{"id":"70b07aa4-d47d-4231-a212-5ab118c0fcb3","arxiv_id":"2511.10500","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A learnable per-pixel regularization map inside an unrolled total-variation solver improves low-dose CT denoising by ~3.7 dB PSNR over FBP+U-Net.","lead":"This paper couples a total-variation denoiser with a small network that learns a per-pixel smoothing strength map, then trains the whole system end-to-end on simulated low-dose CT images. The result is a denoising method that beats a standard CNN baseline by about 3.7 dB PSNR on a CT test set while keeping the TV-based structure visible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed scalar data-fidelity weight w_data is an unstated, non-adaptive bottleneck; on real CT noise the learned λ-map may not compensate, so gains may not transfer.","rationale":"The reader's weakest_assumption identified the same cluster of issues: the simulation realism, the unspecified LambdaNet input, and the fixed scalar w_data. My concern is the specific internal consequence of Eq. (3): even if the simulation is perfectly realistic, the method's adaptivity is incomplete because the data-fidelity weight is global while CT noise is spatially varying. This is not an external validity threat only; it is a limitation in the model's design. The proposed concrete test would directly quantify the impact by comparing against oracle or per-pixel w_data. If the test shows little difference, the concern would be retired. I agree with the reader's conditional verdict: the paper is promising but needs additional validation and transparency before acceptance. My concern does not change that verdict, so I recommend leaving it CONDITIONAL.","tokens_in":4742,"tokens_out":6424,"duration_ms":63118,"concrete_test":"On the same DeepLesion/LoDoPaB-CT setup, train three variants: (a) LTV as reported (fixed scalar w_data), (b) LTV with oracle pixel-wise w_data(x) = 1/σ_data(x) computed from the known noise map and inserted in Eq. (3), and (c) LTV with w_data as a learned per-pixel map alongside λ. Compare PSNR/SSIM on the same test split with multiple random seeds and error bars. If (b) or (c) outperform (a) by more than ~0.5 dB PSNR, the fixed scalar w_data is a load-bearing bottleneck. Additionally, evaluate the reported model on a real low-dose CT dataset (e.g., AAPM Low-Dose CT Grand Challenge) to check whether the +3.7 dB gain reproduces.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is consistent gains over classical TV and FBP+U-Net on simulated low-dose CT. The method's only spatial adaptivity is the λ-map (Eq. 1), but the data-fidelity term in Eq. (3) uses a single scalar w_data = 1/σ_data for all pixels. Low-dose CT noise, even in the LoDoPaB-CT simulation, is not iid Gaussian: FBP reconstruction of Poisson-corrupted sinograms yields noise whose variance depends on local attenuation and dose. The paper never states the value of σ_data nor how it was estimated. If w_data is a global constant, the model over-trusts high-attenuation (high-noise) pixels and under-trusts low-noise pixels. The λ-map can only modulate the TV regularization strength; it cannot reweight the data fidelity. Thus the method is only partially adaptive — it cannot fully correct the noise-model mismatch. This is an internal limitation of the variational formulation, not merely an external-domain shift. If real clinical CT has different attenuation statistics, the learned fixed-w_data behavior may not transfer, and the reported gains may shrink or vanish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Learnable Total Variation (LTV) framework for low-dose CT image-domain denoising. The method couples a U-Net-style LambdaNet, which predicts a per-pixel regularization map λ, with a T=20-step unrolled Chambolle–Pock primal–dual TV solver that also learns step sizes and a relaxation parameter. The pipeline is trained end-to-end with a composite loss combining MSE, SSIM, a TV smoothness term on the λ-map, and a distribution-level term that rewards a non-trivial λ standard deviation. Experiments on ~1000 DeepLesion slices with LoDoPaB-CT-simulated 10% dose noise report mean PSNR/SSIM of 30.15 dB/0.854 for the full LTV, vs 27.99/0.816 for classical TV and 26.48/0.784 for FBP+U-Net. The conclusion claims consistent gains over both baselines while preserving variational interpretability, and acknowledges that validation is limited to a single simulated dataset.","tokens_in":5102,"tokens_out":4900,"duration_ms":54962,"significance":"If the claimed results hold, the contribution is a modest but useful extension of classical TV: replacing a scalar regularization parameter with a learned, spatially adaptive map in an unrolled solver that remains differentiable and interpretable. The paper ships code, describes a fully differentiable unrolled architecture, and presents a reproducible simulation pipeline, which are strengths. However, the significance is tempered by the lack of comparison to the closest prior work on learned regularization parameter maps (Kofler et al., ref. [9]), the absence of statistical variability analysis, and the use of a weakly performing FBP+U-Net baseline. The interpretability claim is also somewhat overstated: the learned λ-map shows only a weak correlation with image gradients (r≈0.15), and the fixed scalar data-fidelity weight limits the spatial adaptivity to the regularization term only. Overall, the idea is plausible but the current evidence does not yet establish a clear advantage over existing adaptive TV or unrolled methods.","major_comments":[{"comment":"The central empirical claim of 'consistent gains' rests on single mean PSNR/SSIM values with no error bars, no multiple runs, and no statistical significance test. Moreover, the FBP+U-Net baseline (26.48 dB) scores below classical TV (27.99 dB), which is atypical for this task and suggests the baseline was not tuned to its potential. Please report mean±std over multiple seeds/runs, include per-slice variance, and provide evidence that the FBP+U-Net was trained adequately (e.g., training curves, validation performance). Without these, the relative ranking of LTV versus a properly tuned CNN baseline is uncertain.","section":"Table I and Section III-A"},{"comment":"The data-fidelity term uses a fixed scalar weight w_data = 1/σ_data, but the value of σ_data and how it was estimated are never stated. Low-dose CT noise, even under the LoDoPaB-CT simulation, is spatially varying because it stems from Poisson corruption of sinograms followed by FBP. A constant w_data cannot adapt to local noise levels; the λ-map can only modulate TV strength, not data fidelity. This is an internal limitation of the proposed formulation, not just an external domain-shift concern. Please state σ_data, describe its estimation, and ideally compare against a version with a learned per-pixel data-fidelity map or a spatially varying w_data. This is directly relevant to the claim of spatial adaptivity.","section":"Eq. (3) and Section II-C"},{"comment":"The manuscript cites Kofler et al. [9], which learns spatial regularization parameter maps for variational reconstruction using algorithm unrolling, but does not compare against it empirically. This is the closest learned-parameter-map method, and the paper's novelty over it is not demonstrated by the current experiments. Please include a direct comparison, at least on the same simulated dataset, or clearly argue and demonstrate why the image-domain denoising setting and the chosen unrolled solver are sufficiently distinct that comparison is unnecessary.","section":"Related Work and Contribution (Section I)"},{"comment":"The text repeatedly refers to 'extensive ablation studies' and gives specific degradations (e.g., removing TV(λ) or Ldist(λ) costs approximately 0.6 dB PSNR and 0.01 SSIM), but no ablation table is provided. Without quantitative results, the design choices regarding the minimal λ-regularization are not verifiable. Please add a proper ablation table with means and variances. Also, the reported λ-map statistics (mean≈0.18, median∼1e-3) imply that most pixels receive essentially no TV regularization; discuss whether the method effectively reduces to data fidelity in smooth regions and why the weak gradient correlation (r≈0.15) supports the interpretability claim.","section":"Section II-D and ablation claims"}],"minor_comments":[{"comment":"The primal update is written as 'x(k+1) = x(k) +τ∇^T p(k+1) +τ wdata y /(1+τ wdata)', but the intended formula is (x^k + τ∇^T p^{k+1} + τ w_data y)/(1+τ w_data). Please add parentheses for clarity.","section":"Eq. (3)"},{"comment":"The input to LambdaNet is never explicitly stated. Is it the noisy image y only, or also the current iterate and/or intermediate features from the unrolled solver? Clarify this in the text, since it affects the reproducibility of the method.","section":"Section II-B and Fig. 1"},{"comment":"The symbol λ(e)max and the annealing schedule (0.01→5.0 over 25 epochs) appear only in the text and are not formalized in the equation. Please define the schedule explicitly, including the interpolation rule.","section":"Eq. (1) and annealing schedule"},{"comment":"The attention-gated skip connections and multi-scale fusion head are mentioned but not described in sufficient detail (e.g., how the weighted average is computed). The GitHub link is helpful, but the paper should be self-contained enough for a reader to reimplement without reverse-engineering code.","section":"Section II-A"},{"comment":"The abstract states 'up to +3.7 dB PSNR and 8% relative SSIM improvement' without specifying that these numbers are relative to FBP+U-Net, the weaker baseline. Please clarify to avoid overstatement.","section":"Abstract and Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant problem and the proposed method is plausible, but the evidence base is thin: a single simulated dataset, no error bars, an undertuned CNN baseline, and no empirical comparison to the closest prior work (Kofler et al. [9]). The fixed scalar data-fidelity weight is an unstated hyperparameter that limits the claimed adaptivity and should be addressed head-on. I could see this becoming a solid contribution after adding the missing experiments and ablations. The paper is within the scope of a CV/medical-imaging journal, but the novelty is incremental."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the method is legitimate: an unrolled Chambolle–Pock TV denoiser with a learned per-pixel λ-map, plus learnable step sizes and relaxation, applied to image-domain low-dose CT denoising. Second, the evidence is thinner than the 'consistent gains' claim: one simulated dataset, no error bars, a U-Net baseline that scores worse than classical TV (which suggests it wasn't tuned), and no comparison to the closest prior method, Kofler et al. The core idea of learning regularization maps is not new, but moving it to denoising and showing that a lightweight λ regularizer works better than strong structure-aligned priors is a real, if incremental, contribution.\n\nCredit where due: the paper is well-written, the method is clearly described, the code is available, and the authors explicitly acknowledge the single-dataset limitation in their conclusion. The ablation story about structure-aware λ priors destabilizing training is interesting and plausible, though the numbers are only described in prose.\n\nSoft spots, in rough order. (1) Fairness of baselines. FBP+U-Net at 26.48 dB PSNR is below classical TV (27.99 dB), which is unusual for a well-tuned network. This makes the headline +3.7 dB gain less convincing. (2) No error bars or significance tests. With ~1000 slices, run-to-run variance could affect the rankings. (3) Kofler et al. is cited but never benchmarked; a reader can't tell if LTV adds value beyond that existing framework in a new domain. (4) The scalar w_data issue raised in the stress-test is valid: a fixed data-fidelity weight can't fully adapt to spatially varying noise, and the λ-map only modulates TV strength. The paper doesn't state σdata or discuss this choice. This is a limitation for clinical transfer, not a fatal internal flaw, but it should be acknowledged. (5) Minor: the ablation results should be tabulated.\n\nThe paper deserves peer review. It's a plausible and reproducible method, and the questions it raises—how much adaptivity is needed, and what kind of λ regularization is stable—are useful for the unrolled-optimization community. I'd recommend sending it out, with referees asked to require stronger baselines, error bars, a Kofler comparison, and a clear statement about w_data.","headline":"A clean incremental extension of Kofler et al. to image-domain CT denoising, with a plausible method but thin evidence—worth refereeing, not a breakthrough.","tokens_in":5544,"tokens_out":4272,"would_cite":true,"duration_ms":44806,"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 replacing the single regularization strength in total-variation denoising with a learned per-pixel map—predicted by a small U-Net and fed into an unrolled primal–dual solver—outperforms both classical TV and a U-Net b","keywords":["low-dose CT","total variation","lambda mapping","unrolled optimization","primal-dual solver","image denoising","spatially adaptive regularization","deep learning"],"falsifier":"Take the trained LTV model and run it on real low-dose CT acquisitions (e.g., from a hospital) with paired full-dose references; if the PSNR/SSIM advantage over classical TV and U-Net shrinks to near zero, or if the predicted λ-map does not correlate with locally measured noise variance (estimated from repeated scans or noise-only regions), the central claim fails. A simpler laboratory test: feed the model a synthetic phantom with known spatially varying noise levels and check whether the λ-map tracks the noise pattern rather than anatomy.","tokens_in":4675,"feed_emoji":"🩻","tokens_out":4340,"duration_ms":40788,"temperature":0.7,"pith_summary":"Low-dose CT denoising is a trade-off between smoothing noise and keeping fine anatomy. Classical total variation (TV) uses one global strength parameter, so it either over-smooths or leaves noise. This paper claims the missing piece is spatial adaptivity: a learned per-pixel λ-map, produced by a U-Net-style LambdaNet, tells the TV solver where to denoise strongly and where to preserve edges. The whole system—LambdaNet plus a 20-step unrolled Chambolle–Pock solver with learnable step sizes—is trained end-to-end on simulated 10%-dose CT slices. If the claim holds, clinicians get an interpretable, variational alternative to black-box CNNs that is both better and explainable.","feed_headline":"Learned per-pixel TV beats U-Net by 3.7 dB on low-dose CT","feed_subtitle":"A tiny network predicts a regularization map for every pixel, making classical total-variation denoising adaptive and interpretable.","key_machinery":"The central object is the λ-map: a per-pixel regularization field predicted by LambdaNet (a four-stage attention-gated U-Net with a multi-scale fusion head), normalized to [λ_min, λ_max^(e)] with an exponential contrast-enhancing mapping and an annealed upper bound that ramps from 0.01 to 5.0 over the first 25 epochs. The λ-map enters a 20-iteration unrolled Chambolle–Pock primal–dual solver through the dual update: p ← Π_{||·||≤λ}(p + σ_d ∇ x̄), a pixel-wise projection onto an ℓ2 ball whose radius is λ. This locally bounds the TV penalty: small λ means strong smoothing, large λ preserves edges. The primal update balances the TV-driven smoothing against data fidelity with a weight w_data = 1","core_discovery":"On its own terms, the paper establishes that a spatially varying TV penalty—defined by a per-pixel regularization map λ(x) predicted from the noisy image and used to project the dual variable in each unrolled primal–dual iteration—can outperform a fixed-scalar TV baseline and a FBP+U-Net baseline on simulated low-dose CT. The reported numbers are 30.15 dB PSNR and 0.854 SSIM for LTV versus 27.99 dB/0.816 for classical TV and 26.48 dB/0.784 for FBP+U-Net. The authors further claim that adaptivity emerges naturally: with only light, stability-oriented regularization on λ (a TV-smoothness term and a variance-promoting term), the learned λ-maps end up sparse, non-uniform, and weakly correlated w","pith_inferences":["Beyond the paper: a testable extension is to train LTV on real paired low-dose/full-dose clinical CT and compare λ-maps against measured local noise variance; if the weak gradient correlation (r≈0.15) reflects true noise statistics rather than anatomy, λ could double as a noise-activity map for dose monitoring.","Beyond the paper: the same unrolled-TV-plus-λ-map design could be applied to other image-domain restoration tasks (MRI denoising, PET, ultrasound), where scalar regularization parameters are also a bottleneck.","Beyond the paper: the learned λ-map's dynamic range (mean 0.18, median ~1e-3) implies that most pixels are strongly denoised with a few edge pixels spared; this sparsity pattern may itself be a useful prior for understanding where black-box CNNs concentrate their corrections.","Beyond the paper: the annealing schedule for λ_max (0.01→5 over 25 epochs) is a training-stabilization trick that could be adopted in other learnable-regularizer frameworks; its removal would likely collapse training."],"forward_implications":["Spatially adaptive regularization can be learned end-to-end without hand-designed structural priors, so the same recipe may transfer to other variational priors (e.g., total generalized variation, wavelet sparsity) for inverse problems.","The learned λ-map provides a built-in interpretability channel: radiologists could inspect where the model chooses to denoise strongly versus preserve edges, unlike a black-box CNN.","The reported +0.73 dB gain from learning solver step sizes suggests that unrolled optimization dynamics are a meaningful source of performance, not just the adaptive map.","The minimal, stability-driven λ-regularization (TV(λ) plus a variance term) is sufficient and necessary; stronger structure-aligned λ-priors destabilize training, a finding that can guide future unrolled-method design.","If the approach generalizes to clinical dose levels and 3D volumes, it offers a practical upgrade path for existing TV-based CT pipelines, requiring only a forward pass of LambdaNet."],"fun_headline_variants":["Per-pixel TV map learns adaptivity, beats U-Net by 3.7 dB on CT","Adaptive TV via learned lambda outdoes U-Net in low-dose CT","Interpretable denoising: learned per-pixel TV tops U-Net by 3.7 dB","Tiny LambdaNet makes TV adaptive, +3.7 dB over U-Net on CT","Spatially varying TV: leaner than U-Net, better by 3.7 dB on CT"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire reported gain rests on the assumption that the simulated 10%-dose noise injected into DeepLesion slices (following the LoDoPaB-CT methodology) matches the noise statistics of real low-dose CT; if it does not, the λ-map will learn to fit simulation artifacts rather than clinical photon noise.","fun_headline_variants_meta":{"raw":{"variants":["Per-pixel TV map learns adaptivity, beats U-Net by 3.7 dB on CT","Adaptive TV via learned lambda outdoes U-Net in low-dose CT","Interpretable denoising: learned per-pixel TV tops U-Net by 3.7 dB","Tiny LambdaNet makes TV adaptive, +3.7 dB over U-Net on CT","Spatially varying TV: leaner than U-Net, better by 3.7 dB on CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1606,"prompt_tokens":695,"completion_tokens":911,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":787}},"tokens_in":439,"tokens_out":911,"duration_ms":8841,"temperature":1.0,"reasoning_tokens":787,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:23:30.813436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained LTV model and run it on real low-dose CT acquisitions (e.g., from a hospital) with paired full-dose references; if the PSNR/SSIM advantage over classical TV and U-Net shrinks to near zero, or if the predicted λ-map does not correlate with locally measured noise variance (estimated from repeated scans or noise-only regions), the central claim fails. A simpler laboratory test: feed the model a synthetic phantom with known spatially varying noise levels and check whether the λ-map tracks the noise pattern rather than anatomy.","supporting_citations":[],"review_version":1}