{"id":"c01b507a-7df0-44f0-80da-3f19adfcd4b4","arxiv_id":"2502.02771","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Diffusion priors outperform classical priors in sparse-view CT only below about 15 projections, then plateau while classical methods catch up and surpass them.","lead":"In low-dose chest CT, diffusion model priors beat classical reconstruction only when very few X-ray projections are available, roughly 10 to 15, after which classical methods win. The finding matters for clinical imaging because it marks where AI-generated realistic images help and where they may mislead in fat quantification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed crossover and plateau depend on hyperparameters fixed by tuning at 360 projections; if diffusion guidance weight or classical regularization were re-optimized per nproj, the interval and plateau could shift, so the comparison is not yet evidence about the priors themselves.","rationale":"This is the most load-bearing concern because it directly challenges both parts of the central claim: the low-projection superiority of diffusion priors and the high-projection plateau. The reader's weakest assumption identified the same risk—hyperparameters chosen at 360 projections may not be representative across nproj—and the proposed conditional verdict is appropriate. I considered whether the use of a specific, naive gradient-guidance scheme (rather than diffusion priors in general) is an even more fundamental issue; however, that concern is essentially a manifestation of the hyperparameter sensitivity, since the guidance weight is the key control for how strongly data consistency is enforced. The paper is honest about the plateau's possible cause, but that does not resolve the confound: the experimental design fixes λ at a single value, so the observed trends cannot disentangle the prior's capability from the chosen algorithm's tuning. The proposed concrete test—per-nproj hyperparameter re-optimization—is tractable, directly addresses the confound, and would settle whether the central numeric claims (interval bounds, plateau level) survive. If the qualitative pattern persists after fair tuning, the paper's contribution is solid; if not, the conclusions would need to be substantially revised. Given the current evidence, the CONDITIONAL verdict is right: the findings are useful but not yet robust until this test is performed. No ad hominem is intended; this is a methodological design issue, not an attribution of fraud or incompetence.","tokens_in":4949,"tokens_out":7587,"duration_ms":73413,"concrete_test":"Re-run the main experiments at a subset of projection counts (e.g., nproj ∈ {3, 5, 10, 15, 25, 40, 60, 120, 360}) while performing a hyperparameter sweep at each nproj for both diffusion guidance weight λ (e.g., {0.001, 0.01, 0.1, 1, 10}) and classical regularization (e.g., {1, 10, 100}). Select the best-performing hyperparameters per method per nproj (using a validation set from the training cohort or cross-validation), then recompute the performance curves, the plateau location, and the interval I from Table 1. If the diffusion plateau disappears or the upper bound of I shifts substantially (e.g., beyond 25 projections), the central claim is an artifact of fixed hyperparameters; if the pattern persists under per-nproj tuning, the claim is empirically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that diffusion priors are statistically better only for ~5–15 projections and plateau afterward—requires that the compared methods are fairly configured at every projection count. Section 2 states: \"We performed a grid search to determine the best optimization, regularization, and guidance parameters for L1 (λ = 10), L2 (λ = 10) and diffusion (λ = 0.1). These corresponded to the lowest mean squared error at 360 projections.\" This creates a double confound. First, the diffusion guidance weight λ=0.1 is chosen at 360 projections. As nproj increases, the data-consistency gradient carries more information, so the optimal λ should generally increase to trust the data more; a fixed small λ may artificially cap diffusion performance, producing the observed plateau. Second, the classical regularization strengths (λ=10) are also fixed at 360 projections; at very low nproj, such strong TV/L2 smoothing may be over-regularizing, inflating diffusion's apparent advantage. Thus both ends of the interval I in Table 1—the low-nproj superiority and the high-nproj crossover—could be artifacts of hyperparameters tuned at a single projection count rather than intrinsic properties of the priors. The paper's own discussion attributes the plateau to \"non-optimal convergence\" of naive gradient guidance, which is precisely the mechanism controlled by λ; a larger λ might allow the data term to dominate and eliminate the plateau. Without per-nproj hyperparameter re-optimization, Fig. 2 and Table 1 cannot support the paper's general conclusions about when diffusion priors are helpful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares diffusion-prior sparse-view CT reconstruction against classical L1/TV and L2/Tikhonov regularization over a wide range of projection counts, using PSNR, SSIM, fat-content accuracy, and fat Dice on a 206-patient test set. The authors report that diffusion priors are statistically superior only for an extremely small number of projections (roughly 5–40 depending on the metric), that classical priors become superior once projections are 'sufficient', and that diffusion performance plateaus after about 10–15 projections. They introduce an interval-based definition of statistical superiority based on Mann-Whitney U tests and discuss clinical implications for low-dose thoracic fat quantification.","tokens_in":5285,"tokens_out":5327,"duration_ms":51771,"significance":"If the empirical comparison is accepted, the paper makes a valuable contribution by providing a cautionary calibration for the sparse-reconstruction community: it uses a large clinical dataset and downstream fat-segmentation metrics to challenge the assumption that diffusion priors are universally superior, and it highlights failure modes such as performance plateaus and hallucinated detail. The multi-metric evaluation and the explicit focus on a clinically relevant downstream task are strengths. However, the central claims about the crossover point and the plateau are only as strong as the fairness of the compared configurations at every projection count, and the current hyperparameter and statistical-testing choices leave this unresolved.","major_comments":[{"comment":"The hyperparameters for all three methods are fixed by a grid search that minimized MSE at 360 projections (L1 λ=10, L2 λ=10, diffusion λ=0.1). This single-point tuning is load-bearing for the central claims in Section 3 and Table 1: at low nproj the classical regularization strength may be over-smoothing, which would inflate the diffusion advantage, while at high nproj the diffusion guidance weight λ=0.1 may be too weak for the data-consistency gradient to dominate, which would artificially create the observed plateau. The authors' own explanation in Section 4, that naive gradient guidance gets 'stuck' balancing the prior and the external loss, is exactly the mechanism controlled by λ. The paper therefore does not yet establish that the crossover and plateau are intrinsic to diffusion priors; it establishes them for one particular configuration. A re-optimization of λ per nproj, or at least a sensitivity sweep across λ at several nproj values, is needed, and the use of a held-out validation set for this tuning should be stated explicitly.","section":"Section 2, Inference"},{"comment":"The Mann-Whitney U tests use 618 test slices, but these are nested within 206 patients (3 slices per patient). Treating all slices as independent overstates the effective sample size, so the nominal p-values in Table 1 and the resulting interval bounds are anti-conservative. A patient-level analysis, such as averaging metrics per patient before testing or using a cluster-robust permutation test, is required to support the quantitative UB values reported in Table 1.","section":"Section 2 (Dataset and Metrics); Section 3"},{"comment":"The lower bound LB is chosen as the visible 'corner' of the performance curves, which is a subjective criterion not specified quantitatively, while the upper bound is determined by p < τp. The resulting interval I therefore conflates two different statistical procedures. In addition, the U-test is applied separately at each nproj without multiple-comparison control; because adjacent nproj comparisons are highly correlated, the reported p-values are not a valid basis for claiming that the interval endpoints are exact. The authors should specify an automatic rule for LB and state whether any multiplicity correction was applied.","section":"Section 3, interval definition"}],"minor_comments":[{"comment":"The header 'LB UB (p-value)' with rows 'L1 Prior L2 Prior' is ambiguous; for the MSE row it is unclear which UB corresponds to L1 and which to L2. Use separate L1 and L2 columns.","section":"Table 1"},{"comment":"The definition of the interval I uses 'LB, UB ∈ R × R', but LB and UB are real scalars; the notation should be LB, UB ∈ R, or I ⊂ R².","section":"Section 3"},{"comment":"The label 'IQR5,95' is nonstandard; the figure is plotting the range between the 0.05 and 0.95 quantiles, so call it the 90% quantile range and define it in the caption.","section":"Figure 2"},{"comment":"The definition of fat content accuracy as 'the percentage of correct pixels classified as fat' should specify whether true negatives are included in the denominator; without this, the metric is not clearly defined as a content accuracy.","section":"Section 2, Metrics"}],"recommendation":"major_revision","confidential_remarks":"This is a useful empirical study, but the central quantitative claims rest on tuning and statistical-testing choices that need revisiting. A revised version with per-nproj sensitivity analysis and cluster-aware statistics would be worth another round; the questions raised here are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Matt,\n\nShort version: this paper will be useful to anyone working on sparse-view CT or diffusion-based reconstruction, and it deserves a proper peer review. The headline finding — that diffusion priors only beat classical L1/L2 priors in a narrow, very-few-projection window and then plateau — is clearly presented and backed by sensible pixel, structural, and downstream fat-quantification metrics on a clinically relevant task. That is genuinely new: prior work mostly looked at either very sparse or moderately sparse regimes, not a wide sweep with downstream metrics.\n\nWhat the paper does well is that it asks a clean question and reports honestly. The interval definition (I = [LB, UB]) is transparent, the Mann-Whitney tests are appropriate, and the discussion admits the plateau might be a convergence artifact rather than an intrinsic property. The figures actually support the claims as plotted.\n\nNow the soft spot, and it's a real one. The comparison fixes hyperparameters by tuning at 360 projections: L1 and L2 regularization strengths λ=10 and the diffusion guidance weight λ=0.1 are chosen for lowest MSE at 360 projections. That creates a double confound. At low nproj, strong classical regularization may be over-smoothing, inflating diffusion's apparent advantage. At higher nproj, a fixed small guidance weight may cap diffusion from leveraging extra data, producing the plateau. The authors themselves attribute the plateau to “non-optimal convergence” of gradient guidance, which is exactly what λ controls. So the crossover point and the plateau could shift if λ were re-optimized per projection count. The stress-test note is on target. This doesn't destroy the paper, but it does mean the claims about “when diffusion priors are helpful” are conditional on the tuning scheme rather than being robust properties of the priors.\n\nOther, minor weaknesses: single-cohort dataset (post-mastectomy chest wall CT only), no held-out validation for the hyperparameter selection, and the fat segmentation is a simple HU threshold, which limits the clinical realism. None of these are deal-breakers.\n\nNet: this is a well-executed empirical study with an important caveat. It should go to peer review, but the authors need to either re-tune per nproj, run a sensitivity analysis showing the conclusions are stable across reasonable λ values, or soften the claims. I'd bring it to our reading group because the methodology discussion is worth having, and I'd probably cite it as a cautionary data point.\n\nRecommendation: accept for peer review with the expectation of major revision around the hyperparameter issue.\n\nBest,\n[You]","headline":"A useful, honest empirical study with a real crossover finding, but the central comparison is undercut by hyperparameters tuned only at 360 projections.","tokens_in":5827,"tokens_out":908,"would_cite":true,"duration_ms":11189,"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":"Diffusion priors beat classical priors for sparse CT reconstruction only in the 5-to-15-projection range, then plateau.","keywords":["diffusion priors","sparse-view CT","sparse reconstruction","image reconstruction","total variation regularization","Tikhonov regularization","thoracic fat quantification","low-dose CT"],"falsifier":"Repeat the comparison with a separate grid search over the diffusion guidance weight (and classical regularization strengths) at each projection count, covering at least 5, 10, 15, 30, 60, 120, and 360 projections on the same dataset; if the diffusion metrics keep improving past 15 projections or overtake classical priors at higher counts under projection-count-specific tuning, the plateau claim as stated would need revision.","tokens_in":4783,"feed_emoji":"🩻","tokens_out":6059,"duration_ms":53273,"temperature":0.7,"pith_summary":"This paper asks when diffusion-model priors genuinely help sparse-view CT reconstruction relative to classical analytic priors. Using a large chest-wall CT dataset, the authors compare an unconditional diffusion prior to $L_1$ (total-variation) and $L_2$ (Tikhonov) regularized reconstruction across projection counts from a handful to 360, scoring pixel, structural, and downstream fat-quantification metrics. Their central finding is a crossover: diffusion priors are statistically better only for extremely few projections (roughly 5–15 in most metrics), and their performance plateaus beyond that, while classical priors keep improving and overtake them. The authors conclude that diffusion priors are attractive for ultra-low-dose imaging, but they do not extract all structural detail even when many projections are available.","feed_headline":"Diffusion CT priors win only below 15 projections","feed_subtitle":"Beyond ~15 projections classical priors outperform diffusion, whose results plateau and miss fine detail.","key_machinery":"The load-bearing mechanism is guided diffusion inference with projection-consistency gradients: at each of 50 timesteps, the predicted reconstruction is forward-projected, compared with the measured projections by mean squared error, and the gradient of that error is added back into the diffusion step with guidance weight $\\lambda=0.1$. The classical baselines are total-variation-regularized least squares ($L_1$, $\\lambda=10$) and Tikhonov-regularized least squares ($L_2$, $\\lambda=10$), each run for 1000 iterations. The crossover analysis uses the formula $I=[\\min\\{n_{\\text{proj}}: M \\ge \\tau_M\\},\\ \\max\\{n_{\\text{proj}}: p \\le \\tau_p\\}]$ for increasing metrics, where the Mann-Whitney U-test provides the significance bound, and the lower bound is read from the corner where the diffusion performance curve starts to plateau.","core_discovery":"On an unconditional 2D U-Net diffusion model trained on 128x128 CT slices and guided at inference by projection-consistency gradients, the authors identify an interval $I=[LB,UB]$ of projection counts in which diffusion priors outperform classical priors with statistical significance. The upper bounds depend on metric and baseline: for pixel MSE the interval ends at $n_{\\text{proj}}=15$ against both priors; for SSIM it ends at 25 against $L_1$ and 60 against $L_2$; for fat content it ends at 40 against both; for fat localization it ends at 25 against both. Beyond these counts, classical priors are superior on all metrics. Diffusion performance plateaus after roughly 10–15 projections rather than converging to the ground truth, and the reconstructions miss low-level structures such as vasculature even at high projection counts.","pith_inferences":["Editorial inference: the plateau may be a property of the fixed guidance schedule rather than an intrinsic limit; projection-count-dependent guidance weights could push the crossover to higher $n_{\\text{proj}}$.","Editorial inference: the same crossover analysis could be run on other anatomies and downstream tasks, such as nodule detection or bone density, to test whether the 10–15 projection boundary generalizes.","Editorial inference: the radiation-dose framing suggests a direct prospective study comparing fixed planar X-ray systems with standard CT for fat quantification, using true low-dose acquisitions rather than simulated sparse projections."],"forward_implications":["For CT reconstruction with more than a few tens of projections, classical $L_1$/TV or $L_2$/Tikhonov priors are the safer choice on pixel, structural, and downstream fat metrics.","In the 5–15 projection range, diffusion priors can deliver clinically usable fat quantification with up to roughly 97% dose reduction relative to full projection scans.","Adding projections beyond about 15 does not sharpen diffusion-prior reconstructions, so extra scans confer little image-quality benefit for this method.","Downstream fat-content and fat-localization metrics follow the same crossover pattern as pixel and structural metrics, so the trade-off is not an artifact of one scoring choice.","Because diffusion priors can look realistic while missing low-level detail like vasculature, high-stakes clinical use should not rely on perceptual quality alone."],"supporting_citations":[{"why":"Sparse-view CT diffusion reconstruction work this study extends and uses as the comparison context.","marker":"[3]"},{"why":"Supplies the classical $L_1$/$L_2$ regularization formulations used as baselines.","marker":"[4]"},{"why":"Guided diffusion methodology that the inference procedure follows.","marker":"[5]"},{"why":"Training recipe and noise schedule for the unconditional diffusion model.","marker":"[10]"},{"why":"Projection operator implementation used to compute forward projections for guidance and evaluation.","marker":"[11]"},{"why":"Plug-and-play diffusion prior inference approach on which the guided reconstruction is based.","marker":"[12]"},{"why":"Optimization analysis of gradient guidance invoked to explain the performance plateau.","marker":"[13]"},{"why":"Clinical motivation that obese patients receive higher radiation doses, supporting the low-dose framing.","marker":"[9]"}],"fun_headline_variants":["Diffusion CT priors plateau at 15 projections","Classical priors beat diffusion past 15 CT projections","Sparse CT: diffusion only helps under 15 projections","Diffusion priors lose to classical after 15 CT projections","Diffusion CT priors: benefit vanishes after 15 projections"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that hyperparameters tuned once at 360 projections, especially the diffusion guidance strength $\\lambda=0.1$, remain the fair settings at every projection count, so the observed plateau and crossover could shift if the guidance weight were re-optimized separately for each number of projections.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion CT priors plateau at 15 projections","Classical priors beat diffusion past 15 CT projections","Sparse CT: diffusion only helps under 15 projections","Diffusion priors lose to classical after 15 CT projections","Diffusion CT priors: benefit vanishes after 15 projections"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000293,"raw_usage":{"total_tokens":1706,"prompt_tokens":943,"completion_tokens":763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":681}},"tokens_in":559,"tokens_out":763,"duration_ms":7323,"temperature":1.0,"reasoning_tokens":681,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:10:15.058385+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the comparison with a separate grid search over the diffusion guidance weight (and classical regularization strengths) at each projection count, covering at least 5, 10, 15, 30, 60, 120, and 360 projections on the same dataset; if the diffusion metrics keep improving past 15 projections or overtake classical priors at higher counts under projection-count-specific tuning, the plateau claim as stated would need revision.","supporting_citations":[{"cited_title":"Regularization strategies in statistical image reconstruction of low-dose x-ray ct: A review,","cited_arxiv_id":null,"evidence_quote":"Supplies the classical $L_1$/$L_2$ regularization formulations used as baselines."},{"cited_title":"Diffusion models beat gans on image synthesis,","cited_arxiv_id":null,"evidence_quote":"Guided diffusion methodology that the inference procedure follows."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Training recipe and noise schedule for the unconditional diffusion model."},{"cited_title":"Tomosipo: fast, flexible, and convenient 3d tomography for complex scanning geometries in python,","cited_arxiv_id":null,"evidence_quote":"Projection operator implementation used to compute forward projections for guidance and evaluation."},{"cited_title":"Diffusion models as plug- and-play priors,","cited_arxiv_id":null,"evidence_quote":"Plug-and-play diffusion prior inference approach on which the guided reconstruction is based."},{"cited_title":"Increased radiation dose to overweight and obese patients from radiographic examinations,","cited_arxiv_id":null,"evidence_quote":"Clinical motivation that obese patients receive higher radiation doses, supporting the low-dose framing."}],"review_version":1}