablations comparing different CG initializations, iteration budgets, and reconstruction settings. Importantly, Table 6 compares RELAY against initialization from the prior output and the current data-consistency anchor, rather than considering only zero initialization.

The results show that reusing previous data-consistency solutions can improve reconstruction quality under tight computational budgets. The additional analyses of operator applications and solver residuals further support the practical relevance of the approach.

Critical weaknesses

1. Limited methodological novelty and insufficient justification for the proposed solver strategy as a distinct contribution.

The central mechanism of RELAY is to initialize the current CG solve using the previous data-consistency solution and recompute the residual for the current linear system.

As explicitly acknowledged in Section 3, this procedure is algebraically equivalent to standard CG initialized with the previous iterate. Warm-starting iterative solvers for sequences of related linear systems is a well-established technique in numerical linear algebra.

Consequently, the primary algorithmic contribution appears to be the application of a standard warm-start strategy to few-step plug-and-play reconstruction, rather than a fundamentally new optimization method.

The theoretical analysis does not fully resolve this novelty concern. The cross-step error recursion in Theorem 1 follows from the classical CG convergence bound combined with an estimate of the change between consecutive linear systems. The resulting guarantee is conditional on the carried iterate being sufficiently close to the new solution, and the iteration-savings result characterizes bounds derived from the classical CG estimate rather than establishing an unconditional improvement.

More importantly, the practical necessity of carrying the previous data-consistency solution is not fully established relative to other natural initializations.

The authors do investigate this issue in Table 6, which is appreciated. However, the results indicate that much of the improvement over zero initialization can already be obtained through simpler initialization choices.

For example, on OrganAMNIST with 60 projection angles, initializing CG from the current data-consistency anchor achieves 37.34 dB, compared with 32.13 dB for zero initialization and 38.02 dB for RELAY. Thus, while RELAY improves over the anchor initialization by 0.68 dB, a substantial portion of the total improvement is already obtained without reusing the previous data-consistency solution.

This distinction is important because the paper emphasizes large improvements over cold-started CG, whereas the more relevant methodological question is whether the proposed reuse strategy offers a significant advantage over strong, straightforward initialization alternatives.

Overall, the effectiveness of warm-starting is well demonstrated, but it remains unclear whether the particular realization proposed as RELAY constitutes a sufficiently novel and necessary methodological contribution.

2. Insufficient evaluation against alternative solver acceleration strategies and limited scope of the theoretical claims.

Although the paper provides extensive comparisons with reconstruction methods, the most relevant alternatives to RELAY are efficient numerical strategies for solving the same sequence of data-consistency systems.

The paper compares RELAY with zero initialization, prior-output initialization, DC-anchor initialization, and an eight-dimensional Krylov-subspace recycling method. These comparisons are valuable, but the initialization study is concentrated on OrganAMNIST, and the Krylov-recycling comparison is conducted in a controlled synthetic solver setting.

It remains unclear whether the observed advantages persist against competitive solver configurations, such as preconditioned CG, improved initialization combined with preconditioning, or alternative warm-start strategies under strictly matched operator-application budgets.

This is particularly relevant because RELAY requires additional operations to compute the residual of the carried iterate. The authors report 276 operator applications for RELAY versus 258 for PnP-CM at the same six neural function evaluations. Although this overhead is relatively small, a more comprehensive computational comparison would help establish whether RELAY provides the best accuracy–efficiency trade-off.

The theoretical analysis also provides a conditional guarantee for the inner data-consistency error rather than a convergence or reconstruction-error guarantee for the complete nonlinear plug-and-play algorithm. The conditions involve the error inherited from the previous step and the drift of the exact data-consistency solutions, but it is not clear how consistently these conditions hold during actual reconstruction trajectories.

Finally, the CT experiments use simulated parallel-beam measurements, and the MRI experiments use retrospectively undersampled acquired k-space. While these benchmarks are appropriate for controlled evaluation, the robustness of the proposed strategy under acquisition-model mismatch and realistic system variations remains less clear.

These limitations matter because the proposed method primarily claims improvements in computational efficiency and solver accuracy. Its advantages should therefore be established against competitive numerical alternatives, rather than primarily against repeated zero initialization.

Questions for authors

1. Methodological novelty and distinction from standard warm-start CG (clarification).

The paper explicitly states that RELAY is algebraically equivalent to applying standard CG initialized with the previous data-consistency solution.

Given that warm-starting sequences of related linear systems is well established, what is the specific algorithmic contribution beyond applying this known strategy to few-step plug-and-play reconstruction?

In particular, which aspect of RELAY is specific to consistency-model reconstruction, rather than being a direct application of standard warm-start CG to any iterative inverse solver?

2. Necessity of carrying the previous data-consistency solution (additional evidence).

Table 6 shows that initialization from the current data-consistency anchor already achieves substantial improvements over zero initialization.

For example, at 60 projection angles, the anchor initialization achieves 37.34 dB compared with 38.02 dB for RELAY.

Could the authors provide similar comparisons across the other datasets, acquisition settings, and CG budgets?

This would help establish whether the previous DC solution consistently provides meaningful benefits over simpler initialization strategies, particularly given the substantially smaller gap between RELAY and anchor initialization than between RELAY and zero initialization.

3. Comparison with competitive numerical solvers (additional evidence).

Could the authors compare RELAY against alternative accelerated data-consistency solvers, particularly preconditioned CG and preconditioned warm-start CG, under matched operator-application and wall-clock budgets?

The comparison with Krylov-subspace recycling in Table 8 is useful, but appears limited to a controlled synthetic solver setting. It would be informative to understand whether the same efficiency advantage holds for the actual CT and MRI reconstruction systems.

4. Practical applicability of the theoretical guarantees (clarification).

Theorem 1 establishes an error recursion involving the drift between consecutive exact data-consistency solutions, with a contraction condition depending on the CG convergence factor and the penalty schedule.

Could the authors provide empirical measurements of these quantities during actual reconstruction trajectories and clarify how frequently the sufficient conditions are satisfied?

Furthermore, can the theoretical analysis establish any relationship between reduced inner-solve error and the final reconstruction error after the nonlinear prior updates, or is the guarantee restricted to individual data-consistency subproblems?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.