{"id":"e785f0d0-1870-4aa3-bac0-7f386cb344e7","arxiv_id":"2508.14364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PET-DPC, a conditional diffusion model with posterior physical correction and a GTP-image input, reconstructs PET images with accuracy close to iterative QuanTOF at reduced reconstruction time.","lead":"This paper proposes a new PET image reconstruction method that combines a diffusion model with physics-based corrections, aiming to match the accuracy of standard iterative reconstruction while running faster. The method is tested on clinical brain scans, whole-body scans, a physical phantom, and simulations, showing improved accuracy over end-to-end deep learning models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Posterior correction in Algorithm 1 Step 5 is not derived from the Poisson likelihood and scatter/random estimation is unspecified, so the quantitative accuracy claim is not supported.","rationale":"The reader identified the posterior correction in Algorithm 1 Step 5 as the weakest assumption, and I agree. This is the most load-bearing concern because the entire novelty of PET-DPC is that physical information is incorporated via posterior correction; if that correction is not a valid likelihood-based update, the method reduces to a conditional diffusion model with an ad hoc post-processing step. The paper's own text claims Eq. (14) is used to compute a gradient, but the algorithm implements a different multiplicative update. The lack of specification for scatter and random sinogram estimation is equally serious, as these terms enter the forward model and directly affect the correction. While the phantom study provides some independent evidence of quantitative accuracy, it does not remedy the missing derivation; the correction could be tuned to the phantom specifically. The simulation result showing the recovered gray/white ratio (1.868) is far from the true value (3.846) reinforces that the method's 'quantitative' performance is calibrated to QuanTOF rather than to ground truth. I considered whether the evaluation circularity (using QuanTOF as both training target and reference) is the more fundamental issue, but the posterior-correction derivation is upstream: even if the evaluation were independent, the method's mechanism would still be unsupported. The verdict remains CONDITIONAL because the concern is addressable: the authors could provide a derivation, specify the scatter/random estimation, or release code. I do not see evidence of fraud or deception; the paper simply omits crucial details. Therefore, no verdict change is needed beyond the reader's conditional acceptance.","tokens_in":14578,"tokens_out":6605,"duration_ms":73552,"concrete_test":"Analytically derive the update in Algorithm 1 Step 5 from the Poisson log-likelihood (Eq. 14). Show that the multiplicative update x̂0 ← C_IM x̂0′ * (Σ_i y_i/ȳ_i P_ij)/(Σ_i P_ij) is either a valid EM iteration or a gradient ascent step with a specified step size. If no such derivation exists, or if it requires unknown scatter/random estimates, the 'physics correction' is not grounded in the likelihood, and the central claim of quantitative accuracy is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 Step 5 applies a multiplicative update x̂0 ← C_IM x̂0′ × (Σ_i y_i/ȳ_i P_ij) / (Σ_i P_ij), which is a MLEM-style ratio update. However, the paper claims this implements the gradient of the Poisson log-likelihood in Eq. (14). The gradient of Eq. (14) w.r.t. x_j is Σ_i P_ij (y_i/ȳ_i − 1), a different expression; no step size or equivalence proof is given. The update is also not derived from the likelihood via EM. Moreover, ȳ_i in Eq. (10) includes random and scatter means r̅_i and s̅_i, but the paper never specifies how these are estimated during posterior correction. Since the central claim of quantitative accuracy rests entirely on this correction step, the method is currently an unexplained heuristic. The simulation results in Table III (gray/white ratio 1.868 vs ground truth 3.846) further suggest the method inherits QuanTOF's biases rather than achieving absolute quantitative accuracy, but the core issue is the unjustified correction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PET-DPC, a conditional denoising diffusion probabilistic model for PET reconstruction. List-mode data are first back-projected into a Geometric TOF Probabilistic Image (GTP-image) that serves as the conditioning input; during the reverse diffusion sampling process, a posterior correction step updates the current estimate using the measured list-mode data and the Poisson forward model (Eq. 10), intended to apply scatter, attenuation, and random corrections. The model is trained on 300 brain and 50 whole-body clinical datasets and evaluated on 100 brain validation cases, 5 whole-body validation cases, a NEMA phantom, and 20 simulated brain datasets, with QuanTOF reconstructions used as training labels and as the main reference. The paper reports that PET-DPC outperforms FastPET and a conditional DDPM without correction, closely matches QuanTOF, and reduces reconstruction time substantially.","tokens_in":14922,"tokens_out":3460,"duration_ms":39014,"significance":"If the posterior correction mechanism is sound, the paper would make a useful contribution by embedding explicit physical corrections into a fast diffusion-based reconstruction pipeline, addressing a recognized limitation of purely data-driven PET reconstruction. The evidence base is unusually broad for this literature: clinical brain and whole-body data, a physical phantom scanned on the same system, and GATE simulations, plus ablations on the number of correction steps and a runtime comparison. However, the validity of the central correction step is not established, scatter/random estimation is unspecified, and the quantitative claims are largely measured as agreement with QuanTOF rather than against physical ground truth, so the significance is contingent on resolving these issues.","major_comments":[{"comment":"The paper states that Eq. (14) is used to compute the gradient with respect to x̂0 and that the measurement data update implements the posterior correction. However, Step 5 applies x̂0 ← C_IM x̂0' · (Σ_i y_i/ȳ_i P_ij)/(Σ_i P_ij), which is an MLEM-style multiplicative update. The gradient of Eq. (14) with respect to x_j is Σ_i P_ij(y_i/ȳ_i − 1), a different expression. No step size, derivation, or equivalence argument is provided. Since the entire quantitative-accuracy claim rests on this correction step, the method is at present an unexplained heuristic and the name 'posterior correction' is not justified.","section":"Section II-B, Algorithm 1 Step 5"},{"comment":"The forward model in Eq. (10) includes random and scatter means r̄_i and s̄_i, but the paper never specifies how these are estimated during posterior correction. The correction step requires an accurate ȳ_i, including these components; without specifying the scatter and random estimation method (e.g., SSS, delayed-window, or simulation-based), the algorithm is not reproducible and the risk of bias from inaccurate estimates is not assessed. The brief mention in the Discussion of treating each 10 s of whole-body acquisition as a bed position for correction is not a methods-level description.","section":"Section II-B, Eq. (10)"},{"comment":"In the simulation study, the ground-truth gray-to-white matter activity ratio is 3.846, but QuanTOF recovers 1.898 and PET-DPC 1.868. The text calls this 'strong quantitative consistency,' which conflates agreement with QuanTOF with absolute quantitative accuracy. By this measure PET-DPC simply inherits QuanTOF's substantial bias. The central claim of quantitative accuracy is therefore not supported by the simulation study; the authors should either use the true activity map to compute errors or moderate the claim to 'close agreement with the QuanTOF reference.'","section":"Table III and Section IV-D"},{"comment":"The whole-body study uses only 5 validation cases, and the reported NRMSE standard deviations are large (e.g., 0.663 ± 0.248 for FastPET). The strong conclusions about whole-body performance and generalization rest on a very small sample. Confidence intervals or per-case results should be provided, and the corresponding claims should be framed as preliminary.","section":"Section III.A.2 and Table II"}],"minor_comments":[{"comment":"The definition of the GTP-image is unclear: the formula includes a seemingly gratuitous factor '1 × N_L', and the normalization by the sensitivity map S_v is introduced without explaining how S_v is computed. Please clarify.","section":"Section II-C, Eq. (15)-(16)"},{"comment":"The intensity-matching coefficient C_IM is described as 'pre-calculated' but no formula or method for its determination is given. This is a free parameter in the method and should be specified for reproducibility.","section":"Section II-D and Algorithm 1"},{"comment":"The index i is reused for the sampling step and for the LOR index in Step 5; although the text notes the distinction, the notation is confusing and should be changed (e.g., use k for LORs).","section":"Algorithm 1"},{"comment":"Typo: 'across across seven reconstructed NEMA phantom images'.","section":"Figure 6 caption"},{"comment":"The baseline 'Conditional-DDPM' is defined only as PET-DPC without posterior correction; this should be stated explicitly in the Methods, and the correction-free conditional model should be described as a separate baseline before being used in comparisons.","section":"Section II-B and III-D"},{"comment":"Reference [29] for QuanTOF is a conference abstract; if the method is used as the training label and reference standard, please provide a fuller citation or a public implementation/algorithm description so that the reference results are reproducible.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on QuanTOF, a self-cited method, both as training label and as evaluation reference. This is not hidden, but it means the 'quantitative accuracy' claim is largely a measure of agreement with QuanTOF, not with physical truth. The simulation study actually reveals a large bias relative to the true activity ratio. I would ask the authors to either provide a valid derivation or justification for the posterior correction update, or reposition the paper's claim as 'fast diffusion reconstruction that approximates QuanTOF' rather than absolute quantitative accuracy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real step forward for fast quantitative PET reconstruction, but the central physics-correction step is a heuristic wearing a likelihood's clothes, and the claims about absolute quantitative accuracy need a careful look.\n\nWhat's genuinely new: the GTP-image input (TOF + geometric backprojection with sensitivity normalization) is a sensible way to feed raw list-mode information into a network, and the idea of alternating a DDPM predictor with a measurement-based corrector during sampling is a clean combination of existing pieces (DPS, MLEM-style updates) that hasn't been applied to PET this way before. The experimental work is serious: 300 brain and 50 whole-body clinical datasets, a physical NEMA phantom, 20 GATE simulations, ablations on the number of correction steps, and OOD generalization tests. That's more than most papers in this space do. The ablation showing performance degrading from 5 correction steps down to 1 is convincing evidence that the correction term is doing something real.\n\nThe soft spots are in the physics and the evaluation. Algorithm 1 Step 5 is written as a multiplicative MLEM-style update, but it's not derived as a gradient of the Poisson log-likelihood in Eq. (14) — the gradient would be an additive term, and no step size or equivalence argument is given. The paper also never says how the scatter and random sinograms (r̅ and s̅ in Eq. 10) are estimated for the correction step. That matters because if those estimates are wrong, the \"physics correction\" could be baking in errors. The strongest evidence of this is Table III: on simulated data where the true gray/white matter ratio is 3.846, QuanTOF recovers 1.898 and PET-DPC recovers 1.868. The paper calls this \"strong quantitative consistency,\" but it's actually consistency with QuanTOF, not with the ground truth. The method is inheriting QuanTOF's bias. That's fine if you only claim to match QuanTOF, but the abstract says \"quantitative accuracy\" and \"surpassing traditional iterative methods,\" which overreaches. Also, the whole-body validation is only 5 cases, and no code or data is provided, so the reproducibility is limited.\n\nWho is this for? Anyone working on deep learning for PET reconstruction, especially on making diffusion models clinically practical. It deserves peer review — the flaws are addressable and the empirical finding that the correction steps matter is useful even if the derivation tightens up. My recommendation: send it out, but ask the authors to either derive the update or explicitly label it as an empirical heuristic, specify the scatter/random estimation, and soften the quantitative-accuracy claims to be relative to QuanTOF rather than absolute.\n\nI'd bring this to a reading group and likely cite it as related work, though I'd want the correction step clarified before building on it.","headline":"PET-DPC is a promising diffusion-based PET reconstruction with a physics-correction step that works empirically, but the correction is under-derived and the quantitative accuracy claim is anchored to QuanTOF rather than ground truth.","tokens_in":15367,"tokens_out":1500,"would_cite":true,"duration_ms":19168,"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":"PET-DPC, a conditional diffusion model with posterior physical correction, reconstructs PET images with quantitative accuracy matching iterative OSEM while cutting brain reconstruction time by about half and whole-body time by about 85%.","keywords":["PET reconstruction","diffusion model","posterior correction","quantitative accuracy","time-of-flight","list-mode","scatter correction","attenuation correction"],"falsifier":"In a Monte Carlo simulation with known true scatter/random sinograms, reconstruct the same list-mode data with PET-DPC using the correct estimate and with deliberately perturbed scatter/random estimates; if image metrics shift systematically with the perturbation, the correction is not robustly likelihood-based. More directly, compare the update direction in Algorithm 1 Step 5 against the numerical gradient of Eq. (14) at the same x̂0; a mismatch would show the 'posterior correction' is not the gradient it claims to be.","tokens_in":14562,"feed_emoji":"⚡","tokens_out":6614,"duration_ms":70096,"temperature":0.7,"pith_summary":"The paper proposes PET-DPC, a conditional diffusion model that reconstructs PET images by combining a learned denoiser with explicit physical corrections applied during the reverse sampling process. The central claim is that this physics-constrained sampling yields images quantitatively comparable to fully corrected OSEM/QuanTOF reconstructions, while outperforming end-to-end deep learning baselines in PSNR, SSIM, NRMSE, and tumor contrast, and generalizing to out-of-distribution phantom and simulation data. If true, the method would let clinical PET workflows replace slow iterative reconstruction with a fast diffusion-based pipeline—about 50% faster for brain scans and about 85% faster for whole-body scans—without sacrificing quantitative fidelity. The authors also report that the posterior correction, not the diffusion prior alone, is what restores background uniformity and tumor-to-background ratios in the NEMA phantom.","feed_headline":"Diffusion PET reconstruction matches iterative accuracy at half the time","feed_subtitle":"Posterior physics correction closes the quantitative gap of end-to-end deep learning on brain and whole-body scans.","key_machinery":"The central mechanism is Algorithm 1's posterior correction step: after Tweedie denoising produces x̂0′, the image is multiplied by an intensity-matching coefficient C_IM and multiplicatively updated by back-projecting the ratio of measured to expected prompt events, Σ Pᵀ (y / ȳ), a step claimed to inject scatter, attenuation, and random corrections into the sampling trajectory. The conditioning input is the GTP-image, a back-projected image where each list-mode line of response is spread by a TOF Gaussian and a geometric overlap weight, then normalized by a sensitivity map; this supplies raw-data geometry without storing sinograms. The correction step, not the neural denoiser, is what the a","core_discovery":"The paper's discovery is that a diffusion model whose sampling steps are interleaved with a posterior correction based on a Poisson measurement model can produce PET images with quantitative fidelity comparable to fully corrected iterative reconstruction. During sampling, the model uses list-mode data to apply scatter, attenuation, and random corrections in the image domain, and this correction step is what distinguishes PET-DPC from a plain conditional DDPM. On 100 brain validation datasets, PET-DPC reached PSNR 47.703, SSIM 0.994, and NRMSE 0.066, versus FastPET (37.749/0.941/0.213) and Conditional-DDPM (44.782/0.992/0.098); similar gains held on whole-body data. The authors further report","pith_inferences":["Editorial extension: The multiplicative update in Algorithm 1 Step 5 has the same ratio structure as MLEM updates. If validated as a likelihood-gradient surrogate, similar corrections could be grafted onto faster generative samplers, such as consistency or flow-matching models, to enforce PET physics at even lower step counts.","Editorial extension: Since the correction stage back-projects list-mode data in the image domain and the diffusion prior supplies regularization, the scheme may extend naturally to SPECT or to joint activity/attenuation estimation when a μ-map is available.","Editorial extension: A direct test is to run PET-DPC on simulated data with known true scatter/random sinograms and compare against reconstructions using the estimated ones; this would separate errors from the correction approximation from errors in scatter/random estimation.","Editorial extension: The ablation's large drop from one correction step to zero suggests the correction steps carry most of the physics; a systematic study trading denoiser steps for correction iterations could yield a faster protocol."],"forward_implications":["Clinical PET could be reconstructed in roughly half the time for brain scans and one seventh of the time for whole-body CBM scans, compared with the QuanTOF iterative baseline.","End-to-end deep learning PET methods that omit explicit physical modeling should expect degraded quantitative performance when applied out of distribution; physics-corrected diffusion sampling is the proposed remedy.","The number of posterior correction steps is a practical control knob: more steps improve PSNR/SSIM/NRMSE on brain validation data, at increased compute cost.","A single model trained on brain and whole-body data can transfer to NEMA phantom and simulated data without retraining, provided the posterior correction is active.","The quantitative gap between conditional diffusion and iterative reconstruction can be closed without giving up the speed advantage of deep learning."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model forward/reverse process and ancestral sampling that the method builds on.","marker":"[24]"},{"why":"Supplies the SDE formulation and predictor-corrector sampling strategy used to interleave posterior correction.","marker":"[25]"},{"why":"Provides the approximation ∇ log p(y|x_t) ≈ ∇ log p(y|x̂0) that turns detector data into a usable correction term.","marker":"[28]"},{"why":"QuanTOF reconstructions are the training labels and the iterative reference standard for quantitative comparison.","marker":"[29]"},{"why":"FastPET is the supervised end-to-end baseline; its histo-image input is compared with the proposed GTP-image.","marker":"[17]"},{"why":"Shepp–Vardi MLEM supplies the Poisson log-likelihood model underlying the correction equations.","marker":"[7]"},{"why":"GATE simulation toolkit generates the 20 simulated brain datasets used as out-of-distribution test data.","marker":"[33]"},{"why":"BrainWeb phantom provides the anatomical model and activity assignments used in the simulation study.","marker":"[30]"}],"fun_headline_variants":["Physics-corrected diffusion PET hits iterative accuracy in half the time","Physics-aware diffusion PET: iterative accuracy at deep-learning speed","Diffusion PET with posterior correction: fast and quantitatively faithful","Physics correction closes quantitative gap in diffusion PET"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the multiplicative update using measured-to-expected event ratios, combined with an intensity-matching coefficient, correctly approximates the Poisson-likelihood gradient; the paper does not derive this update and does not specify how the scatter and random sinograms used in the forward model are estimated.","fun_headline_variants_meta":{"raw":{"variants":["Physics-corrected diffusion PET hits iterative accuracy in half the time","Physics-aware diffusion PET: iterative accuracy at deep-learning speed","Diffusion PET with posterior correction: fast and quantitatively faithful","Physics correction closes quantitative gap in diffusion PET"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002123,"raw_usage":{"total_tokens":8123,"prompt_tokens":830,"completion_tokens":7293,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":7227}},"tokens_in":574,"tokens_out":7293,"duration_ms":53022,"temperature":1.0,"reasoning_tokens":7227,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:35:52.689347+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a Monte Carlo simulation with known true scatter/random sinograms, reconstruct the same list-mode data with PET-DPC using the correct estimate and with deliberately perturbed scatter/random estimates; if image metrics shift systematically with the perturbation, the correction is not robustly likelihood-based. More directly, compare the update direction in Algorithm 1 Step 5 against the numerical gradient of Eq. (14) at the same x̂0; a mismatch would show the 'posterior correction' is not the gradient it claims to be.","supporting_citations":[{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model forward/reverse process and ancestral sampling that the method builds on."},{"cited_title":"QUANTOF: multi -GPU-based list -mode fully quantitative TOF PET image reconstruction,","cited_arxiv_id":null,"evidence_quote":"QuanTOF reconstructions are the training labels and the iterative reference standard for quantitative comparison."},{"cited_title":"FastPET: near real-time reconstruction of PET histo-image data using a neural network,","cited_arxiv_id":null,"evidence_quote":"FastPET is the supervised end-to-end baseline; its histo-image input is compared with the proposed GTP-image."},{"cited_title":"Maximum likelihood reconstruction for emission tomography,","cited_arxiv_id":null,"evidence_quote":"Shepp–Vardi MLEM supplies the Poisson log-likelihood model underlying the correction equations."},{"cited_title":"GATE: a simulation toolkit for PET and SPECT,","cited_arxiv_id":null,"evidence_quote":"GATE simulation toolkit generates the 20 simulated brain datasets used as out-of-distribution test data."},{"cited_title":"A new improved version of the realistic digital brain phantom,","cited_arxiv_id":null,"evidence_quote":"BrainWeb phantom provides the anatomical model and activity assignments used in the simulation study."}],"review_version":1}