{"id":"44f7c217-42e2-4e2d-abc0-c5b06b8e0444","arxiv_id":"2506.03804","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Registering other subjects' PET images into a target subject's MR anatomy creates training data that, when used to pre-train a diffusion model, improves low-count PET reconstruction accuracy.","lead":"This paper makes fake, patient-specific PET images by warping other people's PET scans into the shape of a new patient's MRI, then uses them to train a diffusion model that reconstructs low-dose PET scans. The approach improves image quality and lesion visibility compared with standard diffusion-model methods, especially when only a small training set is available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's core assumption — that MR-derived deformation fields are valid for PET tracer distributions — is not directly tested; the simulated experiments guarantee it by construction, so the real-data result carries the empirical weight.","rationale":"Agree with the reader's weakest assumption; the concern is load-bearing because the method's novelty and advantage is precisely using MR-derived deformations to transplant PET data across subjects. In simulation, the assumption is guaranteed by the data-generation process, so those results cannot validate it; they only validate the downstream machinery. The real-data experiment is the relevant test, but its metrics are relative to OSEM and hyperparameters are selected on the test set, making it weaker than an independent validation. A targeted experiment with a non-MR-derived PET ground truth, or a real dataset with functional/pathological PET-MR mismatch, would settle whether the claimed trade-off survives when the assumption is violated. The central idea remains plausible and worth conditional acceptance; no internal inconsistency was found.","tokens_in":15361,"tokens_out":7364,"duration_ms":81274,"concrete_test":"Use a real or realistically simulated dataset with an independent PET ground truth that is not generated from the same MR segmentation (e.g., real full-count FDG PET with manual lesion masks, or a simulation in which a functional uptake component is displaced relative to MR anatomy). Apply the full pipeline (VoxelMorph MR-MR registration, pseudo-PET synthesis, diffusion training, PET-DDS reconstruction) with hyperparameters fixed on a separate tuning fold, and compare NRMSE in PET-MR mismatch regions and CRC-versus-noise against standard PET-DDS. If the proposed method no longer improves the trade-off, the claim is conditional on PET uptake tracking MR anatomy; if it still improves, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that pseudo-PET images, formed by warping other subjects' PET volumes with a deformation field learned from MR-to-MR registration (Section IV-A1), are plausible samples from the target subject's PET distribution. This is the load-bearing step: if PET uptake does not follow MR anatomy, the prior learned from pseudo-PET is biased and the reported improvement in NRMSE/CRC and bias-variance would not generalize. The paper's simulated data do not test this assumption because the simulated ground-truth PET is itself derived from real T1 MR segmentations (Section IV-B1), so PET-MR correspondence is perfect by construction and the MR-derived warp is the correct PET warp. The only evidence against this concern is the real [18F]FDG experiment, but there the reference is a 100%-count OSEM image rather than an independent ground truth, the number of subjects is small (~36), and the reported comparison selects the best hyperparameter setting on the test data. The paper acknowledges the challenge of PET-MR mismatch, but the lesion experiments insert lesions at test time only and never into training data, so they do not expose the deformation assumption to PET-only structures within the learned training distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for generating subject-specific 'pseudo-PET' training images by registering other subjects' paired PET-MR images into the target subject's MR space using VoxelMorph MR-to-MR deformation fields, optionally summing several warped PET images. A diffusion model is trained on these pseudo-PET images and used as the prior within the PET-DDS reconstruction algorithm, which is then applied to low-count (2.5%) simulated and real [18F]FDG PET sinograms. The authors evaluate the method against OSEM, Bowsher, PET-DDS, and PET-DDS+MR, reporting improved NRMSE/SSIM in PET-MR mismatch regions, better lesion CRC versus background noise in simulations, and improved bias-variance and NRMSE on real data. The central claim is that personalizing the diffusion prior with MR-registered pseudo-PET images improves low-count PET reconstruction accuracy and the trade-off between PET-specific and MR-shared features.","tokens_in":15635,"tokens_out":6960,"duration_ms":66243,"significance":"If the quantitative claims hold, the method is a simple and practical way to inject MR information into diffusion-based PET reconstruction without training a conditional model, and it addresses the PET-MR mismatch problem by learning a softer MR-guided prior. The approach is original in its use of registration-based pseudo-PET synthesis rather than deep generative synthesis, and the authors show that it can be useful in low-training-data regimes. The scope of the experiments is commendable: both simulated and real FDG data, multiple baseline algorithms, and a bias-variance analysis. However, the evaluation has several methodological weaknesses—test-set hyperparameter selection, missing error bars in the simulated lesion experiments, and the use of full-count OSEM as the real-data reference—that currently prevent the quantitative claims from being fully established. The paper does not provide code, but the method is described in sufficient detail to be reproduced.","major_comments":[{"comment":"The real-data comparison in Fig. 9 reports the 'best NRMSE' for each method across hyperparameter choices, and Fig. 10 selects hyperparameters to maximize SSIM on the test reconstruction. Because the regularization strength lambda_DDS and the number of iterations were chosen after seeing the test data, the reported improvements may reflect post-hoc selection rather than a reproducible operating point. Please report hyperparameters chosen by a validation set or a pre-specified rule, and show the full trade-off curves with the selected operating point marked.","section":"V-D, Fig. 9 and Fig. 10"},{"comment":"The bias-variance analysis and the real-data NRMSE in Figs. 9 and 11 use the 100%-count OSEM reconstruction as the ground truth. This is not an independent ground truth, and because all diffusion models are trained on OSEM images, the metric may favor reconstructions that are close to OSEM rather than accurate to the true tracer distribution. An independent reference (for example, a phantom with a known activity distribution, or a high-count reconstruction with a different algorithm) is needed to support the real-data claims.","section":"IV-D5 and V-D"},{"comment":"The central premise that an MR-derived deformation field is a valid transformation for PET tracer distributions is not directly tested. In the simulated data the ground-truth PET is analytically derived from MR segmentations, so the premise holds by construction; in the real data there is no independent ground truth and the subject population is not described with respect to pathology. The lesion experiments in Section V-B insert lesions only at test time, so the learned training distribution never contains PET-only structures. Please add a direct validation of the deformation assumption, for example by comparing warped PET images to native PET images of the same subject when available, or by reporting quantitative agreement between MR-registered and PET-registered pseudo-PET.","section":"IV-A1 and IV-B1"},{"comment":"Figures 5-7 report trade-off curves for the simulated lesion study without error bars or significance tests. It is not stated whether the metrics are averaged over multiple Poisson noise realizations, and several of the curves are close. Please provide means and standard deviations over at least 10 independent noise realizations (or an equivalent statistical summary) and a significance test for the claimed differences.","section":"V-B, Figs. 5-7"}],"minor_comments":[{"comment":"The sentence 'Over-fitting is not a concern for this task, as the test data is available during training' is confusing and, taken literally, describes a circular evaluation of the registration network. Please rephrase to clarify what is being personalized and why using the target MR for registration training is valid despite using that same MR for evaluation.","section":"IV-D3"},{"comment":"The heuristic for summing pseudo-PET images (P(N=n) proportional to 1/n and weights w_i ~ U[0,1]) is not justified, and since MR-reg & sum + PET-DDS 'performed similarly to MR-reg + PET-DDS' in Section V-B, the benefit of the summing step is not demonstrated empirically. Please provide supporting evidence or present the summing as an optional variant with limited effect.","section":"IV-A1"},{"comment":"The hyperparameters for PET-DDS+MR are not fully specified: lambda_MR is introduced as the guidance strength, but the text does not state which values were used or how it was varied. Please report the values or state explicitly that it was fixed.","section":"IV-C3"},{"comment":"There are typographical and formatting issues, including 'V oxelMorph' with extra spaces, 'bourne out' in Section V-C1, and the caption of Fig. 2 using 'blues' instead of a consistent term. These should be corrected.","section":"II-E and IV-A1"},{"comment":"The real-data evaluation omits MR-reg & sum + PET-DDS, even though random summing is part of the proposed method as described in Section IV-A1. The omission should be noted in the main text or the method should be evaluated on real data as well.","section":"V-D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for TRPMS and builds on the authors' previous MIC 2024 work. The main risk is the empirical rigor of the real-data evaluation; if the authors address the test-set selection, the OSEM-reference issue, and the absence of error bars in the simulated lesion study, the contribution would be publishable. There are no concerns about novelty disclosure or citation pattern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is real: instead of training a diffusion prior on generic multi-subject PET, warp other subjects' PET into the target subject's MR space via MR-to-MR registration, and train a personalized prior on those pseudo-PET images. That is a simple, sensible way to inject MR information without a large training set or a black-box generator, and it appears to help. The paper also deserves credit for what it does not do: it does not oversell PET-DDS+MR, it reports that classifier-free MR guidance failed to help on real data, and it includes ablations (PET-PET registration, MR side-channel) that mostly strengthen the case for the proposed method.\n\nThe main soft spot is the evaluation. The simulated data are built from MR segmentations, so PET-MR correspondence is perfect by construction; the simulated experiments therefore cannot test the load-bearing assumption that MR-derived deformation fields are valid for PET tracer distributions. The real-data experiment is the only test of that assumption, and there the reference is a 100%-count OSEM image, not ground truth, with hyperparameters selected on the test data (min NRMSE/max SSIM). The simulated figures also lack error bars. The paper openly states that the test data are used for registration training, which is legitimate for this task, but the overall evaluation still carries a selection-bias smell. None of this is fatal; the real-data results do show the method winning on bias-variance against the other diffusion baselines, and the paper is transparent about its choices. But the reader should not walk away thinking the deformation assumption is independently validated—it is not, on the evidence here.\n\nIf I were refereeing this, I would ask for repeated-realization error bars on the simulated metrics, a predefined hyperparameter selection protocol (e.g., validation-set selection), and ideally a real-data comparison against an independent high-quality reconstruction method instead of OSEM alone. The central claim is plausible and the method is simple enough to reproduce, but the evidence is currently conditional rather than conclusive.\n\nThis paper is worth engaging with: it gives a practical alternative to MR-conditioned diffusion training, and the pseudo-PET synthesis idea could have legs beyond PET. A serious referee should get it, though with the expectation of revision.","headline":"The pseudo-PET registration idea is a genuine, useful contribution; the evaluation is honest but has enough selection bias and weak-reference issues that the headline claim is supported rather than proven.","tokens_in":16150,"tokens_out":1417,"would_cite":true,"duration_ms":16527,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Warping other patients' PET scans into a target's MR space yields a personalized diffusion prior that reconstructs low-count PET better than generic priors","keywords":["PET reconstruction","diffusion models","MR-guided reconstruction","pseudo-PET synthesis","image registration","low-count PET","score-based generative models","personalized priors"],"falsifier":"On a cohort whose FDG-avid lesions are partly invisible or misaligned on T1 MR, generate pseudo-PET by MR-to-MR registration and reconstruct 2.5% count data; if lesion contrast recovery falls below that of PET-DDS trained on unregistered PET, the MR-transfer assumption is the cause.","tokens_in":15170,"feed_emoji":"🧠","tokens_out":10593,"duration_ms":105441,"temperature":0.7,"pith_summary":"Positron emission tomography at low dose is noisy, and using MR to guide reconstruction can silently impose anatomy that the tracer signal does not support. The paper proposes a softer form of MR guidance: warp other subjects' PET images into the target subject's MR-defined space using MR-to-MR registration, and train a diffusion model on these synthetic 'pseudo-PET' volumes as a personalized prior. Reconstructing with the PET-DDS algorithm, the paper reports that this prior improves accuracy from 2.5%-count FDG data, with lower error where PET and MR disagree, retained accuracy where they agree, better lesion contrast recovery relative to background noise, and a better bias-variance frontier on real data. The method needs no generative network for synthesis and works with fewer than about fifty paired PET-MR scans. If correct, it provides a practical route to personalized MR-guided PET reconstruction that preserves PET-only features such as lesions.","feed_headline":"Warped PET scans build a personalized prior for low-dose imaging","feed_subtitle":"Registering other patients' PET-MR pairs into one brain space improves lesion contrast and cuts noise at 2.5% counts.","key_machinery":"The central object is the pseudo-PET image: another subject's measured PET volume deformed into the target brain's coordinate frame by an affine alignment followed by a deep-learning deformable registration field learned from MR-to-MR alignment [24]. A random weighted sum of one or more such deformed images forms each training sample, so a small library of paired PET-MR scans yields a large, diverse, subject-specific training set without any generative model. The diffusion model is trained on these volumes with slice conditioning, since all volumes already share the target's slice geometry. Reconstruction then runs PET-DDS (Decomposed Diffusion Sampling), which alternates the diffusion denoiser with Poisson log-likelihood gradient steps on sinogram subsets, and the personalized prior is what anchors the iterate. The registration step is the load-bearing mechanism: it transfers anatomical context from MR into the PET training distribution while leaving the PET intensities measured, not hallucinated.","core_discovery":"The central claim is that a diffusion prior personalized to one subject's anatomy can be built by registering other subjects' measured PET images into that subject's MR space, and that this prior beats both a generic PET prior and an MR-conditioned prior for low-count 3D PET reconstruction. The pseudo-PET images mix real tracer measurements with MR-visible anatomy, so the learned score function becomes a joint anatomical-functional prior rather than an MR-imposing one. Using PET-DDS, the paper reports lower normalized root-mean-square error in mismatch regions, unchanged or better accuracy in agreement regions, and better lesion contrast versus background noise on simulated data with out-of-distribution lesions. On real FDG brain scans at 2.5% of full counts, it reports the best bias-variance trade-off among diffusion-based methods and improved agreement with full-count OSEM. The paper also reports that direct MR conditioning, either as an extra channel or through classifier-free guidance, does not consistently deliver these gains on small real datasets.","pith_inferences":["We infer that the random weighted summing step is itself a lightweight generative mechanism: it can synthesize an unbounded supply of plausible pseudo-PET volumes from a few registered scans, an idea the paper notes but leaves for future work.","We infer a boundary condition: any tracer distribution that is invisible on MR, such as early amyloid deposition or tumors without structural correlate, will stress the MR-transfer assumption, and mixing in unregistered native PET training images may hedge against that failure.","We infer that the shared target coordinate frame makes slice conditioning meaningful, so a patch-based or slab-wise training variant could cut the per-subject training cost while preserving the personalization benefit.","We infer that the deformation fields, being MR-driven, are tracer-independent; a library of precomputed fields could be reused for new tracers without retraining the registration network."],"forward_implications":["If the central claim holds, low-count brain PET can be reconstructed with a prior built from tens of paired scans instead of large corpora of high-quality images.","Out-of-distribution lesions should remain detectable: lesion contrast recovery per unit background noise improves without sacrificing accuracy in MR-PET agreement regions.","MR guidance becomes softer: the reconstruction can use anatomical information without requiring the tracer to match every MR boundary, reducing the blending artifacts seen with stronger MR priors.","The improvement should transfer across tracers and dose levels, because the registration uses MR anatomy and is therefore tracer- and dose-invariant.","On real data, the approach should dominate classifier-free MR conditioning when only limited paired training data are available, since the registration-based prior was the only diffusion variant that improved the bias-variance frontier."],"supporting_citations":[{"why":"Supplies the PET-DDS reconstruction algorithm and the MR-conditioned baseline it personalizes and compares against.","marker":"[9]"},{"why":"Supplies the deep-learning deformable registration framework used to map other subjects' MR and PET into the target MR space.","marker":"[24]"},{"why":"Supplies the simulation pipeline for FDG ground truth and the real brain PET-MR dataset used for evaluation.","marker":"[26]"},{"why":"Supplies the affine registration tool used to bring MR images into a common affine space before deformable registration.","marker":"[33]"},{"why":"Supplies the anatomically guided relative-difference prior baseline and its hyperparameter settings for comparison.","marker":"[17]"},{"why":"Supplies classifier-free guidance, the mechanism behind the PET-DDS plus MR comparison method.","marker":"[23]"},{"why":"Supplies the U-Net diffusion architecture used as the score network in all diffusion-based reconstructions.","marker":"[34]"},{"why":"Supplies the bias-variance evaluation methodology used to rank the reconstruction methods on real data.","marker":"[38]"}],"fun_headline_variants":["Warped PET-MR pairs craft a personalized diffusion prior","Synthetic pseudo-PET from MR-registered patients sharpens low-dose PET","Personalized PET prior from warped scans improves low-count reconstruction","MR-guided synthetic PET data boosts low-dose image quality","Register patients' PET-MR images to make a tailored low-dose PET prior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that warping another subject's brain into the target's MR space also carries that subject's PET tracer distribution into a plausible PET image for the target; if PET uptake does not track MR anatomy, the personalized prior is biased rather than helpful.","fun_headline_variants_meta":{"raw":{"variants":["Warped PET-MR pairs craft a personalized diffusion prior","Synthetic pseudo-PET from MR-registered patients sharpens low-dose PET","Personalized PET prior from warped scans improves low-count reconstruction","MR-guided synthetic PET data boosts low-dose image quality","Register patients' PET-MR images to make a tailored low-dose PET prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":3049,"prompt_tokens":983,"completion_tokens":2066,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1976}},"tokens_in":599,"tokens_out":2066,"duration_ms":16821,"temperature":1.0,"reasoning_tokens":1976,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:54:25.061153+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a cohort whose FDG-avid lesions are partly invisible or misaligned on T1 MR, generate pseudo-PET by MR-to-MR registration and reconstruct 2.5% count data; if lesion contrast recovery falls below that of PET-DDS trained on unregistered PET, the MR-transfer assumption is the cause.","supporting_citations":[{"cited_title":"Global image registration using a symmetric block- matching approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the affine registration tool used to bring MR images into a common affine space before deformable registration."},{"cited_title":"Evaluation of parallel level sets and Bowsher’s method as segmentation-free anatomical priors for time-of-flight PET reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the anatomically guided relative-difference prior baseline and its hyperparameter settings for comparison."},{"cited_title":"Bootstrap-optimised regularised image reconstruction for emission tomography,","cited_arxiv_id":null,"evidence_quote":"Supplies the bias-variance evaluation methodology used to rank the reconstruction methods on real data."}],"review_version":1}