{"id":"1167a4e8-f586-43e3-8b52-89a2a5f0294f","arxiv_id":"2501.12249","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A diffusion-based method that jointly reconstructs motion-corrected sparse-view 4DCT images and estimates respiratory motion, outperforming three baselines on XCAT phantoms.","lead":"This paper trains a diffusion model to reconstruct sparse-view 4DCT images while simultaneously estimating respiratory motion, using a wavelet-based diffusion process and an adaptive forward model. The method is tested on simulated XCAT phantoms and reports higher image quality scores than three baseline methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never validates the estimated motion fields, so the reported gains may reflect the diffusion prior and multi-frame data rather than the claimed motion correction.","rationale":"Table 1 and Figure 2 provide genuine support for better image quality on XCAT phantoms, and the use of a wavelet diffusion prior is a reasonable engineering choice. My concern is not with the measured metrics themselves but with the attribution: the paper's novelty claim is the adaptive motion calibration, and that component is never directly evaluated. Eq. (1) is a strong assumption, and while the XCAT volumes are generated independently of it, the paper does not report whether the estimated (phi, s) actually matches the true deformations. Therefore a reader cannot distinguish a successful blind motion estimate from a method that effectively performs non-gated diffusion reconstruction with all frames and benefits from the larger data volume. The reader's weakest_assumption (Eq. (1) expressiveness) is a plausible source of this gap, but the more load-bearing issue is the absence of any motion-estimation accuracy metric. My proposed test directly supplies that metric. If the test shows accurate motion estimation and a meaningful PSNR difference from a no-motion control, the concern is resolved and the conditional verdict could eventually become acceptance; if not, the central claim is unsupported. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":6530,"tokens_out":10854,"duration_ms":115116,"concrete_test":"Register the reconstructed reference image x_hat to each ground-truth XCAT phase volume x_k to derive the true per-time deformation field, and compare it with the estimated W_{s_k phi} in Algorithm 1 line 7, reporting mean endpoint error across all 170 times. If the mean error exceeds roughly one voxel (2.6 mm) while Table 1 PSNR/SSIM remain high, the claimed motion correction is not validated; if the error is small, the Eq. (1) model is adequate for XCAT data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that JRM-ADM outperforms existing techniques for motion-corrected sparse-view 4DCT. The load-bearing uncertainty is that Algorithm 1 (lines 7 and 13) estimates the B-spline DVF phi and scalar surrogates s_k with only 5 RMSprop iterations per diffusion step, and the paper reports no quantitative comparison between the estimated deformations and the true XCAT phase-to-phase motion. Because the evaluation is restricted to the end-inhale reference phase, high PSNR/SSIM could be achieved through a strong diffusion prior and by using all 170 time frames, even if the estimated motion is inaccurate or the model in Eq. (1) is misspecified. The comparison against JRM-TV, which uses the same all-frame data and the same motion model, does not isolate the contribution of the adaptive diffusion calibration. Section 4 acknowledges the lack of real 4DCT validation, but the missing internal check is more basic: the method's motion estimates are never scored. Without such a check, 'motion-corrected' and the claimed 'solving' of the blind inverse problem are not established by the reported image metrics alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JRM-ADM, a diffusion-model-based framework for joint reconstruction and motion estimation (JRM) in sparse-view 4DCT. Respiratory motion is modeled by a single deformation field φ scaled by a scalar surrogate signal s_k per time frame (φ_k = s_k·φ), and the forward model is calibrated adaptively during diffusion posterior sampling. The method operates in a wavelet latent space to reduce memory and computation, and it alternates between data-consistency updates of the image and RMSprop-based updates of (φ,s). Experiments on five XCAT phantoms report PSNR and SSIM for the end-inhale phase, showing improvements over gated-FBP, gated-DPS, and JRM-TV. The paper claims artifact-free, high-resolution reconstructions under irregular breathing conditions, while acknowledging the lack of real 4DCT validation.","tokens_in":6835,"tokens_out":5868,"duration_ms":57756,"significance":"If the results are substantiated, the paper makes a useful contribution by combining diffusion priors with motion-corrected 4DCT reconstruction and by addressing the computational cost through the wavelet-domain formulation. The pseudo-code and the explicit treatment of the blind inverse problem are helpful. However, the current evidence is limited to synthetic XCAT data, and the absence of any quantitative validation of the estimated deformation fields means the central 'motion-corrected' claim is not yet established. The reported image metrics alone cannot distinguish accurate motion estimation from the effect of the diffusion prior and multi-frame data consistency. The paper is clearly written and the computational direction is relevant, but the evaluation gaps are substantial enough to require revision before the findings can be considered reliable.","major_comments":[{"comment":"The estimated deformation field φ and surrogate signal s are never compared to the ground-truth motion used to generate the five XCAT test datasets. Because Eq. (1) imposes a strong model (φ_k = s_k·φ) and Algorithm 1 (lines 7 and 13) performs only 5 RMSprop iterations for both the motion and image updates, the reported PSNR/SSIM gains cannot be attributed to accurate motion estimation. Please report quantitative motion errors (e.g., DVF endpoint error, surrogate correlation) for the test phantoms. The Section 4 limitation about XCAT training does not address this internal validation gap.","section":"Section 3.2 and Algorithm 1"},{"comment":"The evaluation is restricted to a single end-inhale phase, and the paper never states how the end-inhale image was formed from the method's output (x̂0, φ̂, ŝ). If PSNR/SSIM are computed on W_{ŝ_k φ̂}(x̂0) for the end-inhale index k, this warping step should be described explicitly, and the same pipeline should be applied to multiple phases or a time-averaged error. Per-phase metrics are necessary to support the general claim of motion-corrected 4DCT reconstruction across the respiratory cycle.","section":"Section 3.1 and Table 1"},{"comment":"The most relevant baseline, Huang et al. [2] (surrogate-optimized JRM without diffusion), is cited and used as the basis of the motion model but is not included in the comparison. Without this baseline, the abstract's claim that the method 'outperforms existing techniques' overstates the evidence. Please add this baseline or clearly justify its omission.","section":"Section 3.1 and Table 1"},{"comment":"The comparison against JRM-TV does not isolate the contribution of the adaptive forward-model calibration, since JRM-TV uses the same all-frame data and the same motion model. An ablation that fixes (φ,s) to the gated estimates or runs the diffusion prior without the per-step calibration would demonstrate the value of the adaptive component, which is the paper's central novelty.","section":"Section 3.2 and Table 1"}],"minor_comments":[{"comment":"The sentence 'performing the diffusion in a the wavelet domain' should read 'performing the diffusion in the wavelet domain.'","section":"Section 2.2.2"},{"comment":"'This results are confirmed' should be 'These results are confirmed.'","section":"Section 3.2"},{"comment":"The notation is incomplete: the Require line lists {α_t}_{t=1}^{T'} but the update in line 10 uses α_{t−δt}, and the relationship between δt and the α index is not defined.","section":"Algorithm 1"},{"comment":"Hyperparameters that materially affect the method are not reported: the data-consistency weight schedule ζ_t, the number of RMSprop iterations for lines 7 and 13, the B-spline control-point grid, and the initial surrogate signal s. Providing these values is important for reproducibility.","section":"Section 3.1"},{"comment":"The phrase 'standard sinusoidal signal to initialize s' is vague; please specify the amplitude, frequency, and how the phase is aligned with the acquired data.","section":"Section 2.2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope. The main risk is that the method's motion estimates are never verified, so the central 'motion-corrected' claim may not survive on real data even if the phantom results are reproducible. The authors should be encouraged to add the motion validation and the missing Huang et al. baseline, and to clarify the warping used in the evaluation. These are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is clean: take the joint reconstruction and motion estimation (JRM) framework from Huang et al., combine it with the adaptive blind-inversion diffusion approach of Bai et al., run the diffusion in the wavelet domain to save memory, and solve the combined problem with alternating updates. That integration is non-trivial and the empirical results on five XCAT phantoms are decent. The PSNR/SSIM gains over gated-FBP, gated-DPS, and JRM-TV are consistent across datasets, and the qualitative figure shows visibly cleaner images. The paper is honest about the biggest limitation—training and testing on XCAT phantoms only—and does not pretend to have clinical validation. For a short conference-style paper, the writing is clear and the method is reproducible in principle.\n\nThe soft spots are real and load-bearing. The most serious is that the estimated motion fields are never scored. Algorithm 1 updates the B-spline DVF and surrogate signal with only 5 RMSprop iterations per diffusion step, and the paper reports no comparison between the estimated deformations and the true XCAT phase-to-phase motion. Without that internal check, the claim that the method does motion correction is not established; the image-quality gains could come largely from the diffusion prior and from using all 170 frames. The comparison against JRM-TV, which uses the same motion model and full data, isolates the diffusion prior's contribution but not the motion estimation's accuracy. The title says 'solving' the blind inverse problem, which is stronger than what the evidence supports.\n\nThe missing baseline to Huang et al. is a mistake. Since the motion model is taken directly from that paper, readers need to see whether JRM-ADM beats the underlying JRM method with a proper convex prior under the same motion model. The TV-based JRM is a useful proxy but not the same thing. Reporting only the end-inhale phase is also selective, though the authors justify it as the phase most affected by motion. Still, showing one more phase would have strengthened the claim. The lack of code and hyperparameters makes the results hard to verify, and the linear surrogate model in Eq. (1) is an assumption that is never tested against richer motion patterns.\n\nNone of these issues are disqualifying for a workshop or a conference round. The paper is a reasonable contribution that combines existing ideas in a new way and reports positive preliminary evidence. What it needs is motion-field validation, the missing baseline, more phases, and code release. I would send it to peer review with the expectation of major revision. The reader's take is, in my view, correctly calibrated: conditional, moderate confidence. The stress-test note is right that the motion estimates are unvalidated, and that should be the main thing an editor asks for.\n\nFor a colleague working on 4DCT or diffusion-based inverse problems, this is worth a quick read. I would not cite it in my own work yet, until the motion validation appears. But it deserves a serious referee rather than a desk rejection.","headline":"A sensible integration of diffusion priors and surrogate-driven motion estimation for sparse-view 4DCT, with promising phantom results but no validation of the estimated motion fields.","tokens_in":7354,"tokens_out":1229,"would_cite":false,"duration_ms":15435,"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":"This paper claims that a diffusion-model prior, combined with on-the-fly calibration of a surrogate-scaled deformation model, solves blind motion-corrected sparse-view 4DCT reconstruction, reporting artifact-free end-inhale images on XCAT…","keywords":["4DCT reconstruction","motion correction","blind inverse problems","diffusion posterior sampling","joint reconstruction and motion estimation","respiratory motion","sparse-view computed tomography","wavelet diffusion model"],"falsifier":"Run JRM-ADM on a dataset whose ground-truth motion includes spatially varying or inspiratory/expiratory hysteresis—for example, a XCAT phantom with different regional breathing amplitudes or a real patient 4DCT with irregular breathing—and compare end-inhale PSNR and SSIM against JRM-TV and gated-DPS. If JRM-ADM no longer exceeds the comparators or shows residual diaphragm motion artifacts, the central claim that adaptive diffusion models solve blind motion-corrected sparse-view 4DCT is refuted.","tokens_in":6372,"feed_emoji":"🩻","tokens_out":6648,"duration_ms":58610,"temperature":0.7,"pith_summary":"This paper tries to establish that a blind inverse problem—reconstructing a motion-free 3-D CT image from sparse-view, low-dose 4DCT measurements in which the respiratory deformation fields and surrogate signal are unknown—can be solved by combining a diffusion-model prior with an adaptively calibrated motion model. The proposed method, JRM-ADM, alternates wavelet-domain diffusion posterior sampling with short optimization steps that fit a scalar-surrogate-scaled deformation field to the current image estimate. On five XCAT phantom datasets with irregular breathing, it reports end-inhale reconstructions with PSNR 27.05 ± 0.37 and SSIM 0.94 ± 0.01, outperforming gated filtered backprojection, gated diffusion posterior sampling, and TV-regularized joint reconstruction and motion estimation. If this holds, sparse-view 4DCT could be reconstructed without respiratory gating while preserving resolution, which is exactly the regime relevant to low-dose radiotherapy imaging.","feed_headline":"Adaptive diffusion model cleans sparse-view 4DCT motion","feed_subtitle":"Wavelet diffusion prior plus on-the-fly motion calibration beats gating and TV baselines on XCAT phantoms.","key_machinery":"The adaptive forward model is the central object: $A_{\\varphi,s} = [R \\circ T_k \\circ W_{s_k \\varphi}]_k$, a per-time composition of the fan-beam line-integral operator $R$, a slice extractor $T_k$, and a deformation operator $W_{s_k \\varphi}$ driven by a shared B-spline DVF scaled by scalar surrogates. The argument is carried by the alternating loop in Algorithm 1: a wavelet-domain diffusion network $v_\\theta$ proposes a clean coefficient image, a short RMSprop step fits $(\\varphi, s)$ to that proposal through the likelihood, and a data-consistency update in image space produces the next clean estimate before the DDIM step advances. The discrete wavelet transform $V$ is what makes 3-D diffusion tractable, and the jumpstart from a gated FBP image shortens the reverse process.","core_discovery":"JRM-ADM's central claim is that one can jointly reconstruct the image and estimate the motion by treating the forward model itself as an unknown that depends on a B-spline deformation vector field $\\varphi$ and a surrogate vector $s$, with per-time deformations $\\varphi_k = s_k \\varphi$. During DDIM sampling in the wavelet domain, the method alternates between estimating the clean image from the diffusion network, calibrating $(\\varphi, s)$ by minimizing the weighted negative log-likelihood against that estimate, and enforcing image-domain data consistency. The paper reports that this yields artifact-free, high-resolution end-inhale reconstructions under irregular breathing on XCAT phantom data, with higher PSNR and SSIM than gated-FBP, gated-DPS, and JRM-TV on all five test datasets.","pith_inferences":["The single-deformation-field scaling assumption, $\\varphi_k = s_k \\varphi$, is the part most likely to break on real patients; a natural extension would be to parameterize motion with two or more DVF components, for example separate inspiratory and expiratory fields, and let the surrogate vector choose their mixture.","The same alternating calibration trick should transfer to other blind inverse problems in medical imaging where the unknown forward operator has a low-dimensional parameterization, such as joint activity and attenuation estimation in PET.","Because the evidence is currently limited to XCAT phantoms that may conform to the motion model, the decisive test is a comparison on real 4DCT or on a phantom with spatially varying or hysteretic motion; the authors state they are working toward real CT volumes.","The jumpstart strategy and five-step RMSprop calibration suggest a robustness question: how sensitive the final image is to the surrogate initialization and the number of calibration iterations could be checked with an ablation across datasets."],"forward_implications":["The end-inhale phase can be reconstructed without phase gating, removing gating-induced noise amplification and irregular-breathing motion artifacts in the same pass.","Surrogate signals need not be measured during acquisition; the scalar surrogate vector is recovered as part of the reconstruction.","Diffusion priors give the resolution preservation that TV regularization lacks, while still suppressing streak and photon-counting noise.","The Poisson statistical weighting in the data-consistency term carries over to low-dose photon-counting CT, where counts per view are small.","Wavelet-domain diffusion keeps the computationally heavy 3-D problem within feasible memory and runtime budgets."],"supporting_citations":[{"why":"Supplies the surrogate-driven motion model and the assumption that per-time deformations are a scaled shared DVF (Eq. 1).","marker":"[1]"},{"why":"Extends the motion model to optimize the surrogate signal as hyper-parameters, which JRM-ADM adopts.","marker":"[2]"},{"why":"Provides the diffusion posterior sampling conditional score that links the diffusion prior to the measurement likelihood.","marker":"[3]"},{"why":"Wavelet diffusion model that compresses the 3-D diffusion into a low-memory latent space.","marker":"[6]"},{"why":"DDIM sampler used for the reverse-time update with $\\sigma_t = 0$ and step $\\delta_t > 1$.","marker":"[9]"},{"why":"Blind inversion approach that inspires calibrating the unknown forward model by likelihood minimization in Eq. (12).","marker":"[10]"},{"why":"Plug-and-play data-consistency update used in the image-domain step of Eq. (13).","marker":"[12]"},{"why":"XCAT phantom software used to generate training and test 4DCT datasets.","marker":"[15]"}],"fun_headline_variants":["Adaptive diffusion unblinds motion in 4DCT","Diffusion calibrates motion for cleaner 4DCT","Joint motion estimation with diffusion improves 4DCT","Wavelet diffusion beats baselines in sparse 4DCT","Blind inverse problem solved: adaptive diffusion 4DCT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes every respiratory deformation is a single reference deformation field scaled by a scalar surrogate; if true breathing has spatially varying or hysteretic motion that this one-parameter scaling cannot represent, both the reconstructed image and the estimated motion will be biased.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive diffusion unblinds motion in 4DCT","Diffusion calibrates motion for cleaner 4DCT","Joint motion estimation with diffusion improves 4DCT","Wavelet diffusion beats baselines in sparse 4DCT","Blind inverse problem solved: adaptive diffusion 4DCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1603,"prompt_tokens":892,"completion_tokens":711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":628}},"tokens_in":508,"tokens_out":711,"duration_ms":7192,"temperature":1.0,"reasoning_tokens":628,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:21:15.449283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run JRM-ADM on a dataset whose ground-truth motion includes spatially varying or inspiratory/expiratory hysteresis—for example, a XCAT phantom with different regional breathing amplitudes or a real patient 4DCT with irregular breathing—and compare end-inhale PSNR and SSIM against JRM-TV and gated-DPS. If JRM-ADM no longer exceeds the comparators or shows residual diaphragm motion artifacts, the central claim that adaptive diffusion models solve blind motion-corrected sparse-view 4DCT is refuted.","supporting_citations":[{"cited_title":"A generalized framework unifying image registration and respiratory motion models and incorporating image recon- struction, for partial image data or full images","cited_arxiv_id":null,"evidence_quote":"Supplies the surrogate-driven motion model and the assumption that per-time deformations are a scaled shared DVF (Eq. 1)."},{"cited_title":"Resolving Variable Respiratory Motion From Unsorted 4D Computed Tomography","cited_arxiv_id":null,"evidence_quote":"Extends the motion model to optimize the surrogate signal as hyper-parameters, which JRM-ADM adopts."},{"cited_title":"WDM: 3D wavelet diffusion models for high-resolution medical image synthesis","cited_arxiv_id":null,"evidence_quote":"Wavelet diffusion model that compresses the 3-D diffusion into a low-memory latent space."},{"cited_title":"Blind Inversion using Latent Diffusion Priors","cited_arxiv_id":"2407.01027","evidence_quote":"Blind inversion approach that inspires calibrating the unknown forward model by likelihood minimization in Eq. (12)."},{"cited_title":"Denoising diffusion models for plug-and-play image restoration","cited_arxiv_id":null,"evidence_quote":"Plug-and-play data-consistency update used in the image-domain step of Eq. (13)."}],"review_version":1}