{"id":"a5a17233-8ce7-437d-84de-52cb35b3b5e4","arxiv_id":"2501.05085","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A dual-domain CNN that estimates projection noise inside the measured region and extrapolates outside it, followed by image-domain refinement, improves low-dose interior CT reconstruction over image-domain CNNs in simulated AAPM data.","lead":"The paper proposes a two-stage deep learning method for X-ray CT that fixes both low-dose noise and truncated-field artifacts at once by first cleaning and extrapolating the raw projection data, then refining the image. The method outperforms image-only neural networks in simulated tests, which matters because combined low-dose and region-of-interest scans reduce radiation but current networks struggle with the mixed artifacts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rank argument depends on an unverified spectral-separation premise; Eq. 9 is applied beyond its known validity, so the theoretical reason image-domain CNNs fail is not established.","rationale":"The reader's weakest_assumption identified the same load-bearing concern: the spectral separation of cupping (low-frequency) and noise (high-frequency) is assumed but not established, and the Hankel-rank inequality in Eq. 23 depends on it. My pass confirms this is the most load-bearing issue because the paper's significance statement and Section 4.1's theoretical argument are the primary justification for why a projection-domain network is needed. The concern is concrete: Eq. 9 is an exact identity only under restrictive conditions that real CT image noise likely violates, and the paper provides no spectral analysis of actual artifacts. The empirical comparisons in Tables 2–5 could still support the proposed method as an engineering contribution, but without the rank theory the central claim that image-domain CNNs cannot efficiently solve coupled artifacts is unproven. The end-to-end training contradiction and reproducibility issues are secondary; they affect presentation and confidence but not the core theoretical premise. Therefore the appropriate verdict remains CONDITIONAL, matching the reader's assessment, with the condition being a direct verification of the spectral and rank assumptions. No change to the reader's verdict is needed.","tokens_in":18661,"tokens_out":4691,"duration_ms":51095,"concrete_test":"Generate simulated truncated low-dose CT data with the paper's forward model (fan-beam, 720 views, truncation ratios 58–79%, I0 values from Table 2). For a set of test slices, compute the cupping artifact c_I (difference between truncated FBP and ground truth at I0=∞) and the image noise n_I (difference between low-dose FBP and truncated FBP), and the coupled artifact k = c_I + n_I. Compute the Hankel matrix H_d for each with the filter length d used in the U-Net backbone (e.g., 3×3 kernels, so d = 9 along each dimension, or a practical equivalent). Compare RANKH_d(n_I), RANKH_d(c_I), and RANKH_d(k) against the signal length n and the bound rmax. If RANKH_d(n_I) is already close to d, then Eq. 23's premise is false and the theoretical motivation is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is that a coupled cupping-plus-noise artifact k = c_I + n_I has Hankel-matrix rank near the signal length, exceeding the filter-length bound d, so image-domain CNNs cannot remove it (Sec. 4.1, Eq. 23). This rests on two premises: (i) the identity RANKH_d(f) = COUNT(F(f)≠0) in Eq. 9, and (ii) the spectral separation shown in Fig. 4, where image noise is claimed to have dominant high-frequency support and cupping dominant low-frequency support, so their sum has near-full Fourier support. Both premises are questionable for real CT data. For finite-length signals, Eq. 9 is not generally exact; it holds only for signals that are sums of a few complex exponentials with sufficiently large d. Low-dose CT image noise is roughly white and broadband, not high-frequency-bandlimited, so RANKH_d(n) would already be high, undermining the claimed inequality RANKH_d(n) ≤ rmax < RANKH_d(k) ≤ d. Cupping artifacts also contain boundary singularities that spread energy across frequencies. If the spectral-separation premise fails, Eq. 23 does not follow, and the paper's explanation for why image-domain CNNs underperform collapses—leaving only an empirical comparison without the stated theoretical basis. The paper's own Fig. 9 measures singular values of trained feature maps, not of the input artifact, so it does not validate the input-side rank claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript addresses combined interior (region-of-interest) and low-dose X-ray CT reconstruction. The authors argue, via the theory of deep convolutional framelets, that an image-domain CNN cannot efficiently remove the coupled cupping-plus-noise artifact because the combined artifact has a near-full Fourier support and hence a Hankel-matrix rank close to the signal length, exceeding the filter-length bound. They therefore propose a dual-domain network: a projection-domain CNN with two bridge modules, one estimating projection noise inside the measured (truncated) region and one extrapolating projections outside it, followed by FBP and an image-domain CNN. End-to-end training with an FBP layer is described. Experiments on simulated truncated low-dose projections from 10 AAPM Low-Dose CT Grand Challenge subjects (9 training, 1 test) compare the proposed Dual-Net against U-Net, W-Net, TV, nonlocal prior, and extrapolation methods, reporting consistent NMSE/SSIM/PSNR improvements across six truncation ratios and six photon-count levels.","tokens_in":18949,"tokens_out":4316,"duration_ms":43082,"significance":"If the empirical claims hold, the paper would make a useful contribution to combining interior and low-dose CT reconstruction: the dual-domain formulation is a natural and technically sound way to separate a measurement-domain noise problem from a missing-data (extrapolation) problem, and the consistent gains over U-Net and W-Net across all tested configurations are encouraging. The inclusion of an FBP layer in an end-to-end trainable pipeline, the controlled comparison of one-time and two-time unrolled architectures, and the comparison with classical methods (TV, nonlocal prior, extrapolation) are strengths. However, the significance is currently limited by the narrow experimental basis: all quantitative conclusions rest on a single test subject, with no error bars or statistical testing, and the central theoretical explanation, while interesting, is not rigorously established for the actual data.","major_comments":[{"comment":"The rank inequality RANKH_d(k) <= r*_max ≈ d is derived from Eq. (9), RANKH_d(f) = COUNT(F(f)≠0), but this identity is not exact for finite-length arbitrary signals; it holds only for signals composed of a few complex exponentials when the matrix pencil parameter d is sufficiently large. Low-dose CT image noise is roughly broadband, not a sum of a few exponentials, so its Hankel-matrix rank is not necessarily bounded by a small r_max, and the cupping artifact contains boundary singularities at the ROI edge that spread energy across the Fourier domain. The spectral separation in Fig. 4 is only qualitative, with no quantitative bandlimit measurement. Consequently, Eq. (23) does not follow, and the claimed theoretical reason that image-domain CNNs fail on the coupled artifact is not established. This is the load-bearing theoretical claim of the paper, so it needs either a rigorous demonstration on the actual artifact distributions or a substantial softening of the claim.","section":"Sec. 4.1, Eq. (23)"},{"comment":"The experimental check of low-rankness is performed on the Hankel singular values of the last feature maps of the trained backbone networks, not on the input artifacts (the coupled image-domain artifact k or the projection-domain data p_T and its components). Thus Fig. 9 does not validate the input-side rank claim of Eq. (23); it only shows that the projection-domain CNN learns lower-rank feature representations. The paper should either measure the rank of the actual input artifact (e.g., via the Hankel matrix of q_I in the image domain and of p_T in the projection domain) or explicitly state that Fig. 9 is a feature-space observation, not a test of the theoretical premise.","section":"Sec. 6.2, Fig. 9"},{"comment":"All quantitative results are computed on a single test subject (486 slices) from the AAPM challenge, with point estimates only and no standard deviations, confidence intervals, or statistical significance tests. The improvement of the proposed method over W-Net is consistent in the tables, but a journal-level claim of superiority requires at least multiple test subjects or a per-slice variability analysis. The paper should also release code and, where permissible, the test dataset or a reproducible simulation pipeline, because the current evaluation is not independently verifiable.","section":"Sec. 3.1 and Tables 2-5"},{"comment":"The paper claims that 'the capacity of the neural network Q is fixed, but it is possible to apply low-rank constraints to individual regions while satisfying the theory of the deep convolutional framelets.' However, no low-rank constraint is actually enforced in the objective functions of Table 1; the two bridge heads share a single backbone and are jointly optimized with a weighted sum of two losses. The low-rank behavior is only observed post hoc in Fig. 9. The link between the loss decomposition and the low-rank condition is therefore asserted rather than demonstrated. An ablation that trains with explicit rank regularization or that compares separate backbones for the two heads would help support the claim.","section":"Sec. 4.2 and Table 1(e)"}],"minor_comments":[{"comment":"The affiliation contains a typo: 'Bostan' should be 'Boston'.","section":"Sec. 2.1, affiliation footer"},{"comment":"The PSNR formula uses N M ||f^*||_∞ in the numerator, which is not the standard PSNR definition (which uses the maximum pixel value, not the number of pixels times the maximum). This appears to be a typo and should be corrected, since it affects the reported PSNR values in Tables 4 and 5.","section":"Eq. (17)"},{"comment":"The training is described as 'end-to-end,' but the gradient of the second network is blocked from propagating to the first network. This is a specific design choice; clarify that the first network is trained on its own loss plus an image-domain loss term, but not on the final image loss, and discuss the implication for the term 'end-to-end.'","section":"Sec. 3.3"},{"comment":"The statement that 'the U-Net is set up to twice the size of the channels of other networks' is clear from Fig. 5 but could be stated more precisely: the U-Net's first layer has 64 channels while the others have 32, giving approximately twice the parameters.","section":"Sec. 3.2"},{"comment":"The authors note that 'the NMSE and SSIM valuse do not seem to reflect the well-corrected image quality' and then recompute metrics in body regions. This admission suggests that global image metrics may be misleading for texture-preservation; consider including the body-region metrics in the main tables or justifying the choice of global metrics.","section":"Sec. 6.3"},{"comment":"The Fourier spectra of image noise and cupping artifact are shown without a quantitative scale or a quantitative measure of 'dominant low/high frequency support.' A simple percentage-of-energy threshold would make the spectral-separation premise testable.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a medical imaging or computational imaging journal, and the empirical trend is promising. The main concern is that the theoretical narrative, which is prominently used to motivate the architecture, is not rigorously supported by the presented evidence. The single-subject evaluation and lack of released code/data are additional concerns that should be addressed before publication. The self-citation to [23] is legitimate because it is a published framework; no misconduct is indicated. The writing needs careful proofreading."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a useful empirical paper with a weak theoretical wrapper. The dual-domain architecture — a projection-domain CNN with two heads, one estimating noise inside the truncation mask and one extrapolating outside, followed by FBP and an image-domain refinement — is a specific and sensible combination. The experiments are systematic across photon counts and truncation ratios, and the proposed method wins consistently over U-Net and W-Net. If you work on low-dose interior CT, this is worth a look.\n\nThe theoretical story, however, does not hold up. The central claim is that an image-domain CNN cannot remove the coupled cupping-plus-noise artifact because the combined artifact's Hankel-matrix rank is close to the signal length, exceeding the filter-length bound from deep convolutional framelets. That argument depends on Eq. 9, which equates Hankel rank with the number of non-zero Fourier components. That identity is not exact for arbitrary finite-length signals; it applies cleanly only to sums of few exponentials or when the matrix pencil parameter is large enough. Low-dose CT noise is broadband — its Fourier support is essentially full already — and cupping artifacts have boundary singularities that spread energy across frequencies. So the premise that each individual artifact has low rank while their sum has near-full rank is not established for real CT data. The paper's own Fig. 9 measures singular values of learned feature maps, not the input artifact, so it does not validate the input-side rank claim. The theoretical reason for why image-domain CNNs fail is speculative, not proven. The empirical comparison is still informative, but the theory section overstates what it shows.\n\nThere is a wording problem too: the paper calls the training 'end-to-end' but explicitly blocks gradients from the second network to the first. That is not end-to-end learning; it is a two-stage unrolled scheme with gradient blocking. The method may still be fine, but the description is misleading.\n\nThe evaluation is thin in specific ways: one test subject, point estimates without error bars, and no code or data released. The gains are consistent across the tested grid, which is reassuring, but the evidence does not support a claim of general superiority.\n\nNet: the architecture deserves a serious referee, and I would send it out. But the theory needs to be either fixed or clearly labeled as a heuristic, and the authors should release code and data, or at least report multiple test subjects and variance. I recommend engaging with it — it is a plausible step for the subfield, but do not let the framelets story carry more weight than it can bear.","headline":"Useful dual-domain architecture with consistent empirical gains, but the theoretical rank argument is not established and the evaluation is thin; worth sending to review with clear requests for revision.","tokens_in":19451,"tokens_out":2883,"would_cite":true,"duration_ms":28798,"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":"The paper claims that image-domain CNNs cannot efficiently remove coupled cupping-plus-noise artifacts in low-dose interior CT, and that a dual-domain network which decouples the problem in projection space restores the low-rank condition…","keywords":["interior tomography","low-dose CT","deep convolutional framelets","dual-domain CNN","projection extrapolation","cupping artifact","truncated projection","end-to-end learning"],"falsifier":"Compute the Fourier support or Hankel-structured rank of the artifact directly from a real scanner's truncated low-dose projections, and train an image-domain CNN whose convolutional filter length matches the signal length; if the image-domain network then matches the dual-domain network's quality, the filter-length bound is not the limiting factor and the proposed explanation would be wrong even if the method still works.","tokens_in":18455,"feed_emoji":"🩻","tokens_out":8530,"duration_ms":69973,"temperature":0.7,"pith_summary":"This paper argues that an image-domain convolutional network cannot cleanly remove the two artifacts that appear together in low-dose interior X-ray CT: the cupping artifact from truncated projections and the image noise from few photons. The reason comes from deep convolutional framelets: the coupled artifact occupies nearly the whole Fourier spectrum, so the rank of its Hankel-structured matrix approaches the signal length and exceeds the convolutional filter length that a fixed network can represent. The proposed solution decouples the problem in the projection domain before reconstruction, using one branch to estimate the projection noise inside the measured detector region and another to extrapolate the projections outside it, followed by filtered backprojection and a second image-domain network. Trained end-to-end with the backprojection as a differentiable layer, this dual-domain network reports the lowest normalized mean-square error and highest structural similarity across truncation ratios from 0 to 79 percent and photon counts from $10^{5}$ to $\\infty$, and it also beats image-domain networks on the isolated interior and low-dose problems. If correct, the result shifts the explanation of failure from network capacity to the choice of training domain.","feed_headline":"Projection-first network beats image-only CT artifact removal","feed_subtitle":"Denoising inside the mask and extrapolating outside it preserves structures that image-only CNNs blur.","key_machinery":"The central object is the Hankel structured matrix $H_d(f)$ and the rank condition from deep convolutional framelets, together with the decoupling mask pair $T$ and $1-T$ in the projection domain. The framelets identity $H_d(f) = \\tilde{\\Phi}\\Phi^T H_d(f)\\Psi\\tilde{\\Psi}^T$ and the fact that the Hankel rank equals the number of nonzero Fourier components of the image imply that a CNN with filter length $d$ can represent only artifacts whose rank stays below a bound; the coupled cupping-plus-noise artifact violates this because its Fourier support fills the spectrum. The proposed network restores the low-rank condition by splitting the projection data into two regions: a noise-estimation bridge inside $T$ and an extrapolation bridge outside $(1-T)$, so each subproblem has a narrow Fourier support and satisfies the framelets condition.","core_discovery":"The paper's central claim is that image-domain CNNs fail at coupled low-dose interior CT not because of capacity but because of low-rankness: for the coupled artifact $k = c_I + n_I$, the rank of the Hankel structured matrix $H_d(k)$ is close to the signal length, exceeding the filter length $d$ that a practical CNN can implement, so the network's kernels cannot satisfy the projection condition required by deep convolutional framelets. The remedy is that in the projection domain the two artifacts separate: the low-dose noise lives inside the truncation mask $T$, where it corrupts the measured sinogram, and the cupping artifact corresponds to the missing data outside $T$. The proposed projection-domain CNN therefore solves two decoupled low-rank subproblems — denoising the measured projections inside $T$ and extrapolating the projections outside $T$ — after which the filtered backprojection image enters a second image-domain CNN. The authors report that this projection-domain-first design outperforms the same-size image-domain CNN, a U-Net with twice the parameters, and a W-Net unrolled twice in the image domain, and that it also improves the isolated interior-tomography and low-dose problems.","pith_inferences":["Beyond the paper, the same decoupling logic should transfer to other coupled CT distortions, such as sparse-view streaking combined with motion artifacts, whenever the two artifact types occupy disjoint regions of the projection or frequency domain.","The authors' argument implies a testable failure threshold: an image-domain CNN should degrade sharply once the Fourier support of the coupled artifact exceeds the network's effective filter length, so plotting reconstruction quality against filter length would directly probe the proposed mechanism.","If the mechanism holds, projection-domain auxiliary losses could regularize image-domain networks in general, since the paper treats projection data and FBP images as carrying the same information up to a linear transform."],"forward_implications":["On the ten-subject dataset with simulated fan-beam truncation and Poisson low-dose noise, the proposed network achieves the best NMSE and SSIM at every tested photon count and truncation ratio.","A projection-domain CNN used as the first stage outperforms an image-domain CNN with the same architecture, showing the training domain matters for coupled artifacts.","The singular-value spectra of the backbone's last feature maps are lower for the projection-domain CNN, consistent with the claim that it better satisfies the low-rank condition.","Two-times unrolled networks beat a single larger network: the W-Net with fewer parameters outperforms a U-Net with roughly twice the parameters.","The projection-domain-first design remains superior even when the problems are isolated, i.e., when there is no truncation or no projection noise."],"supporting_citations":[{"why":"Supplies the deep convolutional framelets theory and the rank/filter-length bound used to argue image-domain CNNs fail on coupled artifacts.","marker":"[23]"},{"why":"Extends the framelets analysis to encoder-decoder CNNs, supporting the low-rank interpretation of the feature maps.","marker":"[26]"},{"why":"Defines the U-Net backbone used in all compared architectures and in the proposed network.","marker":"[32]"},{"why":"Supplies the ten-subject dataset from which training, validation, and test slices are drawn.","marker":"[33]"},{"why":"Supplies the low-dose simulation method used to generate projection noise from the dataset.","marker":"[34]"},{"why":"The classical projection extrapolation approach for cupping artifacts that the extrapolation branch extends and improves.","marker":"[3]"},{"why":"Provides the compressed-sensing interior tomography baseline with total-variation penalty compared in the results.","marker":"[5]"},{"why":"Previous framing of U-Net via deep convolutional framelets for sparse-view CT that the coupled-artifact analysis builds on.","marker":"[17]"},{"why":"Supports the non-stationary Gaussian noise model for low-dose projection data used in the problem formulation.","marker":"[25]"}],"fun_headline_variants":["Projection-first deep learning cuts CT dose artifacts","Why image-only CNNs fail at low-dose CT","Dual-domain CNN beats image-only for CT artifacts","Split CT artifacts to fix them: projection-first wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cupping artifacts sit in the low-frequency range while low-dose image noise sits in the high-frequency range, so their combined artifact spans nearly the whole spectrum and exceeds what a network with practical filter sizes can represent; if real truncated low-dose scans do not separate this way, the theoretical motivation for the dual-domain design loses its force.","fun_headline_variants_meta":{"raw":{"variants":["Projection-first deep learning cuts CT dose artifacts","Why image-only CNNs fail at low-dose CT","Dual-domain CNN beats image-only for CT artifacts","Split CT artifacts to fix them: projection-first wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1736,"prompt_tokens":1098,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":714,"completion_tokens_details":{"reasoning_tokens":576}},"tokens_in":714,"tokens_out":638,"duration_ms":6684,"temperature":1.0,"reasoning_tokens":576,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:21:13.283776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Fourier support or Hankel-structured rank of the artifact directly from a real scanner's truncated low-dose projections, and train an image-domain CNN whose convolutional filter length matches the signal length; if the image-domain network then matches the dual-domain network's quality, the filter-length bound is not the limiting factor and the proposed explanation would be wrong even if the method still works.","supporting_citations":[{"cited_title":"Deep convolutional framelets: A general deep learning framework for inverse problems","cited_arxiv_id":null,"evidence_quote":"Supplies the deep convolutional framelets theory and the rank/filter-length bound used to argue image-domain CNNs fail on coupled artifacts."},{"cited_title":"Understanding geometry of encoder-decoder cnns","cited_arxiv_id":null,"evidence_quote":"Extends the framelets analysis to encoder-decoder CNNs, supporting the low-rank interpretation of the feature maps."},{"cited_title":"Tu-fg-207a-04: Overview of the low dose ct grand challenge","cited_arxiv_id":null,"evidence_quote":"Supplies the ten-subject dataset from which training, validation, and test slices are drawn."},{"cited_title":"Development and validation of a practical lower-dose-simulation tool for optimizing computed tomography scan protocols","cited_arxiv_id":null,"evidence_quote":"Supplies the low-dose simulation method used to generate projection noise from the dataset."},{"cited_title":"Algorithm to extend reconstruction field-of-view","cited_arxiv_id":null,"evidence_quote":"The classical projection extrapolation approach for cupping artifacts that the extrapolation branch extends and improves."},{"cited_title":"Compressed sensing based interior tomography","cited_arxiv_id":null,"evidence_quote":"Provides the compressed-sensing interior tomography baseline with total-variation penalty compared in the results."},{"cited_title":"Framing u-net via deep convolutional framelets: Application to sparse-view ct","cited_arxiv_id":null,"evidence_quote":"Previous framing of U-Net via deep convolutional framelets for sparse-view CT that the coupled-artifact analysis builds on."},{"cited_title":"Noise reduction with low dose ct data based on a modified rof model","cited_arxiv_id":null,"evidence_quote":"Supports the non-stationary Gaussian noise model for low-dose projection data used in the problem formulation."}],"review_version":1}