{"id":"ef36251a-4f40-4d4d-bcd4-b3007a814773","arxiv_id":"2505.09985","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An ordered-subsets multi-diffusion model trains separate diffusion models on view subsets plus one global model, improving sparse-view CT reconstruction.","lead":"Researchers split CT sinogram data into view subsets, train a diffusion model on each subset, and add a whole-data diffusion model to enforce global consistency. The two-stage scheme improves sparse-view CT reconstruction quality over several diffusion and deep-learning baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical superiority claim is not yet fully supported: Tables I–III lack error bars and contain an internally inconsistent PSNR/MSE entry; independent reproducibility check is needed.","rationale":"The reader's conditional verdict is appropriate. I did not find a deeper logical contradiction in the algorithm that would warrant rejection; the ordered-subset decomposition, while theoretically unmotivated, is a plausible divide-and-conquer strategy, and the ablation in Table IV gives at least some internal evidence that both stages contribute. The load-bearing weakness is the empirical proof chain: the central claim rests on Tables I–III, which lack reproducibility metadata and contain an inconsistent entry. This is a concrete block to full acceptance, not a criticism of the intellectual contribution. I agree with the reader that overclaiming noise resilience is a problem, but I would treat that as a secondary qualifier. Thus the verdict should remain conditional.","tokens_in":15343,"tokens_out":7347,"duration_ms":78631,"concrete_test":"Use the public repository (https://github.com/yqx7150/OSMM), pin a specific commit, and rerun the CIRS phantom experiment under the reported settings for all view counts using at least three random seeds, saving all output sinograms. Recompute PSNR and MSE from the saved outputs and compare to Table II; in particular check whether the 180-view MSE is 0.021e-3 rather than 0.21e-3. Compute the mean and standard deviation of PSNR for OSMM and GMSD; if the OSMM-GMSD gap is smaller than one standard deviation at any view count, the claimed 'consistently outperforms' conclusion is not supported, while a persistent gap with a corrected table would resolve the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is empirical: OSMM achieves the best PSNR/SSIM on three datasets. What must be true for that claim is that the reported point estimates are reliable and that the comparison to single-model diffusion baselines is not an artifact of one seed or split. This is the least secure part of the paper. Tables I–III present single numbers with no standard deviations, no number of repetitions, and no significance test. More concretely, Table II contains an internal contradiction: for the CIRS 180-view case, OSMM is reported as PSNR=46.87 dB and MSE=0.21e-3. For normalized data these quantities are inconsistent, since 46.87 dB corresponds to MSE≈0.021e-3; the printed MSE is ten times too large, indicating that at least one table was not carefully cross-checked. The abstract additionally asserts 'noise resilience,' but no experiment with noisy projections is reported, so that part of the central claim has no direct evidence. The subset-splitting rationale in Section III-A is heuristic, but the method can still be correct; the empirical claim, however, cannot be separated from the reliability of the tables. Until a commit-pinned version of the linked repository reproduces the numbers with error bars, the superiority claim is conditional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OSMM, an unsupervised ordered-subsets multi-diffusion model for sparse-view CT reconstruction. The method divides full-view sinogram data into multiple view subsets, trains a separate score-based diffusion model (MSDM) on each subset, and then applies a whole-sinogram diffusion model (OWDM) as a global consistency constraint, with alternating data-consistency updates during iterative reconstruction. Experiments on AAPM challenge data, a CIRS phantom, and preclinical mouse data compare OSMM against FBP, U-Net, FBPConvNet, patch-based DDPM, and GMSD, reporting that OSMM achieves the best PSNR/SSIM/MSE at all tested view counts. Ablation studies examine the contributions of MSDM and OWDM and the effect of the number of subsets.","tokens_in":15619,"tokens_out":2574,"duration_ms":26840,"significance":"If the reported empirical results hold, OSMM would provide a generally applicable unsupervised sinogram-domain prior for sparse-view CT that improves on single-model diffusion baselines, with generalization demonstrated across a human-CT test set, a phantom, and a preclinical mouse dataset. The paper has several strengths: the evaluation is on held-out data not used in training, the method is compared against established baselines rather than only against the authors' prior work, ablations isolate the contribution of each stage, and the source code is publicly linked. These strengths make the central claim worthy of serious consideration, and the paper could be a useful contribution to the diffusion-based CT reconstruction literature. However, the significance is currently limited by the lack of statistical reproducibility evidence and by an internal inconsistency in the reporting of one of the headline quantitative results.","major_comments":[{"comment":"The central claim that OSMM 'outperforms' all baselines at every view count rests on single-run point estimates with no standard deviations, no number of repeated experiments, and no significance tests. At the highest view counts the margins over GMSD are small (e.g., Table I: 43.69 vs 41.44; Table II: 46.87 vs 42.86; Table III: 40.20 vs 39.07), so without variance information the superiority claim is not statistically supported. Please report means and standard deviations over at least three independent runs, or otherwise provide a significance analysis.","section":"§IV-C, Tables I–III"},{"comment":"The CIRS 180-view entry reports PSNR=46.87 dB and MSE=0.21e-3. For normalized images these two quantities are mutually inconsistent: PSNR=46.87 dB corresponds to MSE≈0.021e-3, while MSE=0.21e-3 corresponds to PSNR≈36.8 dB. This internal inconsistency indicates a reporting or computation error in a central results table and must be corrected and verified before the generalization claim can be accepted.","section":"§IV-C, Table II, CIRS 180-view row"},{"comment":"The abstract claims that OSMM outperforms traditional diffusion models in 'noise resilience,' but no experiment with noisy projections is reported anywhere in the paper. The only mention of noisy data is a qualitative statement in §V that 'when projection data is noisy, using too many subsets can hinder convergence.' Either add explicit experiments with noisy sinograms to support the noise-resilience claim, or remove that claim from the abstract.","section":"Abstract and §V"},{"comment":"The method's rationale rests on the heuristic that 'dividing the projection data into multi-subsets can more effectively represent prior information' and that per-subset diffusion reduces learning difficulty. This premise is not justified theoretically or empirically, and Algorithm 1's alternating MSDM/OWDM scheme is presented without a convergence analysis or a clear objective that the iterations are guaranteed to minimize. Since the paper's contribution is precisely this decomposition, please add either a formal convergence argument for the alternating scheme or an empirical study (e.g., reconstruction error vs. iteration number, and a comparison of per-subset versus whole-data learning curves) that supports the stated motivation.","section":"§III-A and Algorithm 1"}],"minor_comments":[{"comment":"The caption refers to 'GSDM' while the text and tables use 'GMSD'; please harmonize the naming.","section":"§IV-C, Fig. 10 caption"},{"comment":"The AAPM data description states that 'the FBP algorithm was employed to obtain artifact-free images at 720 projection angles, serving as the standard reference,' but it is unclear whether the 720-view FBP reconstruction is truly artifact-free; please clarify the reference standard and its relation to the original high-dose CT images.","section":"§IV-A"},{"comment":"The subset definitions in the text (e.g., x1 = {1, N+1, 2×N+1, ..., (N−1)×N+1}) appear to contain off-by-one inconsistencies; for N subsets and M views, the last element in x1 should be M−N+1 rather than (N−1)×N+1 when M is a multiple of N. Please fix the notation.","section":"§II-C"},{"comment":"The statement in the text that 'Analysis of 90 views reconstructions reveals a similar trend to that observed with 90 views' appears to contain a typo; the second '90 views' should likely be '60 views' or another view count.","section":"§IV-C, Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the method is plausible, but the empirical support for the central superiority claim needs to be strengthened. The internal inconsistency in Table II (46.87 dB vs 0.21e-3) is particularly concerning because it suggests that at least some reported numbers were not carefully cross-checked. I would encourage the editor to request the commit-pinned repository link and to require the authors to provide run-to-run variability information before considering the paper for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the architecture: instead of training one diffusion model on the full sinogram, the authors split projection views into ordered subsets round-robin style (borrowed from OSEM), train a separate score model on each subset, and then run an alternating scheme that also consults a whole-sinogram diffusion model, with data consistency steps in between. That specific two-stage combination is not in the cited literature, and the paper shows it works on three datasets, including generalization from AAPM training to CIRS phantom and preclinical mouse data. The ablations in Table IV are useful because they show both the subset models and the global model contribute. The code is public. That is real value for a medical imaging audience.\n\nThe soft spots are recognisable. Tables I-III report single means, no standard deviations, no repetition count, no significance tests, so the superiority claim is narrower than the abstract suggests. Table II has an actual inconsistency: for the CIRS 180-view case, OSMM is listed as PSNR=46.87 dB and MSE=0.21e-3; those numbers do not match for normalized data (the MSE should be about ten times smaller). The abstract also claims noise resilience, but no noisy-projection experiment is reported. The motivation in Section III-A, that subset splitting reduces learning difficulty, is heuristic and not supported by any formal argument; the alternating algorithm in Algorithm 1 has no convergence analysis. None of these are fatal, but they need to be addressed before the performance claim is fully trustworthy.\n\nI want to be clear about what holds up. The math is standard score-based SDE theory correctly applied. The method is plausible and internally coherent, and the empirical pattern is consistent across all three datasets. The table typo is probably a transcription error rather than a sign of fabricated numbers, but it needs fixing. The citation pattern looks fine: the paper builds on the authors' own GMSD work and on OSEM, and it compares against external baselines.\n\nThis paper deserves a serious referee. The idea is useful, the experiments are reasonably broad, and the flaws are reporting problems plus missing uncertainty quantification, not a load-bearing logical break. My recommendation: send it to peer review, ask the authors to correct the table, add error bars or at least multiple runs, run at least one noisy-projection experiment to back the noise-resilience claim, and tone down the abstract until the evidence is there.","headline":"OSMM is a plausible new combination of per-subset diffusion models plus a whole-sinogram diffusion model for sparse-view CT; the empirical claim is likely true, but the paper under-reports uncertainty and contains a concrete table error.","tokens_in":16149,"tokens_out":1438,"would_cite":false,"duration_ms":17065,"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 claims that splitting CT projection data into ordered subsets, training a diffusion model on each subset, and refining with a whole-sinogram diffusion model reconstructs sparse-view CT images with higher PSNR/SSIM than…","keywords":["sparse-view CT","score-based diffusion model","ordered subsets","sinogram reconstruction","unsupervised learning","image reconstruction","generative prior","data consistency"],"falsifier":"A controlled experiment that matches total compute: train a single whole-sinogram diffusion model with the same architecture, parameter count, and total training steps or GPU time as the N subset models together. If the single model matches or exceeds OSMM's PSNR/SSIM on the same sparse-view test data, the claimed benefit of subset division is not real. Also, training OSMM with N=1 inside the same code should reproduce OWDM-only performance; any gap must then be attributed to subset learning.","tokens_in":15134,"feed_emoji":"🩻","tokens_out":5059,"duration_ms":47945,"temperature":0.7,"pith_summary":"This paper proposes OSMM, a two-stage diffusion-model pipeline for sparse-view CT reconstruction. Instead of feeding the full sinogram to one score-based diffusion model, it splits the projection views into ordered subsets, trains a separate diffusion model on each subset (MSDM), and then uses a whole-sinogram diffusion model (OWDM) as a global consistency constraint. The authors claim this reduces the difficulty of learning the projection data distribution, preserves fine details, and corrects errors, leading to better reconstruction quality. Experiments on AAPM, CIRS phantom, and preclinical mouse data report that OSMM beats FBP, U-Net, FBPConvNet, patch-based DDPM, and GMSD in PSNR/SSIM across 50–180 views. If correct, the method offers an unsupervised, sparsity-adaptive alternative to supervised and single-model generative approaches.","feed_headline":"Splitting CT sinograms into subsets beats single-model diffusion","feed_subtitle":"Ordered subsets each get a diffusion model; a whole-sinogram pass ties them together and lifts PSNR.","key_machinery":"The operative mechanism is the ordered-subsets division of the sinogram: projection views $1,\\dots,M$ are partitioned into $N$ disjoint subsets in round-robin order, with view $n$ assigned to subset $n$. Each subset is then handled by an independent variance-exploding SDE score model. The MSDM stage applies the $N$ subset predictors with per-step data consistency; the OWDM stage merges the refined subsets and runs one whole-sinogram predictor, again with data consistency. This alternating scheme carries the argument by decomposing the learning task.","core_discovery":"OSMM's core discovery is that ordered-subsets partitioning of the projection data turns one hard diffusion-learning problem into several easier ones: each subset diffusion model (MSDM) reconstructs its own sparse sinogram, recovering fine details, while the one-whole diffusion model (OWDM) keeps the merged sinogram globally consistent. Iterating the two stages with data-consistency updates produces sinograms that, after filtered back projection, have higher PSNR and SSIM than those from FBP, U-Net, FBPConvNet, patch-based DDPM, and GMSD. The authors further claim that the unsupervised training on full-view sinograms transfers across datasets and sparsity levels.","pith_inferences":["If subset division is the real driver, the same ordered-subsets trick could be applied to other domains with large, redundant measurement spaces, such as MRI k-space or PET sinograms, where per-partition score models might ease training.","The paper's 'reduced learning difficulty' rationale predicts that the gain over a single model should grow with the redundancy of the full sinogram; a testable extension is to vary the amount of angular overlap in the projection data and measure the PSNR gain.","Because the alternating MSDM/OWDM scheme lacks a convergence guarantee, a practical extension would be to analyze it as a fixed-point iteration or replace the Predictor with a predictor-corrector sampler to see whether the data-consistency steps actually contract.","The subset-count effect (N=2 vs N=3) hints at a bias–variance trade-off: more subsets lower per-model difficulty but increase the chance of inconsistent merges; finding the automatic rule for N would turn the heuristic into an algorithm."],"forward_implications":["Sparse-view CT reconstructions from as few as 50–60 views can reach PSNR values above 37 dB on AAPM data, roughly 3 dB higher than the next best generative baseline.","The method transfers across scanners and anatomies: models trained only on AAPM data improve CIRS phantom and preclinical mouse reconstructions without retraining.","At higher view counts (120, 180, 100) OSMM still gains, implying the subset prior remains useful even when the inverse problem is less ill-posed.","Ablation shows both stages matter: removing either MSDM or OWDM lowers PSNR, so the two-stage alternation is a genuine part of the improvement.","Increasing the number of subsets from 2 to 3 does not consistently improve results, suggesting there is an optimal split count."],"supporting_citations":[{"why":"Supplies the VE-SDE diffusion framework and Predictor sampling used in the MSDM and OWDM stages.","marker":"[18]"},{"why":"Introduces ordered-subsets partitioning for reconstruction, the direct inspiration for the subset-division scheme.","marker":"[25]"},{"why":"GMSD, a sinogram-domain score-based baseline, provides the main generative comparison that OSMM must beat.","marker":"[14]"},{"why":"Patch-based DDPM is a diffusion baseline whose detail preservation is compared against OSMM.","marker":"[12]"},{"why":"Supplies the AAPM Challenge dataset used for training and most of the quantitative evaluation.","marker":"[26]"},{"why":"Siddon's ray-driven algorithm generates the fan-beam sinograms used in the AAPM experiments.","marker":"[27]"},{"why":"Cited in the discussion to support the claim that too many subsets can hinder convergence.","marker":"[31]"}],"fun_headline_variants":["Ordered-subsets diffusion sharpens sparse-view CT","Split sinograms, multi-diffuse, better sparse CT","Subset diffusion models boost sparse CT quality","Partitioned sinogram diffusion outperforms for sparse CT","Multi-diffusion on ordered sinogram subsets lifts CT quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that splitting the projection data into ordered subsets genuinely lowers each diffusion model's learning difficulty while preserving the information needed for fine details; the paper gives this as a heuristic with no statistical or convergence analysis.","fun_headline_variants_meta":{"raw":{"variants":["Ordered-subsets diffusion sharpens sparse-view CT","Split sinograms, multi-diffuse, better sparse CT","Subset diffusion models boost sparse CT quality","Partitioned sinogram diffusion outperforms for sparse CT","Multi-diffusion on ordered sinogram subsets lifts CT quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001033,"raw_usage":{"total_tokens":4331,"prompt_tokens":904,"completion_tokens":3427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":3347}},"tokens_in":520,"tokens_out":3427,"duration_ms":27483,"temperature":1.0,"reasoning_tokens":3347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:18:55.958319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that matches total compute: train a single whole-sinogram diffusion model with the same architecture, parameter count, and total training steps or GPU time as the N subset models together. If the single model matches or exceeds OSMM's PSNR/SSIM on the same sparse-view test data, the claimed benefit of subset division is not real. Also, training OSMM with N=1 inside the same code should reproduce OWDM-only performance; any gap must then be attributed to subset learning.","supporting_citations":[{"cited_title":"Score-based generative modeling through stochastic differential equations,","cited_arxiv_id":null,"evidence_quote":"Supplies the VE-SDE diffusion framework and Predictor sampling used in the MSDM and OWDM stages."},{"cited_title":"Accelerated image reconstruction using or- dered subsets of projection data","cited_arxiv_id":null,"evidence_quote":"Introduces ordered-subsets partitioning for reconstruction, the direct inspiration for the subset-division scheme."},{"cited_title":"Generative modeling in sinogram domain for sparse-view CT reconstruction,","cited_arxiv_id":null,"evidence_quote":"GMSD, a sinogram-domain score-based baseline, provides the main generative comparison that OSMM must beat."},{"cited_title":"Accessed: Apr","cited_arxiv_id":null,"evidence_quote":"Supplies the AAPM Challenge dataset used for training and most of the quantitative evaluation."},{"cited_title":"Fast calculation of the exact radiological path fora three- dimensional CT array,","cited_arxiv_id":null,"evidence_quote":"Siddon's ray-driven algorithm generates the fan-beam sinograms used in the AAPM experiments."},{"cited_title":"Statistically regulated and adaptive EM reconstruction for emission computed tomography","cited_arxiv_id":null,"evidence_quote":"Cited in the discussion to support the claim that too many subsets can hinder convergence."}],"review_version":1}