{"id":"5b6abecc-c6e0-42a1-936d-92ae3ed445ad","arxiv_id":"2411.10787","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A Patch-GAN based stepwise reconstruction model achieves state-of-the-art average SSIM of 0.9907 and 0.9799 on the two CMRxRecon2024 cardiac MRI reconstruction tasks.","lead":"This paper presents a deep learning model that reconstructs high-quality cardiac MRI images from heavily undersampled data, using a Patch-GAN-based stepwise approach validated on the CMRxRecon2024 challenge. It reports leading scores on both challenge tasks, which could make cardiac MRI scans faster and more comfortable for patients while preserving image quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Task 1's reported SSIM averages three separately trained models (one per acceleration factor, §3), contradicting the 'all-in-one' claim and making Table 1's comparison to single-model baselines potentially inequitable.","rationale":"The paper contains credible components: a public dataset, a GitHub repository, ablation studies, and a clear technical pipeline. The reader's weakest assumption identifies baseline provenance as the key risk, and that is indeed important. But the sharper, more self-inflicted issue is the paper's own Task 1 setup: Section 3 explicitly says three models were trained, one per acceleration factor, and their metrics were averaged. That contradicts the 'all-in-one' framing in the title and abstract. If the comparison baselines were run as single models across accelerations, the proposed approach receives an unfair advantage simply by having more capacity and task-specific specialization. The reader's concern about identical preprocessing and data splits is subsumed under this: even with identical splits, the number of models per task is part of the evaluation protocol. The paper should state plainly that Task 1 uses three models and should either compare against baselines with the same allowance or report a true single-model version. Because the authors may have a legitimate challenge-submission strategy, the verdict remains conditional rather than rejection: the claim is plausible but currently unverified under a matched protocol. I agree with the reader partially because we both center on comparison fairness, but I emphasize an internal inconsistency rather than external provenance.","tokens_in":8736,"tokens_out":6439,"duration_ms":67442,"concrete_test":"Train the proposed generator exactly as in Task 2 (one model) on the Task 1 data covering all three acceleration factors, and evaluate it on the same held-out set; then compare its Task 1 average SSIM against PromptMR under the same data split. Also require per-acceleration SSIM for all methods and a disclosure of whether baselines used one or three models. If the single-model variant of the proposed method does not beat PromptMR, the Table 1 margin is explained by model-count advantage, and the 'all-in-one' framing collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is described as an 'all-in-one' universal reconstruction model, and the headline numbers in Table 1 are presented as method-level comparisons against PromptIR, E2E-Varnet, and PromptMR. However, Section 3 states that for Task 1, 'three different models for three acceleration factors of 4, 8, and 10x' were trained, and the reported Task 1 results are 'the average results of the three trained models.' This means the Task 1 SSIM of 0.9907 is not produced by a single all-in-one model but by an oracle selection among three acceleration-specialized models. The paper does not disclose whether the baselines were likewise allowed one model per acceleration factor or were evaluated as single models; if the latter, the margin in Table 1 is an artifact of model count, not architecture. This is an internal inconsistency between the abstract/title and the experimental setup, and it directly bears on the fairness of the comparison, independent of baseline provenance. A matched protocol (same number of models per task for all methods, or a true single-model run for the proposed method) is needed before the superiority claim can be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a stepwise reconstruction approach for accelerated multi-contrast cardiac MRI, built on a Patch-GAN architecture with an attention-based prompt UNet (APUNet) as the generator, a discriminator, curriculum learning, and a combined loss consisting of k-space physical (MSE of magnitude and phase) and image-domain SSIM terms computed at each reconstruction step. The method is evaluated on the CMRxRecon2024 challenge dataset for Task 1 (uniform undersampling at accelerations 4x, 8x, 10x) and Task 2 (multiple trajectories and accelerations up to 24x). The authors report average SSIM of 0.9907 (Task 1) and 0.9799 (Task 2), outperforming PromptIR, E2E-Varnet, and PromptMR, with ablation studies on cine contrasts supporting the proposed components. Code is made available on GitHub.","tokens_in":8999,"tokens_out":5051,"duration_ms":45716,"significance":"Reconstructing highly undersampled CMR k-space in a single model across contrasts, views, trajectories, and acceleration factors is a timely and valuable goal. The paper combines several plausible ingredients—stepwise unrolled updates, physical and structural losses per step, curriculum learning, and adversarial training—and reports consistent improvements over strong baselines across the contrast/view columns in Table 1. The public release of the code is a concrete strength that should facilitate verification. The significance is currently tempered by the evaluation protocol (three separate models for Task 1 vs. the 'all-in-one' claim), undisclosed baseline provenance, and the absence of statistical uncertainty; if these are addressed, the contribution would be a solid advance for the CMR reconstruction community.","major_comments":[{"comment":"Task 1 results are averages of three separately trained models, one for each acceleration factor (4, 8, and 10x), as stated in the text: 'three different models for three acceleration factors of 4, 8, and 10x have been trained.' The abstract and contributions, however, claim an 'all-in-one' approach. This inconsistency is load-bearing: the superior Task 1 SSIM of 0.9907 is not achieved by a single universal model but by an oracle selection among three specialized models. Unless the baselines were likewise allowed one model per acceleration factor (which is not stated), the comparison is inequitable. The authors should either train and report a single model for Task 1 or explicitly state that the baselines were also evaluated with per-acceleration models.","section":"Section 3, Table 1"},{"comment":"The source of the baseline numbers (PromptIR, E2E-Varnet, PromptMR) is not disclosed. It is not stated whether these numbers are taken from the CMRxRecon2024 official leaderboard or recomputed by the authors under the same data splits, masks, coil compression, and preprocessing pipeline. Without this information, the reported margins cannot be verified as method-level comparisons rather than artifacts of protocol differences. Please specify the provenance of every baseline entry and, if feasible, include a detailed reproducibility statement.","section":"Section 3, Table 1"},{"comment":"The paper identifies Gk in Eq. (2) as 'the gradient of the sparsity regularization term R(x)' (text after Eq. (2)). In Algorithm 1 (lines 12–13), Gk is computed as FFT(ISS), where ISS is derived from the APUNet output; no constraint or derivation links this network output to ∇R(x). This unsupported identification underlies the unrolled-optimization interpretation of Eq. (2). Please either provide a formal argument or empirical validation that the learned module approximates the regularizer gradient, or revise the text to describe Gk as a learned k-space correction without the gradient claim.","section":"Section 2.2, Eq. (2) and Algorithm 1"},{"comment":"No error bars, confidence intervals, or significance tests are provided. Many ablation differences are within a few thousandths of SSIM (e.g., Table 2, cine SAX: proposed 0.9903 vs. without physical loss 0.9892; cine LAX: proposed 0.9893 vs. without curriculum learning 0.9845). Without subject-level variance, these differences could be within noise. Report mean ± std over the evaluation subjects and, ideally, paired statistical tests for the main comparison and the ablation study.","section":"Section 3, Tables 1 and 2"}],"minor_comments":[{"comment":"The first number in each triplet (e.g., '2.6') is not labeled; state explicitly that it denotes NMSE in percent (or as a ratio) to avoid ambiguity.","section":"Table 1"},{"comment":"The abstract reports SSIM as 99.07 and 97.99, while Table 1 gives 0.9907 and 0.9799; use a consistent scale across the manuscript.","section":"Abstract and Table 1"},{"comment":"The word 'centeral' appears multiple times (e.g., in Section 2.2 and Algorithm 1); it should be 'central'.","section":"Section 2.2 and Algorithm 1"},{"comment":"The description of CMRxRecon2024 says Flow2D and BlackBlood images are used exclusively for validation as unseen data, but the paper does not report results for these contrasts in Table 1; clarify which contrasts are included in the official evaluation and why these two are omitted.","section":"Section 2.1"},{"comment":"The citation of the authors' own prior works [13,14] for breath-holding artifacts seems unrelated; consider citing a dedicated CMR motion-artifact reference instead.","section":"Introduction"},{"comment":"Figure 4 is referenced in the text but not discussed; please add a brief interpretation of the qualitative differences shown in the reconstruction example.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution with a public code release, but the 'all-in-one' claim is overstated for Task 1. The main results should be reproducible after the authors clarify the experimental protocol. I see no fundamental flaw that would prevent revision within scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible applied cardiac MRI reconstruction paper with one genuine internal inconsistency in its headline claim. The authors present an all-in-one universal model, but their Task 1 results are the average of three separately trained models, one per acceleration factor (4x/8x/10x). That is worth fixing before anyone takes the Table 1 comparison at face value. Task 2, however, is a true single model across mixed trajectories and acceleration factors, so the core idea is not empty.\n\nWhat is actually new: the stepwise Patch-GAN reconstruction, where each of 12 unrolled reconstructor modules produces k-space output and gets its own physical (MSE in k-space magnitude/phase) plus SSIM loss, with a discriminator on zero-filled and ground-truth concatenations. The prompt-based APUNet and curriculum learning schedule are borrowed from recent work but combined here in a way I have not seen. The ablation table supports the design choices: self-attention, physical loss, stepwise loss, curriculum, and discriminator each contribute.\n\nCredits: evaluation on the public CMRxRecon2024 benchmark; results improve on PromptMR consistently across contrasts and views; code is on GitHub. That is enough to take it seriously.\n\nSoft spots, in order:\n- The Task 1 protocol problem is real. Table 1's Proposed Approach row for Task 1 is the average of three acceleration-specialized models. If the baselines were single models, the comparison is apples-to-oranges. The paper never says whether PromptIR/E2E-Varnet/PromptMR were given the same number of models or evaluated under identical splits and preprocessing. This is the biggest issue.\n- Baseline provenance is undisclosed. Reported numbers may come from the challenge leaderboard or reimplementation; the paper does not say.\n- No error bars or statistical tests. Several SSIM gaps are small (e.g., 0.9903 vs 0.9912 between contrasts), so variance matters.\n- Missing details: the lambda coefficient in Eq. (3) is never given; no commit hash; exact reproduction is uncertain.\n- The identification of Gk with the gradient of the regularization term is loose, but that is a standard unrolling heuristic and not a flaw.\n\nBottom line: the method is plausible and the universal model for Task 2 is a real contribution, but the Task 1 comparison must be re-run under a matched protocol or explicitly relabeled as three-model ensemble results. This deserves a serious referee with the expectation of major revision, not a desk reject.","headline":"Solid challenge-based cardiac MRI paper with a real all-in-one inconsistency in Task 1; Task 2's single-model universal result is the genuine contribution.","tokens_in":9548,"tokens_out":2347,"would_cite":true,"duration_ms":24192,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a stepwise Patch-GAN reconstruction that reconstructs highly undersampled multi-coil cardiac MRI k-space across contrasts, views, and trajectories, reporting average SSIM of 0.9907 on Task 1 and 0.9799 on Task 2 of the…","keywords":["cardiac MRI reconstruction","accelerated MRI","Patch-GAN","k-space undersampling","curriculum learning","multi-contrast CMR","CMRxRecon2024","deep learning reconstruction"],"falsifier":"Re-run all four methods on the same held-out CMRxRecon2024 test subjects with identical coil compression, undersampling masks, and preprocessing, then compare per-contrast NMSE, PSNR, and SSIM. If the proposed approach no longer exceeds PromptMR on average SSIM in both tasks, the central outperformance claim is refuted.","tokens_in":8526,"feed_emoji":"🫀","tokens_out":6744,"duration_ms":59572,"temperature":0.7,"pith_summary":"The paper aims to show that a single deep-learning model can reconstruct high-quality cardiac MRI from heavily undersampled k-space data across the multi-contrast, multi-view, multi-trajectory settings of the CMRxRecon2024 challenge. Its proposed stepwise Patch-GAN approach iteratively refines the k-space using an attention-based prompt UNet, sensitivity-map estimation, and a combined physical-plus-SSIM loss, trained with curriculum learning that starts at low acceleration factors and moves to higher ones. On the challenge tasks the authors report average SSIM of 0.9907 for Task 1 (uniform masks at 4x, 8x, 10x acceleration) and 0.9799 for Task 2 (mixed trajectories at up to 24x), above the compared PromptIR, E2E-Varnet, and PromptMR methods. If the claim holds, the practical payoff is markedly shorter cardiac scan times with image quality close to fully sampled reconstructions.","feed_headline":"Patch-GAN pipeline reconstructs cardiac MRI to 99% SSIM","feed_subtitle":"A single stepwise model beats PromptIR, E2E-Varnet, and PromptMR on CMRxRecon2024, promising faster scans.","key_machinery":"The load-bearing mechanism is the stepwise k-space update $k^{(t+1)} = k^{(t)} - \\eta^{(t)} M (k^{(t)} - k^{(0)}) + G_k$, where $M$ is the subsampling mask, $\\eta^{(t)}$ is a learnable step size, and $G_k$ is a gradient-like term produced by an attention-based prompt UNet (APUNet) that receives adjacent k-spaces, a discriminative prompt, and a conjugate-symmetric sensitivity map. At each of the twelve steps, a physical loss (MSE on magnitude and phase of the central k-space) and an SSIM loss in image space are summed into a stepwise loss that is added to the generator's adversarial loss, mitigating vanishing gradients. The Patch-GAN discriminator receives zero-filled or reconstructed images concatenated with the undersampled input, which helps it learn subsampling artefacts. Curriculum learning orders training from 4x acceleration to higher factors.","core_discovery":"The central discovery is that a single integrated Patch-GAN generator, without a separate video-restoration refinement stage, can serve as an all-in-one CMR reconstruction method. The model consumes adjacent undersampled multi-coil k-spaces, estimates sensitivity maps from auto-calibration lines, and refines the central k-space through twelve stepwise reconstructor modules. Each step compares the intermediate result against the ground-truth k-space through a physics-based loss (magnitude and phase MSE) and against the ground-truth image through SSIM, while a Patch-GAN discriminator encourages local realism. Curriculum learning is used to progress from easier low-acceleration reconstructions to harder high-acceleration ones. The authors report that this combination outperforms the compared methods on both tasks of CMRxRecon2024, with per-contrast SSIM values up to 0.9931.","pith_inferences":["The paper does not state whether the baseline numbers for PromptIR, E2E-Varnet, and PromptMR come from the challenge's official results or from the authors' own reimplementation; a standardized re-run on identical splits and preprocessing would clarify how much of the margin is architectural.","If the reported margins reproduce, the most direct clinical consequence is shorter breath-hold times and fewer motion artefacts in cardiac MRI, because 24x undersampling could become practical.","The ablation table decomposes gains per component, but it does not separately test the discriminator or curriculum on the unseen flow2D and BlackBlood contrasts; a targeted extension would quantify generalization beyond the training distribution.","The same stepwise Patch-GAN design could be tested on non-cardiac MR or on other non-Cartesian trajectories outside the challenge to see whether the all-in-one behaviour transfers."],"forward_implications":["A single trained model can cover multiple contrasts, anatomical views, and k-space trajectories, so clinical deployment would not require retraining per imaging protocol.","The stepwise loss design lets a twelve-module generator train without vanishing gradients, pointing to a straightforward way to deepen reconstruction networks.","Curriculum learning from low to high acceleration factors contributes to the final quality, suggesting that easy-to-hard training is an effective ordering for undersampled reconstruction.","At acceleration factors up to 24x, the reported SSIM and PSNR values stay close to fully sampled quality, which would make substantially faster cardiac scans feasible."],"supporting_citations":[{"why":"Supplies the CMRxRecon2024 dataset, tasks, masks, and unseen validation contrasts that define the evaluation.","marker":"[25]"},{"why":"PromptMR is the main previous-state-of-the-art baseline that the proposed approach must outperform.","marker":"[26]"},{"why":"E2E-Varnet provides the iterative k-space update formulation that the stepwise reconstruction builds on.","marker":"[21]"},{"why":"PromptIR is a compared baseline and the source of the prompt-block idea used inside the APUNet.","marker":"[15]"},{"why":"Patch-GAN provides the adversarial discriminator structure and the local-realism objective.","marker":"[9]"},{"why":"Defines the SSIM metric used both as a loss term and as the primary quality evaluation.","marker":"[24]"},{"why":"Gives the compressed-sensing formulation that motivates the optimization problem in Equation (1).","marker":"[5]"},{"why":"CMRxRecon2023 supplies additional training data and undersampling diversity.","marker":"[23]"}],"fun_headline_variants":["All-in-one Patch-GAN cuts cardiac MRI time, keeps 99% SSIM","Single Patch-GAN model accelerates cardiac MRI to 99% SSIM","Stepwise Patch-GAN offers all-in-one cardiac MRI reconstruction","Cardiac MRI: one Patch-GAN beats prior methods, scores 99% SSIM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison with PromptIR, E2E-Varnet, and PromptMR was run under identical data splits, preprocessing, masks, and evaluation protocol; the paper does not state whether the baseline numbers come from the challenge's official results or from the authors' own reimplementation, so the reported advantage could shrink under a strictly equal re-run.","fun_headline_variants_meta":{"raw":{"variants":["All-in-one Patch-GAN cuts cardiac MRI time, keeps 99% SSIM","Single Patch-GAN model accelerates cardiac MRI to 99% SSIM","Stepwise Patch-GAN offers all-in-one cardiac MRI reconstruction","Cardiac MRI: one Patch-GAN beats prior methods, scores 99% SSIM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1545,"prompt_tokens":943,"completion_tokens":602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":518}},"tokens_in":559,"tokens_out":602,"duration_ms":32460,"temperature":1.0,"reasoning_tokens":518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:17:40.574697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all four methods on the same held-out CMRxRecon2024 test subjects with identical coil compression, undersampling masks, and preprocessing, then compare per-contrast NMSE, PSNR, and SSIM. If the proposed approach no longer exceeds PromptMR on average SSIM in both tasks, the central outperformance claim is refuted.","supporting_citations":[{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention– MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part II 23","cited_arxiv_id":null,"evidence_quote":"E2E-Varnet provides the iterative k-space update formulation that the stepwise reconstruction builds on."},{"cited_title":"Advances in Neural Information Processing Systems 36 (2024)","cited_arxiv_id":null,"evidence_quote":"PromptIR is a compared baseline and the source of the prompt-block idea used inside the APUNet."},{"cited_title":"IEEE Transactions on information theory 52(4), 1289–1306 (2006)","cited_arxiv_id":null,"evidence_quote":"Gives the compressed-sensing formulation that motivates the optimization problem in Equation (1)."},{"cited_title":"Scientific Data11(1), 687 (2024)","cited_arxiv_id":null,"evidence_quote":"CMRxRecon2023 supplies additional training data and undersampling diversity."}],"review_version":1}