{"id":"29326a23-1be6-4caa-be60-31c7f50fdc00","arxiv_id":"2507.18012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DEcomp-MoD decomposes dual-energy CT sinograms into material images by combining a supervised U-Net that learns the material decomposition function with a plug-and-play denoising diffusion prior, and reports improved quantitative results on synthetic AAPM data.","lead":"This paper introduces DEcomp-MoD, a deep learning method that turns dual-energy CT projection data into water and bone material images by combining a learned projection-domain decomposition network with a diffusion image prior. It reports higher PSNR and SSIM than several baselines on synthetic low-dose, sparse-view scans from the public AAPM dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inference data-consistency term is defined against Pθ(y), the learned U-Net substitute for h^{-1}(y), rather than against the measured DECT sinogram; since Pθ is trained and validated only on Spekpy-simulated data, the claimed 'guarantees consistency' is unverified outside the training…","rationale":"The paper has genuine strengths: it cleanly decouples projection-domain decomposition from an image-domain diffusion prior, the synthetic training protocol is reproducible in principle, and within the exact simulation distribution the reported gains over DIRECT-Net and MCG are substantial. The decisive question is whether this supports the stronger claims in the Abstract and Section I-A: direct conversion of DECT projections into material images and guaranteed consistency. The reader's weakest assumption identifies the same point I would stress: the only link from raw measurements to the data-consistency term is Pθ, a supervised network trained on simulated (y, p*) pairs. I agree with that read; the diffusion prior cannot compensate for a biased Pθ because every measurement-informed update in Algorithm 1 is computed from Pθ(y) (line 5), not from y directly. The missing validation is concrete: no real DECT sinograms, no measured spectra, no calibration phantom, and no robustness test under a different forward model. This is an omission in the evidence for the central claim, not a disagreement with community consensus or a circularity charge. I also flag Eq. (31) as a formal defect in the published solver, but I do not make it the headline because it is more plausibly a fixable implementation or notation error, whereas the simulator-to-scanner transfer gap is structural. Consequently the reader's conditional verdict should stand: accept the method only after the data-consistency path is validated on a held-out forward model or real scanner data, and after the solver formula is corrected or code is released.","tokens_in":16504,"tokens_out":7355,"duration_ms":75440,"concrete_test":"Generate a held-out test set from the same AAPM ground-truth slices using a perturbed forward model: 100/140 kVp spectra with 2.0 mm Al and 0.3 mm Cu filtration instead of 90/150 kVp with 1.5/1.2 mm, while keeping the trained Pθ and diffusion prior fixed. If DEcomp-MoD's water/bone PSNR on this shifted-spectrum set drops by more than a pre-specified tolerance (e.g., >3 dB) relative to Table I, or if a calibrated polynomial h^{-1} baseline trained on the same geometry becomes competitive, the consistency claim is simulation-bound. If a real DECT phantom (e.g., Gammex 467) is available, repeat the comparison with measured spectra and calibration data on the same algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eqs. (26)–(28). Eq. (26) replaces the true inverse material decomposition h^{-1} by the learned network Pθ, and Eq. (28) then defines the inference data-consistency term as ||Pθ(y) - Ax||^2_B instead of the spectral forward-model NLL in Eq. (14). Thus DEcomp-MoD can only enforce consistency with the output of Pθ; it never compares the final material images against the measured projections y through Beer's law. Pθ is trained on synthetic (y, p*) pairs generated with Spekpy spectra at 90/150 kVp, NIST attenuation coefficients, and threshold-based segmentation (Section VI-A), and no real DECT sinograms or calibrated h^{-1} baseline are used anywhere in the paper. Any mismatch in kVp, filtration, detector response, beam hardening, or segmentation protocol becomes a systematic bias in Pθ, and the diffusion prior cannot remove it because the diffusion sub-problem operates on material images and has no independent access to y. The claimed 'guarantees consistency of the results' therefore holds only if the simulation used to train Pθ matches the deployment scanner, which is not demonstrated. I additionally note that Eq. (31) is dimensionally inconsistent as written: with A ∈ R^{N×M} and B ∈ R^{N×2}, the term A^T(B⊙A) does not form the correct two-material normal equations for (30); this is a separate reproducibility defect, but the transfer gap concerning Pθ is the more fundamental barrier to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DEcomp-MoD is a two-module deep-learning pipeline for dual-energy CT material decomposition. A U-Net Pθ is trained in a supervised way on simulated dual-energy sinograms to approximate the inverse material decomposition h^{-1}, mapping the two energy sinograms into material sinograms. The second module is a pre-trained DDPM/DDIM score-based prior on material images. At inference, HQS alternates between a diffusion denoising step and a quadratic data-consistency step in which the material sinogram produced by Pθ(y) is fitted to Ax under a per-ray covariance B_n, solved with CG. Experiments on synthetic low-dose, sparse-view DECT sinograms simulated from the AAPM low-dose CT dataset compare water/bone PSNR and SSIM against FBP, MCG diffusion, and DIRECT-Net, and include an angular-undersampling generalization study and a parameter study of λ and ξ.","tokens_in":16852,"tokens_out":8740,"duration_ms":91594,"significance":"If fully supported, the method would have practical value: decoupled training avoids end-to-end unrolling, the diffusion prior can be reused across acquisition settings, and the projection-domain decomposition makes the data-consistency subproblem quadratic and relatively cheap (about 10 s per image). The paper gives a reasonably complete description of the simulation pipeline, the network architectures, and Algorithm 1, which helps reproducibility. However, the central 'guarantees consistency' claim is not supported by Eq. (28), the comparison claims are weakened by test-set tuning and by the single-slice MCG baseline, and the closed-form update in Eq. (31) is not reproducible as written.","major_comments":[{"comment":"The data-consistency term fθ(y,x) in Eq. (28) is ||Pθ(y) − Ax||^2_B, not a comparison of x with the measured sinogram y; it only enforces consistency with the output of the learned U-Net. Since Pθ is trained and tested on Spekpy-simulated pairs from a single simulation pipeline (Section VI-A2) and is never validated against a calibrated h^{-1} or any real DECT data, the Abstract and Section I-A claim that DEcomp-MoD 'guarantees consistency of the results' is an overclaim. I recommend either augmenting the inference objective with the spectral forward-model term y ≈ h(Ax), or rephrasing the claim as consistency with the learned projection-domain decomposition and adding a domain-shift experiment (e.g., testing on spectra or filtration different from training).","section":"V-A, Eq. (28)"},{"comment":"The default values λ = 0.001 and ξ = 1.0 are selected 'based on the results in section VI-F1', which are obtained by examining qualitative and quantitative testing performance on the same test set used for Table I and Fig. 8. This is effectively tuning on the test set and makes the reported PSNR/SSIM and the comparisons with FBP, MCG, and DIRECT-Net optimistic. Please introduce a separate validation set, or report both validation and test numbers, before computing the final comparisons.","section":"VI-F1"},{"comment":"Equation (31) is dimensionally inconsistent: B ∈ R^{N×2} is defined as the concatenation of the diagonal entries of B_n, while A ∈ R^{N×M} with M = 256^2 in the experiments, so the elementwise product B ⊙ A is not defined. The closed-form solution of the weighted least-squares problem (30) requires a block-diagonal weight matrix acting on vec(Ax) (e.g., a matrix with blocks B_n along the ray index), leading to normal equations of the form (Ãᵀ D Ã + μI)x = Ãᵀ D p + μz. Please give the correct matrix form and state what was actually implemented; as written, Algorithm 1 is not reproducible.","section":"V-A, Eq. (31)"},{"comment":"The covariance formula in Eq. (29) appears to have the Jacobian orientation reversed. With W_n the inverse covariance of y_n and Pθ(y)_n approximating h^{-1}(y)_n, the first-order covariance of the decomposed sinogram should scale as ∇Pθ(y)_n W_n^{-1} ∇Pθ(y)_nᵀ (or the equivalent forward-model form), whereas [∇Pθ(y)_n]^{-1} W_n [∇Pθ(y)_n]^{-T} would make the weight increase with photon count, which is unphysical. Please correct the formula or provide the derivation from [17] showing the intended convention.","section":"V-A, Eq. (29)"},{"comment":"The MCG diffusion baseline is evaluated on a single slice (as acknowledged in the text below Table I), and its PSNR/SSIM entries have no standard deviation. The claim that DEcomp-MoD outperforms 'state-of-the-art unsupervised score-based model' is therefore not established over the test population. Please run MCG on the full test set or, if the cost is prohibitive, clearly state that the comparison is illustrative on one slice and remove the global comparative claim.","section":"VI-D, Table I"}],"minor_comments":[{"comment":"The description of Pθ is inconsistent: Eq. (26) and Algorithm 1 treat Pθ as mapping R^{N×2} to R^{N×2}, but Section VI-B1 says the U-Net output is a one-channel material sinogram, and Section VII-3 says only one material can be generated at a time. Please clarify whether two networks are trained or whether a two-channel output is used.","section":"VI-B1 / VII-3"},{"comment":"In Section VI-E, 'the CT data is acquired at difference conditions' contains a typo ('difference' should be 'different'), and the sentence 'with fixed photon counts 2·10^6 and different number of sampling angles' is missing a verb; the paragraph should be edited for clarity.","section":"VI-E"},{"comment":"There is a notation inconsistency after Eq. (31), where G is defined as Aᵀ(B ⊙ A) + λI but the preceding equation uses μ; the text should use one symbol consistently (μ or λ).","section":"V-A after Eq. (31)"},{"comment":"The default CG iterations are fixed to 10, but the CG tolerance or stopping criterion is not stated; since the update (35b) is solved approximately at every diffusion step, please report the stopping rule or tolerance used.","section":"VI-C"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a public code release because the dimension and covariance issues in Section V-A cannot be resolved by reading alone. I would also ask the authors to revise the abstract, because the phrase 'guarantees consistency' is stronger than what Eq. (28) actually enforces. Overall this is a solid simulation study with a useful decoupled training idea, but the numerical claims and the algorithmic specification need corrections before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2507.18012. The genuinely new bit is the specific combination: a supervised U-Net that maps dual-energy sinograms to material sinograms, inserted into a Half-Quadratic Splitting loop with a plug-and-play DDPM prior on material images. That integration is not in the cited literature, and the paper is honest that each component is known. The synthetic results are competitive, and the generalization sweep across sampling angles is a useful empirical check.\n\nThe paper also deserves credit for a mostly clear train/inference description and a simulation pipeline that is reproducible in principle: Spekpy spectra, Tomosipo projector, NIST attenuation coefficients. The comparisons against FBP and DIRECT-Net are reasonable, and the reported gains are large enough to take seriously.\n\nNow the soft spots, in rough proportion. The load-bearing issue is the data consistency term. Eq. (26) replaces the true inverse h^{-1} with the learned Pθ, and Eq. (28) defines the NLL against Pθ(y) rather than the measured sinogram. So the inference optimization only enforces agreement with the network's own decomposition; it never checks the final material images against y through Beer's law. Pθ is trained and tested on the same Spekpy simulation pipeline, so the abstract's claim that the method \"guarantees consistency of the results\" is only as good as the simulation-to-scanner transfer. No real DECT data or calibrated phantom appears anywhere. That is not fatal for a simulation study, but the wording overreaches.\n\nSecond, Eq. (31) is dimensionally wrong as written. B is N×2, A is N×M, so B⊙A is undefined except in the degenerate case M=2, and A^T(B⊙A) cannot form the correct M×M normal equations. The CG solve in Eq. (32) inherits the problem. This is a visible, fixable defect, but it needs correcting before anyone can reproduce the algorithm from the paper.\n\nMinor issues: λ and ξ are selected based on testing performance in Section VI-F1, which is tuning on the test set. The MCG baseline is evaluated on a single slice only. The paper cites other diffusion-based DECT decomposition works ([31], [32]) but does not compare against them, which weakens the \"state-of-the-art\" claim.\n\nOverall, this is a plausible integration with solid synthetic evidence, but the central claim is overbroad and there is a concrete derivational error. A serious referee should see it; the core architecture can survive major revision. I would bring it to reading group only after the authors fix the math and either soften or properly support the consistency claim.\n\nRecommendation: send to peer review, expect major revision.","headline":"Plausible integration of a learned sinogram-domain decomposition with a diffusion prior, but the consistency claim is overreaching and Eq. (31) has a real dimensional error.","tokens_in":17364,"tokens_out":1781,"would_cite":false,"duration_ms":20052,"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":"A learned inverse material model turns dual-energy CT projections directly into water and bone density maps.","keywords":["dual-energy CT","material decomposition","denoising diffusion model","projection-domain decomposition","sparse-view CT","model-based iterative reconstruction","score-based prior","quantitative CT"],"falsifier":"Take a calibrated dual-energy CT system with known water and bone phantom inserts, acquire real low-dose and sparse-view sinograms, run the pre-trained DEcomp-MoD without any fine-tuning, and compare the recovered material densities against the known insert values; a systematic bias that grows with tube voltage, filtration, or added phantom material would show the learned $h^{-1}$ was tied to the simulation rather than to the physical scanner.","tokens_in":16279,"feed_emoji":"🩻","tokens_out":8115,"duration_ms":81486,"temperature":0.7,"pith_summary":"The paper sets out to show that dual-energy CT material decomposition can skip image-domain post-processing entirely: a learned network maps the low- and high-energy sinograms directly into water and bone density images, with a denoising diffusion model supplying the image prior and a data-consistency term keeping the result tied to the measured projections. The proposed DEcomp-MoD pipeline is trained and evaluated on synthetic low-dose, sparse-view DECT sinograms generated from real patient CT images, and the reported numbers put it ahead of both a supervised one-step network and a score-based diffusion baseline. If the claim holds, the practical payoff is a fast, calibration-free route from raw DECT projections to quantitative material maps that handles beam hardening and sparse-view artifacts without explicit inversion of the polyenergetic model.","feed_headline":"Dual-energy CT sinograms become water and bone maps in one pass","feed_subtitle":"Learned inverse material model plus diffusion prior beats supervised and unsupervised baselines at low dose and sparse views.","key_machinery":"The load-bearing object is the learned decomposition map $P_\\theta \\approx h^{-1}$ embedded directly inside the negative log-likelihood term (Eq. 28). This substitution is what linearizes the otherwise nonlinear DECT model: instead of inverting the polyenergetic Beer-Lambert relation numerically or by phantom calibration, the U-Net supplies material sinograms $P_\\theta(y)$, and the data-consistency update becomes a weighted quadratic minimization solved with conjugate gradient. Around it, the diffusion prior is applied through a proximal denoising step matching the DDIM reverse update, with parameters $\\lambda$ and $\\xi$ trading prior strength against measurement fidelity. The HQS split makes the two learning modules independent: the U-Net works in sinogram space, the diffusion model in image space, and the CG solve couples them.","core_discovery":"DEcomp-MoD's central claim is that the hard nonlinear inverse of the polyenergetic DECT measurement model, $h^{-1}$, can be replaced by a supervised U-Net $P_\\theta$ trained on simulated (energy sinogram, material sinogram) pairs, turning the data-consistency sub-problem into a quadratic least-squares fit. Around that fit, an unconditional denoising diffusion model acts as a plug-and-play prior on material images, and the two are coupled through half-quadratic splitting with conjugate-gradient updates. The algorithm iterates a DDIM-style reverse diffusion step followed by a projection-domain data-consistency correction, which the paper argues guarantees consistency of the results in the sense that every update pulls the material image back toward the measured sinograms. Quantitative results on simulated low-dose patient data give water PSNR 29.14 dB and bone PSNR 33.88 dB, above the compared baselines.","pith_inferences":["Not stated in the paper: because the diffusion prior is trained on material images rather than sinograms, the same prior could be reattached to different scanner geometries or spectra by retraining only the smaller U-Net; the authors note a future unsupervised direction, but the modular split makes this concrete.","Not stated in the paper: the covariance matrix $B_n$ computed by backpropagating through $P_\\theta$ could be interpreted as an uncertainty map for the decomposition, offering a per-pixel confidence estimate that the current evaluation does not report.","Not stated in the paper: replacing the supervised U-Net with a self-supervised estimator trained on paired noisy sub-sampled sinograms would remove the dependence on simulated ground-truth material sinograms and directly test whether the consistency guarantee transfers to real scanner data.","Not stated in the paper: if $P_\\theta$ were swapped for a differentiable calibration-based decomposition, the same diffusion/HQS loop would still run, making the framework a testbed for whether learned decompositions beat classical polynomial fits independently of the diffusion prior."],"forward_implications":["Material decomposition no longer needs a separate reconstruction step: the energy sinograms are mapped to material sinograms and then to images through a closed-form quadratic solve, so beam-hardening and noise correlations from the polyenergetic model are handled in one pass.","The learned $P_\\theta$ removes the need for explicit calibration of $h^{-1}$, since the U-Net is trained from simulated spectral data instead of a measured calibration phantom.","On the simulated low-dose, sparse-view test set, the method reports water PSNR 29.14 dB and bone PSNR 33.88 dB, outperforming both the supervised one-step network and the score-based diffusion baseline.","The diffusion prior generalizes across sampling angles: trained at 180 angles, DEcomp-MoD keeps higher PSNR and SSIM than the baselines from 60 to 360 angles without retraining.","With 100 sampling steps a 256x256 material image is produced in about 9.4 seconds, which is competitive with unrolled iterative methods and far faster than the manifold-constraint diffusion baseline."],"supporting_citations":[{"why":"Supplies the plug-and-play denoising diffusion formulation, including the parameter $\\xi$ that trades predicted noise against random noise in the reverse sampling.","marker":"[34]"},{"why":"Defines the DDPM forward/reverse process and the noise-prediction loss used to train the material-image diffusion prior.","marker":"[27]"},{"why":"Provides the DDIM accelerated reverse-step update that the DEcomp-MoD sampling loop is built on.","marker":"[37]"},{"why":"Supplies the joint quadratic likelihood and Taylor-expanded covariance model used to turn the data-consistency term into a weighted least-squares problem.","marker":"[17]"},{"why":"Is the supervised one-step mutual-domain network DIRECT-Net that serves as the primary supervised baseline.","marker":"[23]"},{"why":"Is the manifold-constraint diffusion baseline (MCG) for inverse problems that DEcomp-MoD is compared against.","marker":"[46]"},{"why":"Provides the real patient CT slices used to synthesize the low-dose, sparse-view DECT sinograms and material labels for training and testing.","marker":"[35]"},{"why":"Supplies the modeled 90 kVp and 150 kVp X-ray spectra used in the forward simulation.","marker":"[42]"},{"why":"Supplies the energy-dependent mass attenuation coefficients for the water and bone basis materials in the spectral forward model.","marker":"[43]"}],"fun_headline_variants":["Direct DECT decomposition: sinogram to water and bone in one model","Diffusion prior turns dual-energy CT sinograms into material maps directly","One-shot sinogram-to-material decomposition with model-based diffusion","Diffusion with data consistency: DECT projections become water and bone"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on the simulated U-Net $P_\\theta$ being a faithful substitute for the true inverse material decomposition on real scanner data; the experiments only test $P_\\theta$ on synthetic sinograms generated with the same spectra, segmentation thresholds, and attenuation tables used in training, so a real scanner with different beam quality, scatter, or detector response could break the consistency guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Direct DECT decomposition: sinogram to water and bone in one model","Diffusion prior turns dual-energy CT sinograms into material maps directly","One-shot sinogram-to-material decomposition with model-based diffusion","Diffusion with data consistency: DECT projections become water and bone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000397,"raw_usage":{"total_tokens":2090,"prompt_tokens":969,"completion_tokens":1121,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1047}},"tokens_in":585,"tokens_out":1121,"duration_ms":11108,"temperature":1.0,"reasoning_tokens":1047,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:39:01.123921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a calibrated dual-energy CT system with known water and bone phantom inserts, acquire real low-dose and sparse-view sinograms, run the pre-trained DEcomp-MoD without any fine-tuning, and compare the recovered material densities against the known insert values; a systematic bias that grows with tube voltage, filtration, or added phantom material would show the learned $h^{-1}$ was tied to the simulation rather than to the physical scanner.","supporting_citations":[{"cited_title":"Denoising diffusion models for plug-and-play image restoration,","cited_arxiv_id":null,"evidence_quote":"Supplies the plug-and-play denoising diffusion formulation, including the parameter $\\xi$ that trades predicted noise against random noise in the reverse sampling."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Defines the DDPM forward/reverse process and the noise-prediction loss used to train the material-image diffusion prior."},{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Provides the DDIM accelerated reverse-step update that the DEcomp-MoD sampling loop is built on."},{"cited_title":"Model-based iterative reconstruction for dual-energy X-ray CT using a joint quadratic likelihood model,","cited_arxiv_id":null,"evidence_quote":"Supplies the joint quadratic likelihood and Taylor-expanded covariance model used to turn the data-consistency term into a weighted least-squares problem."},{"cited_title":"DIRECT-Net: A unified mutual- domain material decomposition network for quantitative dual-energy CT imaging,","cited_arxiv_id":null,"evidence_quote":"Is the supervised one-step mutual-domain network DIRECT-Net that serves as the primary supervised baseline."},{"cited_title":"Improving diffusion models for inverse problems using manifold constraints,","cited_arxiv_id":null,"evidence_quote":"Is the manifold-constraint diffusion baseline (MCG) for inverse problems that DEcomp-MoD is compared against."},{"cited_title":"Low-dose CT for the detection and classification of metastatic liver lesions: results of the 2016 low dose CT grand challenge,","cited_arxiv_id":null,"evidence_quote":"Provides the real patient CT slices used to synthesize the low-dose, sparse-view DECT sinograms and material labels for training and testing."},{"cited_title":"Spekpy v2. 0—a software toolkit for modeling X-ray tube spectra,","cited_arxiv_id":null,"evidence_quote":"Supplies the modeled 90 kVp and 150 kVp X-ray spectra used in the forward simulation."},{"cited_title":"Tables of X-ray mass attenuation coefficients and mass energy-absorption coefficients 1 kev to 20 mev for elements Z = 1 to 92 and 48 additional substances of dosimetric interest,","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-dependent mass attenuation coefficients for the water and bone basis materials in the spectral forward model."}],"review_version":1}