{"id":"5d2432f4-11eb-4859-ab5f-38ae66815219","arxiv_id":"2607.21737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Adaptive QUBO-based k-space line selection improves reconstruction metrics over static Cartesian sampling on simulated 8-coil MRI data, while a D-Wave hybrid implementation only matched variable-density Poisson-disc sampling.","lead":"A new MRI acceleration method chooses which k-space lines to measure next by solving an optimization problem that can also run on quantum-annealing hardware. In simulations it improved reconstructed image quality over fixed sampling patterns, but it has not been validated on a real scanner and needs stronger statistical and reproducibility support.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the adaptive field; reported gains could come from static center-bias/repulsion terms, leaving the central 'adaptive' claim untested.","rationale":"The reader's weakest-assumption points to the signal-energy proxy in the adaptive field. I agree that the adaptive mechanism is the fragile pillar, but the more precise and more load-bearing problem is that the adaptive field's contribution is never isolated. The full QAS objective includes several non-adaptive terms that could plausibly explain the improvement over VDP; without an ablation, even a perfect signal-energy proxy might be irrelevant to the result. The D-Wave experiment, though honestly discussed, does not provide supporting evidence because the pooled QAS mask was not better than VDP. The missing error bars and test-set hyperparameter tuning are additional reasons for caution, but the ablation is the decisive missing experiment. A clean w_a=0 comparison would settle whether the word 'adaptive' in the central claim is justified, or whether the contribution is just a static QUBO sampling pattern. Since the reader already conditioned acceptance on missing validation, this concern reinforces, rather than changes, the CONDITIONAL verdict.","tokens_in":758,"tokens_out":727,"duration_ms":115879,"concrete_test":"Run an ablation on the 256^3 phantom at R=5 and R=10: use exactly the same QUBO solver, batch size, center bias, repulsion, exploration bonus, and reconstruction pipeline, but set the adaptive weight w_a=0 in Eq. 25 (i.e., skip the measured-signal update in Algorithm 1). Compare full QAS, ablated static-QUBO, and VDP over at least 10 random seeds, with hyperparameters chosen on a separate tuning phantom. If full QAS does not significantly beat the w_a=0 variant on PSNR/SSIM/HFEN, the adaptive field is not load-bearing and the central 'adaptive' claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"QAS's objective (Eqs. 25-26) combines a static center bias h_static, pairwise repulsion J, exploration bonus b, and the adaptive signal-energy term w_a h_adaptive. The main experiments compare the full QAS mask against static baselines (VDP, radial, spiral, etc.), which conflates the adaptive mechanism with the non-adaptive QUBO geometry. No experiment sets w_a=0 or suppresses the feedback update in Algorithm 1, so the reported PSNR/SSIM/NMSE/HFEN gains cannot be attributed to adaptation. If a static version of the same QUBO (center bias + repulsion + exploration only) performs equally, the paper's central claim of adaptive sensing collapses to 'a particular static QUBO mask is good.' The reduced-pool D-Wave experiment does not resolve this: it also includes the adaptive field, and Table 3 actually shows VDP matching or beating QAS on most metrics. Hyperparameters were also grid-searched on the same test phantoms, further clouding attribution. Thus the load-bearing premise — that measured-signal feedback drives the improvement — is untested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quantum Adaptive Sampling (QAS), a sequential Cartesian k-space sampling method that selects batches of phase-encode lines by minimizing a QUBO/Ising objective. The objective (Eqs. 25-26) combines a static center-bias attraction h_static, an adaptive field h_adaptive derived from log-compressed, smoothed L2 norms of previously acquired k-space lines, an exploration bonus, and pairwise repulsion terms, with a quadratic cardinality penalty. The authors evaluate QAS retrospectively on simulated 8-coil brain phantoms at 128^3 and 256^3 resolution, 20% and 10% sampling, and several input SNR levels, using parallel tempering as the QUBO solver and SENSE-TV ADMM reconstruction. They report that QAS improves PSNR, SSIM, NMSE, and HFEN relative to static Cartesian strategies such as variable-density Poisson-disc sampling. A reduced-pool implementation using D-Wave's hybrid solver gives reconstruction quality comparable to variable-density Poisson-disc sampling, which the authors correctly interpret as a feasibility demonstration rather than quantum advantage. The main claimed contribution is the adaptive QUBO framework rather than a quantum speedup.","tokens_in":15987,"tokens_out":8107,"duration_ms":73655,"significance":"If the central claim were fully supported, the paper would make a useful contribution: it provides a concrete, solver-independent QUBO encoding for adaptive Cartesian MRI sampling, with pseudocode for the feedback loop, and it is appropriately cautious about quantum advantage. The strengths include the explicit algorithmic description, the fair-budget comparison across multiple resolutions, accelerations, and noise levels, and the transparent discussion of clinical translation and hardware limitations. However, the paper's headline result depends on the adaptive field being the cause of the improvement, and that causal claim is not tested by any ablation. The empirical evidence is also weakened by the absence of error bars, the likely tuning of many free parameters on the same test phantoms, and an internally inconsistent D-Wave table. These issues are fixable within the scope of a revision, but they are load-bearing for the current claims.","major_comments":[{"comment":"The central claim is that signal-energy feedback drives the improvement, but no experiment isolates the adaptive field. The objective is H_total = Σ J_ij s_i s_j − Σ (h_static + w_a h_adaptive + b_i) s_i + P(Σ s_i − M)^2, and Algorithm 4 updates h_adaptive from measured k-space energy. The main comparisons in Tables 1-2 and Figs. 1-3 use the full objective, so an improvement over VDP could come from the static center bias, the repulsion term, the exploration bonus, or the QUBO joint selection, without any contribution from h_adaptive. I request an explicit ablation: set w_a = 0 (or freeze/randomize h_adaptive) while keeping all other terms identical, and compare against the full QAS. A static QUBO mask with the same center-bias, repulsion, and exploration terms should also be included. Without this, the title and abstract's 'adaptive' claim is untested.","section":"Mathematical Formulation, Eqs. (25)-(26); Algorithm 4; Results"},{"comment":"The methods state that 'grid search algorithm was applied to determine the suitable parameters' and specify α, γ, and β, but not the ranges, the number of trials, or whether tuning was performed on the same 128^3 and 256^3 phantoms used for evaluation. The method also has additional free parameters (w_a, η, σ_g, c_b, τ, P) whose values are not given. If the same phantoms were used for both tuning and evaluation, the reported gains are optimism-biased and cannot be assumed to generalize to unseen anatomies. Please provide a data-split description (e.g., tune on one phantom/contrast, test on a held-out phantom), report the chosen parameter values, and include a sensitivity analysis.","section":"Methods, hyperparameter selection; Tables 1-2"},{"comment":"All quantitative results appear to be single realizations. Since mask generation, noise injection, and parallel tempering are stochastic, and some reported differences are small (e.g., Table 2, R=10: QAS PSNR 27.57 vs VDP 26.93; SSIM 0.8542 vs 0.8280), the comparison needs repeated runs with mean ± standard deviation or confidence intervals and, where appropriate, significance tests. Without error bars, the robustness of the claimed improvement, especially at R=10 on the 128^3 phantom, is not established.","section":"Results, Tables 1-2 and Figs. 1-2"},{"comment":"Table 3 is internally inconsistent. For the same reference phantom, Eqs. (30) and (32) imply that PSNR and NMSE should be monotonically related: lower NMSE should give higher PSNR. The noiseless row reports QAS NMSE=0.0533 with PSNR=29.653 and VDP NMSE=0.2259 with PSNR=29.845, so the method with higher error has higher PSNR. At 20 dB input SNR, QAS NMSE increases from 0.0533 to 0.2258 while PSNR slightly increases from 29.653 to 29.852, which is impossible under the stated definitions unless the reference energy or dynamic range changed between rows. In addition, the D-Wave experiment uses a Shepp-Logan phantom, whereas the main experiments use an MPRAGE brain phantom, so the reduced-pool result is not directly comparable to Tables 1-2. These issues must be corrected and the table recomputed before the D-Wave feasibility claim can be interpreted.","section":"D'Wave Experiments and results, Table 3"}],"minor_comments":[{"comment":"Eq. (26) and Eq. (27) are identical; one should be removed. Algorithm 3's title contains the stray text 'candidate pool 3.0.1'; this appears to be a formatting artifact.","section":"Mathematical Formulation; Methods"},{"comment":"The column header uses 'NRMSE' while the text and Eq. (30) define 'NMSE'. Please use consistent terminology.","section":"Table 3"},{"comment":"The conclusion states that 'images can be recovered from randomly undersampled data,' but the proposed method is deterministic and adaptive, not random. This sentence should be rephrased to match the actual method.","section":"Conclusion"},{"comment":"The 'decreasing center bias' branch is described only verbally. An explicit formula for how h_static decreases with sampling progress, and the role of τ, would improve reproducibility.","section":"Algorithm 4"},{"comment":"There is no code or data availability statement. Given the number of free parameters and the empirical nature of the claims, releasing the mask-generation and reconstruction code would substantially strengthen the manuscript.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The D-Wave table inconsistency in Table 3 is the most concrete technical red flag; it should be resolved before any further consideration. The missing adaptive-field ablation is the central methodological gap: the paper's title and abstract promise adaptation, but the experiments only compare the full algorithm against static baselines. I believe both are fixable in revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2607.21737. First, the paper ships a concrete new formulation: sequential Cartesian k-space line selection as a fixed-cardinality QUBO with static center bias, signal-energy adaptivity, pairwise repulsion, and an exploration bonus, solved with parallel tempering and, in a reduced-pool experiment, on D-Wave hardware. That integrated package is genuinely new, even if each ingredient is standard. Second, the paper's headline claim—that the adaptive feedback drives the reconstruction gains—is not actually tested. There is no ablation that sets the adaptive weight to zero or turns off the feedback loop. The full method beats VDP and others, but a static QUBO with center bias and repulsion alone might do just as well. The stress-test note is right, and I did not find anything in the methods or results that answers it.\n\nThe paper does a lot well. The authors are honest: they explicitly say no quantum advantage is claimed, they discuss hardware limits, pooling artifacts, and the need for prospective validation. The experiment design is broad—two resolutions, two accelerations, multiple noise levels, four metrics—and the reconstruction pipeline (SENSE-TV with fixed regularization) is held constant across methods, so the comparison is fair on that axis. The D-Wave experiment, while only matching VDP, at least demonstrates feasibility and the limitations are acknowledged in the text.\n\nThe soft spots are proportionate to their impact. The missing ablation is load-bearing; without it, the title's \"adaptive\" is a hypothesis, not a result. Hyperparameters were grid-searched on the same test phantoms, so some of the gain is fit. There are no error bars or multiple realizations, which matters because the reported improvements at R=10 (e.g., 0.64 dB PSNR on 128³) could easily be noise. Table 3 has an internal inconsistency: the D-Wave mask's NRMSE jumps from 0.0533 (noiseless) to 0.2258 (20 dB) while PSNR and SSIM stay essentially flat—that does not behave like a real reconstruction curve. No code or data is released, so reproducibility is limited.\n\nWho should read this? People working on sampling-pattern optimization for MRI and researchers interested in practical QUBO formulations on quantum hardware. The paper deserves a serious referee because the formulation is new, the limitations section is unusually candid, and the gap between claim and evidence is fixable with an ablation study and more careful statistics. I would not cite it as evidence that adaptive sampling works until that ablation appears.\n\nRecommendation: send it to peer review, but the referee should push hard for a w_a=0 ablation, error bars, and a corrected D-Wave table.","headline":"A serious but unproven adaptive-sampling claim: the integrated QUBO formulation is new and honestly reported, but the paper never isolates the adaptive term, so the central mechanism is untested.","tokens_in":16404,"tokens_out":1273,"would_cite":false,"duration_ms":14173,"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":"Adaptive QUBO-based k-space selection improves accelerated MRI reconstruction over static sampling, with a solver-agnostic formulation ready for quantum-annealing hardware.","keywords":["adaptive MRI sampling","compressed sensing","QUBO","quantum annealing","k-space sampling","Cartesian phase encoding","SENSE-TV reconstruction","variable-density Poisson-disc"],"falsifier":"Run the same reconstruction pipeline on held-out multi-coil data from an anatomy or contrast not used in hyperparameter tuning, comparing the adaptive QUBO mask against variable-density Poisson-disc with equal solver budget and identical reconstruction settings; if the PSNR and SSIM differences shrink below the paper's reported margins, the adaptive field has likely overfit to the phantom family.","tokens_in":1328,"feed_emoji":"🧲","tokens_out":1245,"duration_ms":41008,"temperature":0.7,"pith_summary":"The paper tries to show that accelerated MRI can be improved by choosing which k-space lines to measure adaptively, instead of using a fixed sampling pattern decided before the scan. It casts each batch of line selections as a quadratic unconstrained binary optimization (QUBO) problem whose objective balances central k-space preference, signal energy from previously acquired lines, and spatial dispersion. In retrospective eight-coil three-dimensional simulations, this adaptive selection improved PSNR, SSIM, NMSE, and HFEN relative to static Cartesian strategies, including variable-density Poisson-disc sampling, at both 20% and 10% sampling. The same QUBO ran on a quantum-classical hybrid system and matched variable-density Poisson-disc when the candidate pool was reduced; the authors state clearly that this does not prove quantum advantage. The reason to care is that feedback-driven, optimization-based sampling could let scanners get more useful image content from each measured point, especially when sampling budgets are tight.","feed_headline":"Adaptive QUBO sampling beats static MRI masks at 10% sampling","feed_subtitle":"Live signal energy steers each k-space line choice, improving reconstruction quality when sampling is scarce.","key_machinery":"The central object is a fixed-cardinality QUBO over binary variables that mark candidate k-space phase-encode lines as selected or not. Its linear coefficients encode three attractions: a Gaussian static central bias; an adaptive field obtained by log-compressing the L2 norm of previously acquired k-space lines, diffusing it with a Gaussian kernel, and suppressing it with a hyperbolic tangent; and an exploration bonus favoring un-sampled neighbors. The quadratic coefficients are an anisotropic, softened power-law repulsion between candidate lines, and a penalty term enforces exactly M selections per batch. A Markovian outer loop feeds measured signal energy from each completed batch into the","core_discovery":"The central claim is that casting sequential Cartesian phase-encode-line selection as a fixed-cardinality QUBO yields masks that reconstruct better than static heuristic masks. The QUBO combines a static attraction to the k-space center, an adaptive field built from log-compressed, smoothed L2 norms of already acquired k-space lines, pairwise repulsion that spreads samples, and a penalty that enforces an exact number of lines per batch. The authors report that on simulated eight-coil three-dimensional data with SENSE-TV reconstruction, the method improved PSNR, SSIM, NMSE, and HFEN relative to variable-density Poisson-disc and other baselines at 20% and 10% sampling, with larger gains at hig","pith_inferences":["The signal-energy proxy driving the adaptive field could be swapped for an information-based or learned criterion through the same QUBO interface; the paper itself notes this is a drop-in replacement, leaving the extension open.","If the adaptive field is the true source of the gain, the method's value is not tied to quantum hardware; a testable extension is to compare the full method against a version whose adaptive field is replaced by random noise while keeping the QUBO structure intact.","The same batch-selection QUBO could be adapted to non-Cartesian trajectories or to jointly optimize sampling across time frames in dynamic imaging, where k-space energy shifts over time, although neither is explored here.","Because hyperparameters were grid-searched on the same test phantoms, an honest reader should treat the reported margin as an upper bound until validated on held-out anatomies."],"forward_implications":["If the central claim holds, adaptive QUBO masks can replace static variable-density sampling in Cartesian MRI without requiring new non-Cartesian readouts.","Because the formulation is solver-independent, gains measured with parallel tempering can transfer to any future quantum-annealing backend with enough qubits, connectivity, and precision to handle the full candidate pool.","The reported advantage is concentrated at low sampling rates and moderate noise, so the practical payoff is in high-acceleration scans where static masks lose the most fidelity.","The paper's own timing breakdown suggests mask optimization need not be the bottleneck, but real-time scanner integration still requires asynchronous, warm-started solvers and a precomputed fallback batch.","Clinically, the same masks must be tested prospectively on human data across varied anatomy, contrast, field strength, and coil arrays before translation."],"fun_headline_variants":["Adaptive QUBO sampling outperforms static MRI masks at low rates","Sequential QUBO picks beat fixed masks in accelerated MRI","QUBO-optimized k-space lines boost MRI quality at 10% sampling","Hybrid quantum-classical QUBO matches static MRI sampling in test","Adaptive line selection via QUBO improves undersampled MRI"],"cache_read_input_tokens":17792,"weakest_assumption_plain":"The whole adaptive feedback loop rests on the assumption that the log-compressed, smoothed L2 norm of previously acquired k-space lines predicts where additional samples will carry reconstruction information; if that proxy fails, the method's edge over a well-tuned static variable-density mask can disappear, as it nearly did in the reduced-pool quantum experiment.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive QUBO sampling outperforms static MRI masks at low rates","Sequential QUBO picks beat fixed masks in accelerated MRI","QUBO-optimized k-space lines boost MRI quality at 10% sampling","Hybrid quantum-classical QUBO matches static MRI sampling in test","Adaptive line selection via QUBO improves undersampled MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1541,"prompt_tokens":771,"completion_tokens":770,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":674}},"tokens_in":515,"tokens_out":770,"duration_ms":7727,"temperature":1.0,"reasoning_tokens":674,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:50:10.015511+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same reconstruction pipeline on held-out multi-coil data from an anatomy or contrast not used in hyperparameter tuning, comparing the adaptive QUBO mask against variable-density Poisson-disc with equal solver budget and identical reconstruction settings; if the PSNR and SSIM differences shrink below the paper's reported margins, the adaptive field has likely overfit to the phantom family.","supporting_citations":[],"review_version":1}