{"id":"e85a03b1-d157-4076-9456-f2c5a2d32c18","arxiv_id":"2411.16629","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LegoPET is a diffusion model conditioned on hierarchical features from a learned U-Net, improving sinogram-to-PET reconstruction metrics on simulated brain data.","lead":"LegoPET combines a pretrained U-Net's hierarchical feature maps with a conditional diffusion model to reconstruct PET images from sinograms. On simulated brain PET slices it reports modest PSNR/SSIM gains over a standard conditional diffusion baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised gain over cDPM may be an artifact of classifier-free guidance hyperparameters, not of the PnPNet feature guidance; the ablation that isolates the contribution disables classifier-free guidance.","rationale":"The reader correctly identifies missing error bars and missing final hyperparameters as key weaknesses. My stress-test goes further: the missing pdp and lambda_2 values are not just reporting omissions; they create a direct confound between the proposed feature guidance and the classifier-free guidance mechanism. The ablation in Fig. 3 is the only clean isolation of the proposed module, and it disables classifier-free guidance entirely, so Table 1's improvement cannot be attributed to PnPNet features unless the same guidance settings were used for both cDPM and LegoPET, which is not stated. This is a load-bearing issue because the paper's central claim is about the feature guidance improving cDPM performance. The proposed test is practical and would settle whether the 0.59 dB gain is real under equal guidance. I do not recommend rejection because the method is plausible, the code is promised, and the ablation suggests a real benefit; rather, the manuscript needs a hyperparameter-disclosed, apples-to-apples comparison with significance testing before the claim is accepted.","tokens_in":62,"tokens_out":1267,"duration_ms":72474,"concrete_test":"Rerun Table 1 while fixing pdp = 0 and λ2 = 0 for both cDPM and LegoPET (same random seeds, same T = 1000, same checkpoint selection), and report mean ± std over the 550 test slices. If the PSNR/SSIM gap narrows to less than 0.1 dB / 0.001 SSIM or loses significance under a paired test, the central claim that hierarchical feature guidance improves cDPM performance is not supported by the reported experiments.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that injecting hierarchical PnPNet features into a cDPM improves reconstruction accuracy and convergence (Section 3.5, Fig. 3). The paper's headline comparison, however, is between LegoPET (trained with pdp in {0.1, 0.2, 0.5} and inference guidance λ2) and a cDPM baseline that, by the authors' own description in Section 3.3, is 'LegoPET w/o guidance' — yet the numbers in Table 1 are reported without disclosing which pdp and λ2 were used, and the ablation in Section 3.5 explicitly disables classifier-free guidance (pdp = 0, λ2 = 0) on both sides. This creates a confound: the 0.59 dB PSNR gain in Table 1 could be driven by the classifier-free guidance setup (choice of pdp and λ2, inference-time sampling) rather than by the PnPNet feature injection itself. The paper states 'We chose the final checkpoints ... without any specific selection strategies' but never reports the final pdp and λ2, so the main result cannot be reproduced or attributed. The only experiment that isolates the proposed module (Fig. 3) shows a smaller and uncontrolled benefit, with no error bars, and the test set is only 550 slices from two simulated BrainWeb phantoms, making the 0.59 dB gain potentially within slice-to-slice variance. If the Table 1 numbers were obtained at a high guidance strength, the comparison is not apples-to-apples with the cDPM baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LegoPET, a conditional diffusion model for PET image reconstruction from sinograms, augmented by hierarchical features extracted from a separately trained convolutional U-Net (PnPNet). The features are injected as biases into the diffusion U-Net, and training uses classifier-free guidance with random conditioning dropout. Experiments on simulated 2D 18F-FDG BrainWeb slices report PSNR/SSIM improvements over a cDPM baseline and over cGAN/Pix2Pix, with visual quality claimed to be better than that of DeepPET. An ablation compares LegoPET with cDPM over 500 epochs with classifier-free guidance disabled.","tokens_in":18,"tokens_out":4533,"duration_ms":92070,"significance":"If the reported gains are robust, the work offers a simple recipe for injecting learned multiscale image priors into cDPMs for sinogram-to-image reconstruction, with potential applicability to other inverse problems. The code is publicly available, and the ablation directly addresses the contribution of feature guidance separately from classifier-free guidance. However, the significance is currently limited by the absence of statistical analysis, incomplete hyperparameter reporting, and an overstated comparison to DeepPET on PSNR.","major_comments":[{"comment":"The abstract and Section 1 claim that LegoPET 'surpasses recent DL-based PET image reconstruction techniques in terms of visual quality and pixel-level PSNR/SSIM metrics,' but Table 1 shows DeepPET achieves higher PSNR (28.30 dB) than LegoPET (27.59 dB). The text in Section 3.4 correctly softens this to 'comparable PSNR values,' but the abstract and introduction should be revised to avoid a claim that the table contradicts.","section":"Section 3.4, Table 1"},{"comment":"The paper reports that LegoPET models were trained with pdp in {0, 0.1, 0.2, 0.5} and that inference uses a guidance scale lambda_2, but it never states which pdp and lambda_2 produced the Table 1 results. Because the ablation in Section 3.5 disables classifier-free guidance (pdp = 0, and presumably lambda_2 = 0) on both sides, the 0.59 dB PSNR gain in Table 1 cannot be attributed to the PnPNet feature guidance; it could be driven by the classifier-free guidance hyperparameters. The final pdp and lambda_2 values, and an ablation with guidance enabled on both cDPM and LegoPET, must be reported.","section":"Sections 3.2 and 3.5"},{"comment":"All quantitative claims rest on a single evaluation over 550 slices from two BrainWeb phantoms, with no error bars, multiple training runs, or significance tests. The 0.59 dB PSNR improvement over cDPM and the early-epoch advantage shown in Fig. 3 may be within slice-to-slice or run-to-run variation. Report per-subject metrics and means with standard deviations or confidence intervals, and ideally results from multiple seeds.","section":"Sections 3.4 and 3.5"},{"comment":"The mechanism for injecting bd and bm into the cDPM is described only as adding them 'as extra biases' to the encoder and middle blocks. No equation specifies how the feature maps are transformed, at which resolutions they are added, whether the PnPNet is frozen during diffusion fine-tuning, or how the layers are time-conditioned. This underspecification prevents reproduction and should be formalized.","section":"Sections 2.1 and 2.2"},{"comment":"LegoPET is reported with the same 35.71 M parameters as cDPM, but LegoPET also contains the pretrained PnPNet. If the PnPNet parameters are excluded from the count, the table is misleading and the efficiency comparison with baselines is incomplete; clarify whether PnPNet parameters are included or excluded and report both trainable and total parameter counts.","section":"Table 1"}],"minor_comments":[{"comment":"There are grammatical errors and typographical artifacts, such as 'challenges such as maintain correspondence and consistency' and 'hierarchicaL' in the abstract; these should be corrected.","section":"Abstract and Introduction"},{"comment":"The statement that LegoPET has 'the second lightest color in the squared error maps' is ambiguous without a colorbar or scale; please add a colorbar or describe the error maps quantitatively.","section":"Figure 2"},{"comment":"The baselines are said to be trained 'until full convergence,' but no convergence criterion, number of epochs, or early-stopping rule is provided; please report these details for each baseline.","section":"Section 3.3"},{"comment":"Reference [18] is a conference abstract without a DOI or full peer-reviewed details; the authors should either provide a complete citation or reduce reliance on it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical imaging methods journal, and the public code release is a strength. The main concerns are reporting integrity and statistical rigor rather than fundamental flaws in the proposed idea. I would not recommend rejection, but the claims need to be aligned with the data and the key hyperparameters reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading. First, the core idea is reasonable and honestly presented: take a pretrained U-Net (PnPNet), extract hierarchical features, and inject them as biases into a sinogram-conditioned diffusion model. That is a sensible adaptation of multiscale structure guided diffusion and ControlNet, and the DWT loss is a small but real twist. Second, the central evidence for the method is weaker than the abstract implies. The 0.59 dB PSNR gain over cDPM in Table 1 is reported without disclosing the final pdp and lambda_2 values, and the ablation in Section 3.5 that isolates the feature guidance disables classifier-free guidance entirely. So the headline comparison may be partly driven by guidance hyperparameters, not by the feature injection itself. That is a real confound, and it is the paper's biggest soft spot.\n\nWhat the paper does well: it is clearly written, the method section is easy to follow, and the experimental setup is standard for the subfield. The PnPNet feature extraction is a legitimate contribution that builds on prior work without overclaiming entirely; the authors do acknowledge DiffPET and cite relevant diffusion guidance literature. The ablation (Fig. 3) does show consistent improvement across epochs even with pdp=0, which weakly supports the mechanism, and the visual results are plausible even if subjective. The code is promised, which is good practice.\n\nSoft spots, in proportion. The confound above is load-bearing for the main quantitative claim, and it is not hidden: Section 3.2 says pdp was varied but never reports the final choice, and Section 3.5 explicitly sets pdp=0. The test set is 550 slices from two simulated BrainWeb phantoms, with no error bars or significance tests; a 0.59 dB gain could easily be within slice-to-slice variance. DeepPET actually has higher PSNR (28.30 vs 27.59), so the claim to 'surpass' recent DL methods is only true for perceptual metrics, which are not quantified beyond a subjective figure. No comparison against other diffusion-based PET reconstruction approaches (e.g., contrastive diffusion) is included. These issues are fixable but require additional experiments or softened claims.\n\nThe citation pattern is fine; self-citation to DiffPET is minor and not abusive. The absence of reported hyperparameters for the main result is the most concrete reproducibility problem.\n\nWho this is for: researchers working on diffusion-based medical image reconstruction, particularly PET, who want a concrete example of feature-guided conditioning. The paper deserves a serious referee because the method is plausible and the ablation, despite its limitations, suggests the feature guidance does help. It is not a breakthrough, but with proper hyperparameter disclosure and error bars it could be a solid workshop or short conference paper.\n\nMy recommendation: send it to peer review, but flag the confound and the missing hyperparameters as major issues. The authors should rerun the Table 1 comparison at pdp=0 or report the exact settings, and add error bars or multiple trials. If they cannot cleanly separate the guidance effect from the feature guidance effect, the main claim should be downgraded.","headline":"LegoPET is a plausible incremental method for diffusion-based PET reconstruction, but its headline gain over cDPM is not cleanly attributable to the proposed feature guidance because the main comparison and the ablation use different inference settings.","tokens_in":7203,"tokens_out":782,"would_cite":false,"duration_ms":8832,"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":"LegoPET injects hierarchical learned features into a sinogram-conditioned diffusion model to improve PET reconstruction quality and training efficiency.","keywords":["PET reconstruction","conditional diffusion models","hierarchical feature guidance","sinogram-to-image","classifier-free guidance","plug-and-play prior","high-frequency detail","generative medical imaging"],"falsifier":"Run the LegoPET-versus-cDPM comparison over at least five independent training seeds and report the mean and standard deviation of PSNR/SSIM; if the 0.59 dB gap falls within one standard deviation of the two methods' distributions, the claimed advantage is not established. A second test is to evaluate both methods on held-out clinical or 3D PET volumes and check whether the visual and SSIM gains persist outside the simulated 2D phantom setup.","tokens_in":6134,"feed_emoji":"🧠","tokens_out":9389,"duration_ms":77097,"temperature":0.7,"pith_summary":"The paper tackles a specific failure mode of conditional diffusion models (cDPMs) in PET: when the conditioning input is a sinogram and the target is an image, standard cDPMs struggle to keep the output aligned with the input, tend to miss high-frequency detail, and train slowly. LegoPET's proposal is to train a separate U-Net (PnPNet) on sinogram-PET pairs, then add its multilevel feature maps as biases into a standard cDPM's encoder and middle blocks. On simulated 2D FDG brain-slice data, the paper reports that this guidance lifts PSNR/SSIM from 27.00 dB/0.945 for a vanilla cDPM to 27.59 dB/0.956, yields the most visually realistic reconstructions among the compared methods, and shows strong performance even early in training. The purpose is to show that structured feature guidance is a cheap, effective upgrade for diffusion-based PET reconstruction.","feed_headline":"LegoPET boosts sinogram-to-PET diffusion with multiscale priors","feed_subtitle":"Multiscale features from a learned prior net lift diffusion PET reconstruction fidelity and training speed.","key_machinery":"PnPNet is a conditional convolutional U-Net using the same backbone as the diffusion denoiser, trained end-to-end with $\\mathcal{L}_{MSE} + 0.1\\,\\mathcal{L}_{DWT}$ so that its latent features include high-frequency information. From its latent space, two lists of feature maps, $b_d$ from downsampling blocks and $b_m$ from middle blocks, are added as extra biases to the corresponding cDPM blocks. The diffusion model is trained with the standard noise-prediction objective and classifier-free guidance, randomly dropping conditioning information with probability $p_{dp}$; at inference it denoises with the extrapolated prediction $\\tilde{\\epsilon}_\\theta = (1+\\lambda_2)\\epsilon_\\theta(x_t,t,C) - \\lambda_2\\epsilon_\\theta(x_t,t)$. The role of this machinery is to give the denoiser structured prior information at every scale instead of a single-domain conditioning signal.","core_discovery":"LegoPET's central claim is that a diffusion model conditioned only on a sinogram benefits measurably from being biased, at multiple resolutions, by features extracted from a pretrained regression network. PnPNet is trained with an MSE loss plus a discrete-wavelet-transform loss that emphasizes the high-frequency spectrum, so its downsampling and middle-block features carry both coarse anatomy and sharp edges. Added as biases to the cDPM's encoder and middle blocks, these features act as a multiscale prior that keeps the reconstructed image consistent with the sinogram and preserves fine detail. The paper reports the best SSIM among all compared methods (0.956 vs 0.945 for cDPM), a 0.59 dB PSNR improvement over cDPM, and visually cleaner images than DeepPET, cGAN, Pix2Pix, and cDPM, all with the same 35.71 M parameter count as the vanilla cDPM.","pith_inferences":["If the PSNR/SSIM gains survive repeated training runs, the same feature-bias injection could transfer to other cross-domain inverse problems where conditioning and output live in different spaces, such as low-dose CT or MRI reconstruction.","The discrete-wavelet-transform term in PnPNet's loss is the most likely carrier of the edge-preservation benefit; an ablation that removes only that term would isolate its contribution, and the paper does not report one.","Because the evaluation is slice-wise on simulated 2D phantoms, the design leaves axial correlations unexploited; a 3D extension could either amplify the gains through richer context or erode them through higher memory and data demands.","The identical parameter count suggests the improvement is geometric rather than capacity-driven, implying the guidance could potentially be distilled into a single network to avoid running PnPNet at inference, which the paper does not explore."],"forward_implications":["Because LegoPET matches cDPM's 35.71 M parameter count, the reported gains are attributed to the injected features rather than to added capacity, so the method is a drop-in guidance upgrade.","The ablation study shows LegoPET outperforming cDPM at every checkpoint across 500 epochs, which implies the feature biases accelerate convergence and may allow shorter training schedules.","The highest SSIM and best visual quality among the compared methods position feature-guided diffusion as a baseline for future sinogram-to-PET reconstruction work.","The DWT loss in PnPNet training ties edge and texture preservation directly to the guidance, so the method's high-frequency advantage is a designed consequence rather than an accident.","The paper states that extending the model to an efficient 3D version and validating on patient data are the next steps, meaning clinical relevance depends on those results."],"supporting_citations":[{"why":"Provides the simulated brain phantom dataset from which all training, validation, and test slices are derived.","marker":"[1]"},{"why":"Serves as the regression-based baseline and motivates the need for DL reconstruction that avoids blurry outputs.","marker":"[4]"},{"why":"Supplies the denoising diffusion probabilistic model formulation that LegoPET adapts for its forward and reverse processes.","marker":"[7]"},{"why":"Supplies the classifier-free guidance training strategy and is also the vanilla cDPM baseline that LegoPET claims to improve.","marker":"[8]"},{"why":"Is one of the GAN-based baselines (Pix2Pix) that LegoPET compares against on PSNR/SSIM and visual quality.","marker":"[10]"},{"why":"Is the conditional GAN baseline (cGAN) used in the quantitative comparison.","marker":"[12]"},{"why":"Motivates the discrete-wavelet-transform loss term used to supervise PnPNet's high-frequency features.","marker":"[13]"},{"why":"Motivates the idea of guiding a diffusion model with multiscale structure from a learned prior.","marker":"[15]"},{"why":"Provides the image-to-image diffusion baseline (Palette) counted as a cDPM reference.","marker":"[16]"},{"why":"Shows how pretrained network features can be injected into a diffusion model as biases, the mechanism LegoPET adopts.","marker":"[20]"}],"fun_headline_variants":["LegoPET: multiscale priors sharpen diffusion PET","LegoPET adds hierarchical features to PET diffusion","LegoPET: hierarchical features lift PET diffusion quality","LegoPET: pretrained features improve diffusion PET","LegoPET beats vanilla cDPM with multiscale priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 0.59 dB PSNR and 0.011 SSIM gains measured on 550 simulated 2D FDG brain slices from two phantom volumes reflect a real, reproducible improvement rather than run-to-run variation, and that they will carry over to 3D and patient data.","fun_headline_variants_meta":{"raw":{"variants":["LegoPET: multiscale priors sharpen diffusion PET","LegoPET adds hierarchical features to PET diffusion","LegoPET: hierarchical features lift PET diffusion quality","LegoPET: pretrained features improve diffusion PET","LegoPET beats vanilla cDPM with multiscale priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1896,"prompt_tokens":977,"completion_tokens":919,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":838}},"tokens_in":593,"tokens_out":919,"duration_ms":9066,"temperature":1.0,"reasoning_tokens":838,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:53:19.862563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the LegoPET-versus-cDPM comparison over at least five independent training seeds and report the mean and standard deviation of PSNR/SSIM; if the 0.59 dB gap falls within one standard deviation of the two methods' distributions, the claimed advantage is not established. A second test is to evaluate both methods on held-out clinical or 3D PET volumes and check whether the visual and SSIM gains persist outside the simulated 2D phantom setup.","supporting_citations":[{"cited_title":"Design and construction of a realistic digi- tal brain phantom","cited_arxiv_id":null,"evidence_quote":"Provides the simulated brain phantom dataset from which all training, validation, and test slices are derived."},{"cited_title":"Deeppet: A deep encoder–decoder network for directly solving the pet image reconstruction in- verse problem","cited_arxiv_id":null,"evidence_quote":"Serves as the regression-based baseline and motivates the need for DL reconstruction that avoids blurry outputs."},{"cited_title":"Image-to-image translation with conditional adver- sarial networks","cited_arxiv_id":null,"evidence_quote":"Is one of the GAN-based baselines (Pix2Pix) that LegoPET compares against on PSNR/SSIM and visual quality."},{"cited_title":"The appli- cations of discrete wavelet transform in image processing: A review","cited_arxiv_id":null,"evidence_quote":"Motivates the discrete-wavelet-transform loss term used to supervise PnPNet's high-frequency features."},{"cited_title":"Multiscale structure guided diffusion for image deblurring","cited_arxiv_id":null,"evidence_quote":"Motivates the idea of guiding a diffusion model with multiscale structure from a learned prior."}],"review_version":1}