{"id":"a7c4c6e9-f7fb-4f4d-bebe-f8e923fd8599","arxiv_id":"2411.10403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A cardiac MRI reconstruction model using adaptive unrolling, pattern and contrast prompts, and channel-shifting reconstructs mixed protocols with one network.","lead":"This paper presents a single MRI reconstruction network that adapts to different heart scan protocols, including image contrast, sampling pattern, and acceleration rate, using adaptive unrolling, prompts, and channel-shifting. The authors report improved image-quality scores over their own fixed-iteration baselines on cardiac MRI data from the CMRxRecon dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive unrolling's causal role is untested: no entry-UI schedule is reported, and the Adaptive-vs-Fixed comparison does not isolate the schedule from capacity allocation.","rationale":"The reader's weakest assumption points to the adaptive-unrolling schedule, and that is also the most load-bearing concern here. The paper's headline contribution is a foundation model built on adaptive unrolling, PCP-UNet, and channel-shifting, but the adaptive component is the only one directly tied to the compressed-sensing rationale in Section 2.2. The missing schedule and missing ablation mean the empirical comparison cannot establish that adaptivity, rather than extra capacity or better routing, produces the reported gains. The lack of a protocol-specific baseline is a real limitation, but the fixed-iteration UNet baseline can plausibly be read as the 'conventional ML-based method' named in the abstract, so that concern is less decisive than the unvalidated adaptive mechanism. The proposed permutation ablation would settle whether the schedule matters; until then, a conditional verdict is appropriate.","tokens_in":5954,"tokens_out":9650,"duration_ms":100829,"concrete_test":"Require the authors to release the exact entry-UI schedule for R=4,8,12,16,20,24, and run the following ablation: retrain Adaptive PCP-UNet with a reversed schedule (R=4 enters at the UI used for R=24, and vice versa) and with a randomized schedule, keeping all other hyperparameters and data splits fixed. Also compute the condition number, or the number of CG iterations to reach a fixed residual, of the actual E_Omega for each acceleration rate and sampling pattern. If the reversed/randomized schedules match the reported SSIM/NRMSE within one standard error, or if the schedule does not track the conditioning measure, then adaptive unrolling is not the causal driver and the claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 asserts that acceleration rate affects the conditioning of E_Omega and hence the number of unrolled iterations (UI) needed, and Fig. 1 implements this by routing each sample to a starting UI. Three load-bearing details are missing. (1) The mapping from acceleration rate to starting UI is never reported; the paper states only that up to sixteen UIs exist and that higher rates use more UIs. (2) The unrolled cascade has independent network and CG parameters per UI, so it is not iterating to convergence of a single optimization; compressed-sensing intuition about convergence does not transfer directly. (3) The Adaptive-vs-Fixed comparison in Fig. 6 changes both the per-rate number of active UIs and the allocation of training samples across subnetworks, so the reported SSIM gain cannot be attributed to conditioning-matched adaptivity. If the schedule is arbitrary, or if the gain comes from capacity allocation rather than from matching iteration count to conditioning, the paper's central novelty is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a single reconstruction network intended to work across cardiac MRI contrasts, views, sampling patterns, and acceleration rates. It combines adaptive unrolling, in which the number of unrolled iterations is chosen based on the acceleration rate, with a Pattern and Contrast-Prompt UNet (PCP-UNet) and a channel-shifting input augmentation. Experiments on the CMRxRecon dataset compare four variants—Fixed UNet, Adaptive UNet, Fixed PCP-UNet, and Adaptive PCP-UNet—reporting SSIM and NRMSE boxplots for acceleration rates 8x, 16x, and 24x, three sampling patterns, and multiple contrasts. The authors conclude that Adaptive PCP-UNet significantly improves image quality and outperforms the 'conventional ML-based method.'","tokens_in":6091,"tokens_out":4155,"duration_ms":36601,"significance":"If substantiated, the ability to reconstruct a wide range of CMR protocols with a single trained network would address a real clinical bottleneck, since protocol-specific training is impractical in many clinical settings. The paper uses a large public in vivo dataset, evaluates multiple contrasts and sampling patterns, and includes component-wise ablation among its four variants, which is a useful experimental design. However, the current evidence does not yet establish the central claim: the key adaptive-unrolling schedule is not specified, and the comparison set does not include the 'conventional ML-based method' named in the abstract. The claimed improvements are therefore not yet fully supported.","major_comments":[{"comment":"The abstract claims the proposed model 'outperforms the conventional ML-based method,' but no such baseline appears in Section 3. The four evaluated methods are all variants of the proposed unrolled network (Fixed/Adaptive UNet and Fixed/Adaptive PCP-UNet). A protocol-specific network, a standard published unrolled network such as MoDL or VarNet, or a compressed-sensing baseline is needed to support the claim of outperforming existing ML-based reconstruction. Without such a comparison, the central 'foundation model' claim is not established.","section":"Abstract and Section 3"},{"comment":"The adaptive-unrolling schedule is the paper's linchpin but is never specified. The text states that 'images with the highest acceleration rate go through all sixteen UI' and that lower-acceleration images use 'the latter UI only,' yet the exact mapping from the six evaluated acceleration rates (4, 8, 12, 16, 20, 24) to the entry UI is never reported. Moreover, because each UI has its own network and conjugate-gradient parameters, the cascade is not iterating a single optimization to convergence; hence the compressed-sensing conditioning intuition in Section 2.2 does not directly transfer. Please report the full entry-UI schedule and provide an ablation that varies the mapping while holding total network capacity fixed.","section":"Section 2.2"},{"comment":"The Adaptive-versus-Fixed comparison is confounded. In the Adaptive condition, images of different acceleration rates use different numbers of active unrolled iterations, so the comparison changes both the per-rate architecture depth and the allocation of training samples across subnetworks. The reported SSIM gain of Adaptive over Fixed cannot therefore be attributed specifically to matching iteration count to conditioning. A matched-capacity comparison—for example, Fixed networks with the same per-rate UI counts, or Adaptive networks with shuffled entry schedules—is needed to isolate the mechanism.","section":"Figure 6"},{"comment":"The statement that 'paired t-tests show that the SSIM for each method is statistically different (P = 0.05)' is not accompanied by test statistics, multiplicity-corrected p-values, confidence intervals, or effect sizes. With 4,158 validation images and many subgroup comparisons (3 patterns x 3 rates x multiple contrasts), uncorrected paired t-tests are insufficient. Please provide a quantitative table with means, standard deviations, and a corrected significance analysis.","section":"Section 3.1"},{"comment":"Several training and reconstruction hyperparameters are omitted: the data-consistency weight mu in Eqs. (3)-(4), the number of conjugate-gradient iterations and its tolerance, the optimizer, loss function, learning rate, batch size, number of epochs, and the channel-shift replica configuration (how many shifts and along which axes for each sampling pattern). These details are necessary to reproduce the results and to assess the claim that channel-shifting has only 'minor computational overhead.'","section":"Section 2.5"}],"minor_comments":[{"comment":"There is a typographical error: 'up to sixteeen UI' should read 'up to sixteen UI.'","section":"Section 2.2"},{"comment":"The phrase 'Fixed amount of UP' should be 'Fixed amount of UI.'","section":"Section 2.5"},{"comment":"The text says 'Fixed UNet exhibits the lowest SSIM and PSNR,' but Figure 6 reports NRMSE, not PSNR; please make the metric names consistent.","section":"Section 3.1"},{"comment":"The abbreviation 'NMRSE' appears in the figure and text; the standard abbreviation is NRMSE, and the paper should use one form throughout.","section":"Figure 6"},{"comment":"The phrase 'which we refereed as Pattern and Contrast-Prompt-UNet' should read 'which we refer to as Pattern and Contrast-Prompt-UNet.'","section":"Section 2.3"},{"comment":"There is a typo: 'Any mistmatch' should read 'Any mismatch.'","section":"Introduction"},{"comment":"The phrase 'a wide range if CMR image contrast' should read 'a wide range of CMR image contrasts.'","section":"Conclusion"},{"comment":"The boxplots are difficult to read without numerical axis annotations; please add a supplementary table or annotate the medians and interquartile ranges.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a medical imaging venue, and the public CMRxRecon dataset is a genuine strength. The main concern is that the central claim is under-supported by the missing external baseline and the unspecified adaptive-unrolling schedule; these are fixable with additional experiments and detailed reporting, so I recommend major revision rather than rejection. The 'foundation model' framing is stronger than the evidence currently presented, and the editors may wish to ask the authors to temper that claim or provide the requested comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful empirical paper on a real deployment problem, but it overclaims in the abstract and the marquee component—adaptive unrolling—is not actually tested as a mechanism.\n\nWhat's new and what works: the idea of one network handling mixed contrasts, three sampling patterns, and acceleration rates from 8x to 24x is genuinely practical. The PCP-UNet prompt extension is a sensible incremental step beyond Prompt-UNet, and the channel-shifting trick for enlarging receptive field is simple and cheap—the kind of thing that might be worth stealing. Experiments use the public CMRxRecon dataset, and the internal comparisons are consistent: adaptive PCP-UNet beats the fixed and prompt-free variants across rates, patterns, and contrasts. That is real evidence, and the paired t-tests are not doing much work on thousands of samples, but the ordering itself is credible.\n\nThe soft spots, in proportion: the abstract says the method \"outperforms the conventional ML-based method,\" but no such baseline appears anywhere in the experiments. There is also no protocol-specific network trained per contrast/rate/pattern, which is the natural bar for a foundation model claim. That is the biggest gap. Second, the adaptive-unrolling schedule is never specified: no mapping from acceleration rate to starting UI, no ablation. The stress-test note is right that each UI has independent network and CG parameters, so this is not iterating to convergence of one optimization; the compressed-sensing conditioning story is a loose analogy. And the adaptive-vs-fixed comparison changes both the number of active UIs and how training samples are allocated, so the SSIM gain cannot be attributed to conditioning-matched adaptivity. That doesn't kill the empirical result, but it means the paper's central novelty is under-supported. Third, no code, no hyperparameters, and evaluation appears to be on the validation split only—reproducibility will be limited.\n\nWho this is for: CMR reconstruction researchers, and anyone working on multi-protocol MRI reconstruction. It is not a methods breakthrough, but it is a credible engineering contribution on an important problem.\n\nIf I were the editor, I would send it to peer review, with major revision expected. The authors need to add the missing baselines (especially a protocol-specific model and a conventional compressed-sensing or classical ML method), report the adaptive schedule, and run a proper ablation that isolates the schedule from capacity allocation. The core idea is worth engaging with; the current write-up just doesn't yet back the full claim.","headline":"Plausible empirical result for multi-protocol CMR reconstruction, but the abstract oversells it and the adaptive-unrolling mechanism is untested.","tokens_in":6682,"tokens_out":1808,"would_cite":false,"duration_ms":20261,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single adaptive, prompt-based network reconstructs cardiac MRI across mixed protocols without retraining per protocol.","keywords":["cardiac MRI reconstruction","adaptive unrolling","foundation model","pattern and contrast prompt","channel-shifting","undersampled k-space","SSIM","NMRSE"],"falsifier":"Train a fixed-iteration PCP-UNet with the same total number of unrolled iterations, then compare its SSIM against Adaptive PCP-UNet separately at each acceleration rate; if the fixed model matches or beats the adaptive model at low rates, the adaptive schedule is not doing the claimed work. A second check is to compute the condition number of $E_{\\Omega}$ for each rate and show that it varies enough to justify a different iteration count.","tokens_in":5691,"feed_emoji":"🫀","tokens_out":9785,"duration_ms":77270,"temperature":0.7,"pith_summary":"This paper proposes a single machine-learning model for cardiac MRI reconstruction that works across a wide range of image contrasts, k-space sampling patterns, and acceleration rates, rather than requiring one trained network per acquisition protocol. The central claim is that combining adaptive unrolling (letting each image enter the reconstruction cascade at a point set by its acceleration rate), a Pattern and Contrast-Prompt-UNet (PCP-UNet), and channel-shifting gives the best reconstruction quality among the tested methods. Across in vivo experiments with mixed contrasts, sampling patterns, and acceleration rates, the authors report that Adaptive PCP-UNet achieved the highest structural similarity and the lowest NMRSE of the four compared designs. If the claim is correct, it points toward a general-purpose CMR reconstruction network that could be deployed without knowing the exact protocol in advance.","feed_headline":"One adaptive model reconstructs cardiac MRI across all tested protocols","feed_subtitle":"Pattern and contrast prompts plus acceleration-based iterations beat fixed-iteration networks at every rate.","key_machinery":"The load-bearing mechanism is the acceleration-aware unrolled cascade: up to sixteen unrolled iterations, each with its own neural regularizer and conjugate-gradient data-consistency solver, and an entry point selected by the acceleration rate. PCP-UNet is the regularizer: it concatenates image-domain features with a contrast prompt and a sampling-pattern prompt built from statistical features of the k-space mask, letting one set of weights adapt to different contrasts and undersampling geometries. Channel-shifting widens the receptive field by concatenating circularly shifted copies of the input along the channel dimension, which the paper argues helps distinguish point-spread functions of different sampling patterns at a small computational cost.","core_discovery":"The paper claims that one unrolled reconstruction network can handle mixed CMR protocols without retraining. It frames the task as the regularized problem $\\arg\\min_x \\|E_{\\Omega} x - y\\|_2^2 + R(x)$, solved by variable splitting whose iterations are unrolled into a neural cascade. The adaptive component sends an undersampled image into the cascade at a starting unrolled iteration chosen by its acceleration rate, so the highest-rate images pass through all sixteen iterations and lower-rate images pass through only the latter ones, with each iteration's weights trained for a specific artifact level. The regularizer is PCP-UNet, which adds a k-space sampling-pattern prompt to the contrast prompt of Prompt-UNet, and channel-shifting concatenates circularly shifted copies of the input to widen the receptive field. In the reported in vivo experiments, Adaptive PCP-UNet had the highest SSIM and lowest NMRSE across uniform, Gaussian random, and radial sampling and across the tested acceleration rates and contrasts, with paired t-tests showing statistically significant differences between methods.","pith_inferences":["The acceleration-to-starting-iteration mapping is never spelled out, so the adaptive gain could be tested directly by training a fixed-iteration PCP-UNet with the same per-iteration weights and comparing per-rate SSIM; the paper does not include that ablation.","The pattern-prompt idea should transfer beyond cardiac MRI to other anatomies or to non-Cartesian trajectories, because the statistical mask features are expressed in k-space coordinates rather than anatomy-specific image content.","A full computational-cost comparison, including total inference time per acceleration rate and memory use of the channel-shifting input, would quantify the practical benefit of adaptive unrolling more directly than SSIM alone.","The method's 'foundation model' framing is currently a single-task reconstruction model; its generality is demonstrated only across contrasts and sampling patterns within the same cardiac anatomy, so calling it a foundation model requires further evidence of transfer across anatomies and tasks."],"forward_implications":["A single Adaptive PCP-UNet model can replace a suite of protocol-specific networks, since it maintains high SSIM across mixed contrasts, sampling patterns, and acceleration rates in the same training run.","Low-acceleration scans can be reconstructed with fewer unrolled iterations, which shortens the inference path for the most common clinically requested rates.","The pattern prompt's statistical mask features are sufficient to separate uniform, Gaussian random, and radial sampling patterns, as shown by the improvement of PCP-UNet over the contrast-only Prompt-UNet.","Because paired t-tests at P=0.05 found each method statistically different, the ranking in SSIM is reproducible on this dataset, though the paper does not translate these differences into a clinical acceptance threshold."],"supporting_citations":[{"why":"Supplies the unrolled variational network architecture and the data-consistency formulation that the cascade extends.","marker":"[1]"},{"why":"Provides the compressed-sensing principle that higher acceleration rates need more unrolled iterations, the basis of adaptive unrolling.","marker":"[15]"},{"why":"The Prompt-UNet base that PCP-UNet extends by adding a k-space sampling-pattern prompt.","marker":"[16]"},{"why":"Supplies the in vivo cardiac MRI data used for training and validation.","marker":"[17]"},{"why":"Used to estimate coil sensitivity profiles for the forward acquisition model.","marker":"[18]"}],"fun_headline_variants":["One adaptive model, all cardiac MRI protocols","Cardiac MRI reconstruction without retraining per protocol","Adaptive unrolling network beats fixed networks on every MRI test","PCP-UNet: one network for all CMR sampling patterns and contrasts","From uniform to radial: one model tops cardiac MRI reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the number of unrolled iterations is the right control for the conditioning of the forward operator, but the paper never states the mapping from acceleration rate to starting iteration or demonstrates that this mapping is optimal.","fun_headline_variants_meta":{"raw":{"variants":["One adaptive model, all cardiac MRI protocols","Cardiac MRI reconstruction without retraining per protocol","Adaptive unrolling network beats fixed networks on every MRI test","PCP-UNet: one network for all CMR sampling patterns and contrasts","From uniform to radial: one model tops cardiac MRI reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1312,"prompt_tokens":977,"completion_tokens":335,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":253}},"tokens_in":593,"tokens_out":335,"duration_ms":4106,"temperature":1.0,"reasoning_tokens":253,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:39:22.817529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a fixed-iteration PCP-UNet with the same total number of unrolled iterations, then compare its SSIM against Adaptive PCP-UNet separately at each acceleration rate; if the fixed model matches or beats the adaptive model at low rates, the adaptive schedule is not doing the claimed work. A second check is to compute the condition number of $E_{\\Omega}$ for each rate and show that it varies enough to justify a different iteration count.","supporting_citations":[{"cited_title":"and Lustig,M., 2014","cited_arxiv_id":null,"evidence_quote":"Used to estimate coil sensitivity profiles for the forward acquisition model."},{"cited_title":"and Metaxas, D.N., 2023, October","cited_arxiv_id":null,"evidence_quote":"The Prompt-UNet base that PCP-UNet extends by adding a k-space sampling-pattern prompt."},{"cited_title":"and Knoll,F.,2018.LearningavariationalnetworkforreconstructionofacceleratedMRI data","cited_arxiv_id":null,"evidence_quote":"Supplies the unrolled variational network architecture and the data-consistency formulation that the cascade extends."},{"cited_title":"and Pauly, J.M., 2007","cited_arxiv_id":null,"evidence_quote":"Provides the compressed-sensing principle that higher acceleration rates need more unrolled iterations, the basis of adaptive unrolling."}],"review_version":1}