{"id":"4cce69ea-b214-4e7a-bafc-01a4a448cf5a","arxiv_id":"2506.22012","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-domain diffusion model trained only on normal-dose CT data generalizes to unseen low-dose levels by matching Poisson noise in projections and refining images with double guidance.","lead":"This paper introduces NEED, a two-stage diffusion model that cleans low-dose CT scans using only normal-dose data for training and adapts automatically to dose levels it has never seen. It combines a shifted-Poisson diffusion denoiser on CT projection data with a doubly guided diffusion refiner on images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dose-generalization claim rests entirely on simulated projections generated from the same shifted-Poisson model SPDiff is trained to invert; no real pre-log projection data is used, including in the Mayo 2020 external test.","rationale":"The reader's weakest_assumption matches my main concern. The paper is carefully engineered and the internal ablations (Table 8, Fig. 15) support the contribution of each module, but the central generalization claim is evaluated entirely within simulated Poisson+Gaussian projections. My reading of Section 4.1 suggests even Mayo 2020 is not used as measured pre-log data: the authors synthesize pre-log projections from provided LDCT images, making the 'external dataset' claim weaker than presented. The concern is not that the simulation is worthless; it is a reasonable first validation. However, the paper's own Section 5.5 lists exactly the physics omitted, so a conditional acceptance requiring a real-data check is the right calibration. I do not see an internal inconsistency that would require rejection, and the code release makes the proposed test feasible. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":27997,"tokens_out":7745,"duration_ms":83128,"concrete_test":"Run the released NEED code, without retraining, on the actual pre-log projection data from the Mayo 2020 Low Dose CT Image and Projection Dataset (Moen et al., 2021) instead of forward-projecting reconstructed LDCT images. Estimate I_ld from unattenuated air regions or scanner metadata, apply Alg. 1, reconstruct with FBP, and report PSNR/SSIM/RMSE against the official normal-dose images for chest (10%) and abdomen (25%). If NEED no longer consistently outperforms GDP, Dn-Dp, and SPDiff alone, or if the gap to the simulated Mayo 2016 results exceeds its inter-method margins, the dose-generalization claim is not established on real scanner physics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that NEED generalizes to unseen dose levels depends on SPDiff's shifted-Poisson degradation operator (Eq. 9) being a faithful model of real pre-log low-dose CT measurements: the time-step matching rule (Eq. 16) chooses the sampling start point by equating lambda_t with I_ld. That fidelity is never tested. On Mayo 2016, Section 4.1 states that only post-log data are available, so the authors forward-project reconstructed normal-dose images to synthesize pre-log projections and then add Poisson+Gaussian noise with I0=2.5e5, sigma_e^2=10. This is exactly the degradation family on which SPDiff is trained, so the 'unseen dose' results in Tables 2, 5, and 7 are interpolation within the training degradation schedule lambda_t, not tests against unseen physics. The Mayo 2020 'external validation' is also not measured projection data: Section 4.1 says they use the officially provided LDCT images and the Table 1 geometry to obtain the corresponding pre-log projections. Section 5.5 concedes that real projections include scatter, detector nonlinearities, and other effects absent from this simulation. If real pre-log measurements deviate from the shifted-Poisson model, the entire projection-domain component, and the dose generalization built on it, may not transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NEED, a cascade of two diffusion models for low-dose CT reconstruction. SPDiff denoises pre-log projection data with a shifted-Poisson cold-diffusion operator D(y0,t)=P(λ_t y0 + σ_e^2)/λ_t, and DGDiff refines the FBP reconstruction using a Gaussian diffusion prior with double guidance from the LDCT image and from the initial reconstruction. A time-step matching strategy selects the starting sampling step from the incident photon count I_ld in the projection domain and from σ_ld in the image domain. The method is trained only on normal-dose projections and images, then tested on simulated low-dose data at 50%, 25%, 12.5%, and 10% dose on Mayo 2016 and on Mayo 2020. The paper reports consistent improvements over self-supervised, iterative, and unsupervised diffusion baselines in PSNR/SSIM/RMSE, additional perceptual metrics, downstream MedSAM segmentation, and inference time.","tokens_in":28355,"tokens_out":7315,"duration_ms":75232,"significance":"If the central claims hold, the contribution is practically significant: NEED would offer dose-agnostic low-dose CT reconstruction trained only on normal-dose data, with an explicit physics-inspired projection-domain degradation and faster sampling than single-domain diffusion baselines. The paper has clear strengths: source code is released; the sampling update in Eq. (10) is internally consistent with the shifted-Poisson degradation; the ablation studies isolate the contributions of SPDiff, DGDiff, and time-step matching; and the diffusion baselines share the same pretrained Gaussian diffusion model, which makes the comparison fair at the architectural level. The downstream MedSAM evaluation adds a useful clinically oriented check.","major_comments":[{"comment":"The headline claim of generalization to unseen dose levels is validated only on projections synthesized with the same Poisson+Gaussian model from which SPDiff's degradation operator is constructed. For Mayo 2016, the pre-log projection data are obtained by forward-projecting reconstructed NDCT images and adding Poisson+Gaussian noise with I0=2.5e5 and σ_e^2=10; the SPDiff degradation in Eq. (9) uses the same σ_e^2 and a λ_t schedule spanning the range of I_ld. The 'unseen dose' experiments therefore test interpolation within the λ_t schedule, not robustness to deviations from the assumed noise model. Since Section 5.5 acknowledges that real projections contain scatter, detector nonlinearities, and other effects absent from this model, the clinical transferability of the dose-generalization claim is not established. I ask for validation on real measured pre-log data or, at minimum, a hold-out simulation with different σ_e^2, beam-hardening, or detector response.","section":"Sec. 4.1, Tables 2, 5, 7"},{"comment":"The Mayo 2020 'unseen dataset' test is also not based on measured projection data. The text states that the authors use the officially provided LDCT images and 'leverage the projection geometry outlined in Table 1 to obtain the corresponding pre-log LDCT projection.' Since those LDCT images are themselves simulated (from post-log signal-dependent Gaussian noise with bowtie filtration and automatic exposure control), the resulting pre-log projections are another simulation layer, not independent external validation. The claim that NEED 'can be effectively extended to various unseen dose levels' and the comparison in Table 5 would be substantially strengthened if the official Mayo 2020 raw projection data were used instead of re-projecting reconstructed images.","section":"Sec. 4.1, Mayo 2020 experiments"},{"comment":"There is a contradiction in the description of the Noise2Noise baseline. Section 4.3 says 'we simulate an additional set of 25% and 10% dose CT images from the Mayo 2016 dataset as training labels,' while Section 4.6 says 'Noise2Noise, Noise2Sim, and SSDDNet are trained using only 25% dose data.' If Noise2Noise received 10% paired training data, its 10% row in Table 2 is a seen-dose result, not a zero-shot generalization point, and the comparison at 10% dose is unfair. The authors should clarify exactly which dose levels were used for each baseline and, if 10% labels were used, retrain Noise2Noise on 25% only for the generalization comparison.","section":"Sec. 4.3 vs Sec. 4.6, Table 2"},{"comment":"The doubly guided sampling formula is presented as an exact posterior factorization, but the two guidance terms are conditioned on x_t rather than on x_{t-1}, and p(x_g|x_t) is then replaced by an L1-distance modulation with hand-set weights s1 and s2. This is a heuristic guided-sampling procedure, not a derivation. The ablations in Table 8 show empirical value, but the paper should either derive the approximation more carefully (for example, as a variational or annealed update) or explicitly state that Eq. (13) is an ad-hoc guided sampler. Without this clarification, readers cannot tell how much of the reported gain comes from the noise-aligned SPDiff versus from tuning the guidance weights.","section":"Sec. 3.2, Eq. (13)"}],"minor_comments":[{"comment":"Section 5.2 contains a word-for-word duplicated paragraph: the passage beginning 'Our NEED leverages the number of X-ray incident photons I0...' appears twice in succession; one copy should be deleted.","section":"Sec. 5.2"},{"comment":"The sentence 'SPDiff (τ) and SPDi ff (T) yield identical results at this dose level' is confusing because T is not defined for SPDiff; the total diffusion steps are τ=10, so either use SPDiff (τ) consistently or define T separately.","section":"Sec. 4.8, Table 7"},{"comment":"The 'perceptual loss curve' is referred to but never defined; please state the exact loss used and how the intermediate reconstructed images are compared.","section":"Fig. 12"},{"comment":"The notation P(λ_t y0 + σ_e^2)/λ_t conflates a shifted-Poisson random variable with a constant addition; defining the transition as m_t ∼ P(λ_t y0) + σ_e^2, y_t = m_t/λ_t would avoid ambiguity.","section":"Eq. (9)"},{"comment":"Line 8 uses y_1 as the input to R_phi, but if t*_prj=1 the loop at lines 4-7 is skipped and y_1=y_ld; an explicit note that this is the intended behavior would improve reproducibility.","section":"Algorithm 1, line 8"},{"comment":"The values I0=2.5e5 and σ_e^2=10 are given without explicit units or normalization conventions for y_0; stating these conventions is important for reproducing the simulation.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for Medical Image Analysis and the method is well motivated, but the core generalization claim rests on simulated projections that match the assumed noise model. The authors already acknowledge this in Section 5.5, which is good, but the acknowledgment does not repair the gap between the claim as stated and the evidence. The self-citation pattern is concentrated in the authors' own prior work (CoreDiff, CoCoDiff, PrideDiff); this is relevant but not disqualifying. I believe the paper is fixable with additional validation or a carefully restated claim, and I would encourage the editor to require either real pre-log projection experiments or a substantially more diverse simulation study before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The worthwhile part of this paper is the engineering. The shifted Poisson diffusion process on pre-log projections, trained on normal-dose data only, is a genuine adaptation of cold diffusion to CT physics, and the time-step matching rule that picks the sampling start from the incident photon count I0 is physically motivated. The dual-domain cascade is well thought out, and the ablations are more thorough than most: Table 9 shows DGDiff improves a range of initial reconstructions, Table 7 shows time-step matching matters beyond just speed, and the internal comparisons share the same pretrained diffusion prior for fairness. Code is released. That is real work and real evidence.\n\nThe soft spot is exactly where the stress-test note lands. The Mayo 2016 results are produced by forward-projecting reconstructed normal-dose images and adding Poisson+Gaussian noise with the same parameters the shifted Poisson operator is designed to invert. The Mayo 2020 \"external\" test is not measured projection data either; the authors reconstruct pre-log projections from the official LDCT images. So the phrase \"generalization to unseen dose levels\" currently means interpolation within one simulated noise family. The authors concede this in Section 5.5, which is honest, but the central claim of the paper—that the method transfers to real pre-log measurements—remains untested. Scatter, detector nonlinearities, and bowtie effects are not in the training or test distributions. I would not call this fatal, because the framework does not depend on memorizing a fixed dose; the time-step matching uses a physical input rather than a learned test constant, and the internal logic is sound. But the paper's headline claim is broader than its evidence.\n\nMinor issues: the key weighting formulas for s1 and lambda_t are deferred to a missing supplementary, there is no sensitivity analysis on the hand-set hyperparameters, and the quantitative tables lack statistical significance tests. These are ordinary referee requests, not fatal flaws.\n\nWho is this for? Anyone working on dose-agnostic CT reconstruction or diffusion-based inverse problems in medical imaging. It deserves a serious referee. The right path is conditional acceptance with a request for either real pre-log projection data or a deliberate model-mismatch experiment—for example, testing SPDiff on projections simulated with different electronic noise, added scatter, or a different noise model—to show the method degrades gracefully when the assumed physics is wrong. Without that, the generalization claim is unproven on the evidence presented.","headline":"A well-engineered dual-domain diffusion method whose core dose-generalization claim rests on one simulated noise model; real projection data is the missing test.","tokens_in":28778,"tokens_out":1127,"would_cite":true,"duration_ms":15766,"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":"A diffusion model trained only on normal-dose CT data can reconstruct low-dose scans at dose levels never seen during training.","keywords":["low-dose CT reconstruction","diffusion model","shifted Poisson noise","dose generalization","projection-domain denoising","image-domain refinement","time step matching","dual-domain reconstruction"],"falsifier":"Take a phantom or patient with a normal-dose reference scan, acquire true low-dose pre-log projections at a dose level not used in training, estimate $I_0$ from the scan itself, and run NEED. The central claim fails if the reconstruction develops visible artifacts, CT-value bias, or a significant drop in PSNR/SSIM relative to the simulated-dose results—or if the measured variance of the real projections is not consistent with the linear-in-signal shifted Poisson form assumed by the degradation operator.","tokens_in":27785,"feed_emoji":"🩻","tokens_out":8712,"duration_ms":79477,"temperature":0.7,"pith_summary":"The paper claims that a single CT reconstruction pipeline, trained only on normal-dose projection and image data, can clean up low-dose scans across a range of doses, including doses never encountered in training. The model, NEED, cascades a shifted Poisson diffusion model that denoises raw projections according to the measured physics of low-dose CT with a doubly guided diffusion model that refines the image using both the noisy low-dose scan and the initial reconstruction. A time step matching rule chooses where each test scan enters the diffusion process, effectively adapting the denoising strength to the scan's dose. If the claim is right, clinical and research CT denoising would no longer need paired normal/low-dose data or retraining for every new protocol.","feed_headline":"One model, trained on normal-dose CT, cleans unseen low-dose scans","feed_subtitle":"No paired low-dose training data needed: one diffusion cascade adapts denoising to each scan's dose","key_machinery":"The load-bearing object is the shifted Poisson degradation operator $D(y_0,t)=\\mathcal{P}(\\lambda_t y_0+\\sigma_e^2)/\\lambda_t$, which injects Poisson noise scaled by $\\lambda_t$ plus electronic-noise variance $\\sigma_e^2$ into normal-dose projections, so the forward diffusion mimics real pre-log low-dose measurements. Training a restoration network $R_\\phi$ on normal-dose projections corrupted by this operator gives a denoiser matched to CT physics, and the inference recursion $y_{t-1}=(\\lambda_t y_t+\\mathcal{P}((\\lambda_{t-1}-\\lambda_t)\\tilde{y}_0))/\\lambda_{t-1}$ removes the shifted Poisson noise in steps. For images, the mechanism is double guidance: the noisy sample is modulated by the low-dose image, $\\tilde{x}_t=s_1 x_t+(1-s_1)\\sqrt{\\bar\\alpha_t}\\,x_{\\mathrm{ld}}$, and the denoised estimate is blended with the initial reconstruction, $x_0=s_2\\tilde{x}_0+(1-s_2)\\hat{x}_0$. Time step matching, $t^*_{\\mathrm{prj}}=\\arg\\min_t |\\lambda_t-I_{\\mathrm{ld}}|$ for projections and $t^*_{\\mathrm{img}}=\\arg\\min_t |\\sqrt{1-\\bar\\alpha_t}-\\hat{\\sigma}_{\\mathrm{ld}}|$ for images, converts the known dose level into a dose-adaptive sampling schedule.","core_discovery":"NEED's central claim is that the main barrier to generalizable low-dose CT reconstruction is a mismatch between the noise model assumed by diffusion models and the noise actually present in CT data. The paper addresses this mismatch in two domains: in projection space it replaces Gaussian degradation with a shifted Poisson degradation operator aligned to pre-log measurement statistics, and in image space it uses a normal-dose diffusion prior with double guidance from the noisy low-dose image and the initial reconstruction, which locates the prior more precisely. The time step matching strategy then selects the sampling start point from the incident photon count for projections and from the residual standard deviation for the image, and this is what lets one model cover multiple dose levels. On two public low-dose CT datasets, the paper reports that NEED outperforms self-supervised methods, unsupervised diffusion-based baselines, and supervised diffusion and CNN baselines in PSNR, SSIM, and RMSE, and that its reconstructions improve a downstream segmentation model's Dice scores.","pith_inferences":["A natural extension is to treat $I_0$ as an unknown and estimate it from the test projection itself, which would remove the need for scanner calibration data and make the method applicable when exposure metadata is unavailable.","The same cascade design—physics-matched diffusion denoising in a raw-data domain followed by a clean-data diffusion prior refined with double guidance—should transfer to other Poisson-limited imaging problems such as PET, SPECT, or photon-counting CT, provided pre-log data are available.","If real scanner noise departs from the shifted Poisson model, the time step matching rule could be recalibrated empirically per scanner by fitting $\\lambda_t$ to measured noise-variance curves, preserving the dose-adaptive behavior without retraining the full model.","The paper's evaluation on simulated projections leaves open how much of the improvement survives real scatter and beam hardening; a direct test on real pre-log data would be the decisive next step."],"forward_implications":["A single NEED model trained once on normal-dose data can be applied at multiple low-dose levels without retraining or fine-tuning, including levels not seen in training.","Paired normal/low-dose CT collections are no longer a prerequisite for high-quality denoising; only normal-dose projections and images are needed.","Handling the projection-domain noise with the shifted Poisson model reduces CT-value drift and preserves lesion structure, which carries through to better downstream segmentation.","The time step matching strategy cuts sampling steps and inference time compared with full diffusion sampling, while matching or improving reconstruction quality.","Because the image-domain prior is trained only on normal-dose images, the same pipeline can be transferred to a new dataset with a different noise distribution directly, as demonstrated on the second public dataset."],"supporting_citations":[{"why":"Supplies the cold-diffusion framework of custom degradation operators that the shifted Poisson forward process is built on.","marker":"(Bansal et al., 2023)"},{"why":"Provides the Gaussian DDPM backbone used as the image-domain prior in the doubly guided refinement stage.","marker":"(Ho et al., 2020)"},{"why":"States the Poisson+Gaussian model for pre-log CT measurements that SPDiff's degradation operator is designed to invert.","marker":"(Wang et al., 2017)"},{"why":"Provides the simulation approach for creating low-dose projections with Poisson+Gaussian noise used to generate test data at multiple dose levels.","marker":"(Zeng et al., 2015)"},{"why":"Defines Dn-Dp, the diffusion-prior baseline whose guidance formulation is extended by double guidance and whose results NEED is compared against.","marker":"(Liu et al., 2024)"},{"why":"Defines DR2, a degradation-remover baseline using low-frequency guidance that NEED's adaptive time step matching improves upon.","marker":"(Wang et al., 2023)"},{"why":"Provides the first public low-dose CT challenge dataset used for training and seen/unseen dose evaluation.","marker":"(Chen et al., 2016)"},{"why":"Provides the second public low-dose CT image and projection dataset used to test generalization to an unseen dataset.","marker":"(Moen et al., 2021)"}],"fun_headline_variants":["Noise-matched diffusion adapts to any CT dose unseen","Shifted Poisson diffusion: one model, all dose levels","Train on normal CT, denoise low-dose without paired data","Cascaded diffusion doubles guidance to fix CT noise","Unseen doses? This diffusion knows the noise physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that low-dose projections made by forward-projecting normal-dose images and adding Poisson plus Gaussian noise with known parameters faithfully reproduce real scanner measurements, including scatter, beam hardening, and detector effects.","fun_headline_variants_meta":{"raw":{"variants":["Noise-matched diffusion adapts to any CT dose unseen","Shifted Poisson diffusion: one model, all dose levels","Train on normal CT, denoise low-dose without paired data","Cascaded diffusion doubles guidance to fix CT noise","Unseen doses? This diffusion knows the noise physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1719,"prompt_tokens":1013,"completion_tokens":706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":629,"tokens_out":706,"duration_ms":8103,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:13:19.593856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a phantom or patient with a normal-dose reference scan, acquire true low-dose pre-log projections at a dose level not used in training, estimate $I_0$ from the scan itself, and run NEED. The central claim fails if the reconstruction develops visible artifacts, CT-value bias, or a significant drop in PSNR/SSIM relative to the simulated-dose results—or if the measured variance of the real projections is not consistent with the linear-in-signal shifted Poisson form assumed by the degradation operator.","supporting_citations":[{"cited_title":", author Borgnia, E","cited_arxiv_id":null,"evidence_quote":"Supplies the cold-diffusion framework of custom degradation operators that the shifted Poisson forward process is built on."},{"cited_title":", author Zhou, J","cited_arxiv_id":null,"evidence_quote":"States the Poisson+Gaussian model for pre-log CT measurements that SPDiff's degradation operator is designed to invert."},{"cited_title":", author Huang, J","cited_arxiv_id":null,"evidence_quote":"Provides the simulation approach for creating low-dose projections with Poisson+Gaussian noise used to generate test data at multiple dose levels."},{"cited_title":", author Xie, Y","cited_arxiv_id":null,"evidence_quote":"Defines Dn-Dp, the diffusion-prior baseline whose guidance formulation is extended by double guidance and whose results NEED is compared against."},{"cited_title":", author Zhang, Z","cited_arxiv_id":null,"evidence_quote":"Defines DR2, a degradation-remover baseline using low-frequency guidance that NEED's adaptive time step matching improves upon."},{"cited_title":", author Leng, S","cited_arxiv_id":null,"evidence_quote":"Provides the first public low-dose CT challenge dataset used for training and seen/unseen dose evaluation."},{"cited_title":", author Chen, B","cited_arxiv_id":null,"evidence_quote":"Provides the second public low-dose CT image and projection dataset used to test generalization to an unseen dataset."}],"review_version":1}