{"id":"926336e1-9bbc-4446-9ecd-06a9a7b747d5","arxiv_id":"2412.11776","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Diffusion posterior sampling jointly reconstructs activity and attenuation in PET, outperforming MLAA on 2D XCAT phantoms even without time-of-flight information.","lead":"This paper applies diffusion posterior sampling to jointly reconstruct the radioactive tracer distribution and the attenuation map in PET from emission data alone. On 2D XCAT phantom simulations, it reports that this approach outperforms the standard MLAA method, particularly in non-time-of-flight PET where crosstalk usually degrades results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-TOF out-of-distribution reconstruction is untested; prior artifacts seen in TOF OOD cases may be worse without TOF.","rationale":"The reader's weakest assumption is that the learned prior generalizes to unseen anatomies. Our concern is a specific and untested consequence of that assumption: the paper's OOD tests are only performed with TOF data, leaving the non-TOF OOD regime—where the prior is most needed—completely unexamined. This is load-bearing because the abstract and conclusions make a strong non-TOF claim, and the paper's own TOF OOD results reveal prior-induced artifacts. A non-TOF OOD experiment would settle whether the claim extends beyond the training distribution. We partially agree with the reader: the prior-generalization assumption is indeed central, but the most concrete missing evidence is the non-TOF OOD experiment, not just a general concern about prior accuracy. We also note the small sample size and lack of statistical testing for the DPS-vs-DPS2 advantage, which is a secondary but relevant issue. The verdict remains CONDITIONAL: the idea is promising, but the evidence is insufficient to fully support the non-TOF OOD claim, and the proposed test would either strengthen or refine the conclusion.","tokens_in":6751,"tokens_out":8674,"duration_ms":73748,"concrete_test":"Using the same OOD phantom pairs from Figure 3(g)-(k) (the Gaussian tumor in activity, matched attenuation), generate non-TOF sinogram data with the same forward projector and count level as in Section 3.1. Reconstruct with DPS, DPS2, and MLAA in non-TOF mode, using the same hyperparameters (including τ=5e-1). Compare PSNR/SSIM and the activity profile through the tumor center (as in Figure 4) against the TOF OOD results. If DPS non-TOF either fails to separate activity and attenuation or produces artifacts substantially worse than in the TOF case, the central claim of robust non-TOF reconstruction is not supported for out-of-distribution anatomies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that DPS achieves near-perfect, noise-free joint reconstruction 'even in absence of TOF data'. The supporting experiments use 10 test slices drawn from the same XCAT distribution as the training set (no tumors). The only out-of-distribution (OOD) tests, which include tumors, are performed exclusively with TOF data (Section 3.2 and Figure 3). This matters because in the non-TOF setting the likelihood provides no time-of-flight information to separate activity from attenuation, so the learned prior must carry almost the entire burden of resolving crosstalk. The OOD TOF results already show the prior's strong inductive bias: a Gaussian tumor is reconstructed as a piecewise-constant shape (Figure 3(j) and Figure 4), and the attenuation map inherits the same artifacts. If this bias is present even with TOF data, there is a serious risk that non-TOF OOD reconstructions will exhibit severe crosstalk or activity/attenuation mixing, undermining the 'even in absence of TOF' claim for any anatomy outside the training distribution. The paper does not report non-TOF OOD results, so this critical regime is unverified. In addition, the comparison of DPS vs DPS2 (used to claim benefit from modeling dependencies) is based on only 10 slices with no error bars or significance testing, so that secondary claim is also not robustly established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a joint reconstruction of the activity and the attenuation in PET using diffusion posterior sampling (DPS). The authors train a diffusion model on pairs of 2-D XCAT activity/attenuation images, then use DPS to sample from the posterior conditioned on TOF or non-TOF emission data. Experiments on 10 test slices report higher PSNR/SSIM than MLAA and than a variant with independently trained priors (DPS2), for both TOF and non-TOF data. Out-of-distribution tests with tumors are performed only with TOF data. The paper concludes that DPS mitigates activity-attenuation crosstalk in non-TOF settings.","tokens_in":7015,"tokens_out":5048,"duration_ms":48265,"significance":"The central idea is sound and timely: a learned joint prior over activity and attenuation is a plausible route to resolve the cross-talk that limits non-TOF MLAA. The paper cleanly separates the joint-prior model (DPS) from an independent-prior baseline (DPS2), which helps isolate the effect of modeling dependencies. The mathematical formulation follows the standard DPS derivation and the forward model is physical. If confirmed on larger and more diverse datasets, the results would be an important step toward emission-only attenuation correction. The main weaknesses are the small test set and the absence of non-TOF out-of-distribution experiments, which are needed to support the abstract's central claim.","major_comments":[{"comment":"The claim that DPS 'significantly outperforms' MLAA and DPS2 is not supported by any statistical analysis. The evaluation uses 10 test slices; Figure 2 shows only point clouds and means without error bars, confidence intervals, or significance tests. Several reported differences (e.g., DPS no TOF vs DPS2 no TOF for attenuation, with PSNR 25.11 vs 24.25) appear small relative to inter-slice scatter. Please provide paired comparisons with confidence intervals, per-metric distributions, and a justified sample size, or temper the wording.","section":"Section 3.2, Figure 2"},{"comment":"The out-of-distribution evaluation is performed exclusively with TOF data. The paper's principal claim is that DPS works 'even in absence of TOF data'; in the non-TOF setting the likelihood provides no TOF information to separate activity from attenuation, so the learned prior must carry nearly the entire disambiguation burden. The OOD TOF results already show a strong prior bias (a Gaussian tumor reconstructed as piecewise-constant in Figure 3(j) and Figure 4). Without non-TOF OOD experiments, the central claim is not verified for anatomies outside the training distribution. Please add non-TOF OOD reconstructions or explicitly restrict the claim to in-distribution anatomies.","section":"Section 3.2, Figure 3"},{"comment":"The test set consists of only 10 slices drawn from the same XCAT distribution as the training set, with no tumors. The near-perfect in-distribution results may therefore reflect memorization rather than generalization; the paper itself raises this concern in Section 3.2, but the two OOD TOF cases are too few to resolve it. A larger held-out set with varied anatomies and pathologies is needed to substantiate the feasibility claim.","section":"Section 3.1"}],"minor_comments":[{"comment":"The sentence 'We then performed applied MLAA and DPS' contains a duplicated verb; please revise.","section":"Section 3.2"},{"comment":"The norm is missing a closing parenthesis: it should read ∥s_θ(x_t,t) − ∇_{x_t} log p_t(x_t|x_0)∥₂².","section":"Equation (9)"},{"comment":"The step sizes ζ_t and ξ_t, the number of diffusion steps T, and the noise schedule α_t are not reported; only τ = 5×10⁻¹ is given. Please provide these values to enable reproduction.","section":"Algorithm 1 and Section 3.1"},{"comment":"Adding per-method error bars or interquartile ranges would help the reader assess the overlap between DPS and DPS2.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a credible feasibility study, but the evidence is not yet at the level of the abstract's strongest claims. The limited test size and the absence of non-TOF OOD results are the main barriers. If the authors can add non-TOF OOD experiments and basic statistical comparisons, a revised version would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid feasibility study, not a breakthrough. The genuinely new piece is applying diffusion posterior sampling to the joint estimation of activity and attenuation in PET, with a joint prior over the pair instead of separate priors, and testing the non-TOF regime where MLAA suffers from crosstalk. The math is standard DPS applied to a two-channel image; no new theory, but the application is nontrivial and the comparison between joint and independent priors is a useful experimental point.\n\nThe paper does some things well. The authors are honest about limitations: they flag 2D phantom data, lack of anatomical diversity, and show out-of-distribution TOF results where the prior's piecewise-constant bias distorts a Gaussian tumor. That is real evidence about the method's behavior, not hiding. The forward model is physical, the training and test slices come from different phantoms, so the in-distribution results are not circular.\n\nThe soft spots are real but mostly about overstatement and thin statistics. Only 10 test slices, no error bars or significance tests, yet the abstract says DPS \"significantly outperforms\" MLAA. \"Near perfect-resolution and noise-free\" is a strong claim for a 10-image, in-distribution evaluation. The out-of-distribution tests are only TOF; the stress-test note is right that non-TOF OOD is unverified, and since the prior has to carry more weight without TOF, that's a genuine gap. The paper doesn't explicitly claim non-TOF OOD works, but the abstract's \"even in absence of TOF information\" is too broad without that data. Code and hyperparameters are not fully specified, so replication would take effort.\n\nThe central feasibility claim — that DPS can resolve crosstalk in non-TOF on in-distribution phantoms — holds up on the evidence. The comparison DPS vs DPS2 is suggestive but not robust. This paper is for people working on PET reconstruction or conditional generation for inverse problems; they will find it a useful data point, not a definitive solution. It deserves a serious referee: the idea is important enough and the experiments are honest enough to warrant peer review, but the revision should add error bars, clarify the statistical language, and ideally include non-TOF OOD results or at least soften the claim.","headline":"Feasibility study shows DPS can beat MLAA on non-TOF in-distribution phantoms, but the evidence is thin and the OOD non-TOF regime is untested.","tokens_in":7559,"tokens_out":2396,"would_cite":false,"duration_ms":25503,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffusion posterior sampling jointly reconstructs PET activity and attenuation from emission data alone, outperforming MLAA without time-of-flight information in 2D phantom tests.","keywords":["positron emission tomography","joint activity and attenuation reconstruction","diffusion posterior sampling","non-time-of-flight PET","maximum likelihood activity and attenuation","attenuation correction","score-based generative models","computational phantom"],"falsifier":"Run DPS on real 3D patient PET data for which a CT-derived attenuation map and an independently corrected activity image are available: if non-TOF DPS reconstructions fail to reproduce the CT-based attenuation map or introduce training-set biases on anatomies not present in the phantom data, the central claim that DPS resolves crosstalk without TOF would be refuted.","tokens_in":6573,"feed_emoji":"🩻","tokens_out":7990,"duration_ms":67447,"temperature":0.7,"pith_summary":"This feasibility study claims that diffusion posterior sampling (DPS) can solve joint reconstruction of activity and attenuation from PET emission data alone, without CT or MRI and without time-of-flight (TOF) information. The authors train a diffusion model on pairs of activity and attenuation maps from computational phantoms, then use the learned joint prior to guide reconstruction from raw PET projections. In 2D phantom tests, they report that DPS reconstructions are close to noise-free and close to ground-truth resolution, while MLAA without TOF suffers from severe activity–attenuation crosstalk. They also report that the jointly trained DPS beats a variant trained on the two channels independently, suggesting that modeling the dependency between activity and attenuation is what carries the improvement. The paper is explicitly a feasibility study: it is limited to 2D phantom data, and its own out-of-distribution tumor tests reveal that the learned prior imposes piecewise-constant artifacts on shapes that did not appear in training.","feed_headline":"Diffusion sampling reconstructs PET activity and attenuation maps","feed_subtitle":"On 2D phantom tests, joint diffusion sampling beats MLAA and resolves the non-TOF crosstalk.","key_machinery":"The central object is the joint score function that approximates the gradient of the log-density of noisy two-channel images, trained by score matching on paired activity and attenuation slices. During reconstruction, DPS approximates the conditional score by Bayes' rule, with the likelihood term evaluated at Tweedie's denoised estimate of the clean image; the reverse diffusion step is then followed by gradient updates on the activity and attenuation channels separately. The joint score is what distinguishes DPS from its independent-channel variant: the latter factorizes the score into separate activity and attenuation networks, so it assumes independence, and the paper's comparison isolates the contribution of the joint prior.","core_discovery":"The paper's central claim is that the activity–attenuation crosstalk that breaks non-TOF joint reconstruction can be resolved by replacing the missing joint prior with a diffusion-model score function trained on paired activity and attenuation images. In the DPS framework, the posterior over the two-channel image is sampled by alternating the reverse-diffusion update with a likelihood-gradient step computed from the forward model; the score network supplies the prior, and the joint training is what keeps the two channels consistent with each other. On 2D phantom testing slices, the paper reports that DPS without TOF outperforms MLAA with TOF on PSNR and SSIM for both activity and attenuation, and that DPS without TOF outperforms the independently trained variant with TOF. The authors interpret these results as evidence that explicitly modeling activity–attenuation dependencies is more valuable than adding TOF information, at least within the tested phantom setting.","pith_inferences":["A testable extension is to quantify how much of the gain is prior strength versus joint modeling by comparing the jointly trained score against a score trained on paired images with the channel pairing scrambled, which would isolate the dependency-learning effect.","The out-of-distribution artifacts suggest that DPS may act as a learned shape prior rather than a generic statistical prior; if so, its clinical value hinges on the coverage of the training set, much like other learned reconstruction methods.","Since the paper ignores scatter and random coincidences, a natural next experiment is to include background terms in the forward model; a mismatched likelihood gradient could shift the crosstalk balance and test how much the prior compensates.","The result also hints that TOF information may be partly redundant when a strong joint prior is available; a direct comparison of TOF and non-TOF DPS on the same phantom would quantify the remaining value of TOF under DPS."],"forward_implications":["If the central claim holds, non-TOF PET scanners could perform joint activity and attenuation reconstruction from emission data alone, removing the need for CT or MR attenuation correction in some workflows.","The observed superiority of jointly trained DPS over independently trained channels implies that cross-channel dependencies in the prior are the key mechanism for suppressing crosstalk, a lesson transferable to other joint reconstruction problems.","The success of a learned prior on phantom data suggests that a sufficiently rich training corpus of activity and attenuation pairs could make DPS a practical alternative to MLAA in low-dose or PET-only settings.","The out-of-distribution tumor results indicate that performance is tied to the training distribution, so deployment would require training data covering the relevant clinical variability.","Extending the approach to 3D volumes and real patient data is the stated next step, with the paper reporting encouraging preliminary findings along that path."],"supporting_citations":[{"why":"Supplies the diffusion posterior sampling algorithm used to combine the learned prior with the PET likelihood.","marker":"[6]"},{"why":"Provides the denoising diffusion probabilistic model formulation and the Tweedie-style denoised estimate used in the reverse diffusion update.","marker":"[12]"},{"why":"Supplies the computational phantom and the paired activity/attenuation images used for training and testing.","marker":"[10]"},{"why":"Defines the MLAA algorithm that serves as the baseline and as the framework that DPS extends for joint reconstruction.","marker":"[1]"},{"why":"Reports similar DPS-based joint reconstruction gains in multi-energy CT, cited as supporting evidence for the joint-prior mechanism.","marker":"[13]"}],"fun_headline_variants":["Diffusion sampling outdoes MLAA for joint PET reconstruction","Joint PET activity and attenuation via diffusion prior","Diffusion sampling resolves non-TOF PET crosstalk","PET joint reconstruction via diffusion posterior sampling","Diffusion posterior sampling beats MLAA on non-TOF PET"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's advantage rests entirely on the learned joint prior being a faithful stand-in for real activity and attenuation maps; the paper's own out-of-distribution tumor experiments show that when an anatomy falls outside the training distribution the reconstructed shapes become piecewise-constant artifacts, so this prior fidelity is the assumption most at risk.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion sampling outdoes MLAA for joint PET reconstruction","Joint PET activity and attenuation via diffusion prior","Diffusion sampling resolves non-TOF PET crosstalk","PET joint reconstruction via diffusion posterior sampling","Diffusion posterior sampling beats MLAA on non-TOF PET"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2043,"prompt_tokens":852,"completion_tokens":1191,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":1114}},"tokens_in":468,"tokens_out":1191,"duration_ms":7749,"temperature":1.0,"reasoning_tokens":1114,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:35:25.452782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DPS on real 3D patient PET data for which a CT-derived attenuation map and an independently corrected activity image are available: if non-TOF DPS reconstructions fail to reproduce the CT-based attenuation map or introduce training-set biases on anatomies not present in the phantom data, the central claim that DPS resolves crosstalk without TOF would be refuted.","supporting_citations":[{"cited_title":"Simultaneous reconstruction of activity and attenuation in time-of-flight PET","cited_arxiv_id":null,"evidence_quote":"Defines the MLAA algorithm that serves as the baseline and as the framework that DPS extends for joint reconstruction."}],"review_version":1}