{"id":"7e028b3e-41fb-41e6-8da2-ebe4c22f850d","arxiv_id":"2608.10343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A noise-context-aware CT denoiser that injects autocovariance maps and adaptive convolutions, reporting quality gains over vanilla and noise-augmented baselines at reduced MACs.","lead":"The paper introduces ENCORE, a CT denoising framework that feeds the network an estimate of local noise strength and correlation and uses an adaptive convolution that reshapes itself per image region. It reports higher PSNR, SSIM, and lower hallucination scores than vanilla and noise-map baselines while cutting multiply-accumulate counts, though measured runtime does not clearly improve.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quality gains are demonstrated only under the same synthetic noise model used for training; the real-data evaluation is qualitative and the paper itself flags missing physical factors, so the central denoising claim is not yet established outside simulation.","rationale":"The paper's headline contribution is the injection of noise context into a denoising network. That contribution is validated only against simulated data produced by the same Poisson+Gaussian noise model used at training time. The N=1 autocovariance estimate in Eq. (6) is a practical proxy, and Table III shows N=1 suffices within the simulation, but the real-tabletop test is qualitative and Section VI explicitly lists scatter and beam hardening as unmodeled. The most load-bearing condition for the central claim is therefore that the synthesized context matches real noise statistics; if it does not, the FlyingConv mechanism loses its justification and the claimed quality gains are unverified. I also considered the latency/efficiency contradiction in Table I, where ENCORE is slower than Vanilla in every row while having fewer MACs; this is a real overstatement that the authors should fix, but it is secondary relative to the denoising-quality claim and is partly acknowledged in Section VI. The generalization gap is the concern that would most change the practical value of the contribution, so the reader's conditional verdict remains appropriate.","tokens_in":15318,"tokens_out":10761,"duration_ms":102029,"concrete_test":"Acquire repeated low-dose scans of a stationary phantom (or use Mayo2020 projection data with a higher-dose reference) to estimate the empirical noise autocovariance from multiple realizations; compare this with the synthetic V from Eq. (4) under identical geometry. Then evaluate ENCORE against Vanilla, +NADD, and +COV on the real data using quantitative metrics (PSNR/SSIM against a high-dose reference, or CNR in uniform ROIs) rather than qualitative inspection. If ENCORE's advantage over baselines disappears or reverses, the central quality claim is supported only within the simulation's noise model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of substantial denoising gains rests on the assumption that the autocovariance context V in Eq. (6), computed from a single synthesized noise map (N=1, w=5), correctly characterizes real CT noise. Eq. (4) generates the noise map from P_Lower = P_LD + W(0,P_LD) + N(0,sigma_e^2), so the context is derived from the same Poisson+Gaussian model used to create training pairs. All quantitative evidence in Table I comes from Mayo2016/Mayo2020 projection data simulated with this same noise model. The only real-world evaluation (Tabletop, Fig. 3 third row) is qualitative, has no reference image, and Section VI concedes that the noise model omits scatter, beam hardening, and other physical effects. If real CT noise deviates from the simulated model, the autocovariance maps are miscalibrated and FlyingConv's adaptive weighting may be guided by incorrect noise statistics; the claimed gains could vanish or reverse on clinical data. This is a load-bearing external-validity gap, not a tuning detail, because the method's entire mechanism is to feed noise context into the network.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ENCORE, a low-dose CT denoising framework that explicitly models CT noise statistics. It reformulates noise synthesis using a Cornish-Fisher expansion to capture Poisson skewness, extracts local autocovariance maps as noise context, and introduces a FlyingConv module that adaptively modulates convolution weights based on this context. The method is trained in a Noise2Noise setting and supports zero-shot conditional denoising by scaling the context maps at inference. Evaluations on simulated Mayo2016 and Mayo2020 datasets show consistent PSNR/SSIM/AUHOC improvements over vanilla, NADD, and autocovariance-concatenation baselines, with reduced MACs; a real tabletop phantom study is included for qualitative assessment.","tokens_in":15569,"tokens_out":22373,"duration_ms":174869,"significance":"If the results hold, ENCORE contributes a physically motivated way to inject CT noise statistics into a denoising network, with a clear component-wise ablation and an interesting zero-shot control mechanism. The paper is reproducible in principle (code is provided), and the experimental design is thorough within the simulation setting. The main strength is the coherent pipeline from noise synthesis to context estimation to adaptive convolution. However, the quantitative evidence is limited to simulated data generated with the same noise model used in training, and there is a technical inconsistency in the appendix derivation that must be resolved before the noise-synthesis claims can be accepted.","major_comments":[{"comment":"The definition of the skewness-corrected noise W is internally inconsistent. In Eq. (9), V_target is set to d(1-d)P_ND, and this value is used as the variance of W in Eq. (7). When W replaces N(0,P_ND) in Eq. (3), the injected term d a W has variance (d a)^2 V_target = d(1-d) * d(1-d)P_ND, rather than the required d(1-d)P_ND. The final β formula is correct only if W is intended to have variance P_ND, i.e., V_target should be P_ND, not d(1-d)P_ND. As written, the equations would produce noise with the wrong magnitude if implemented literally. Please correct the definitions and clarify whether V_target refers to the variance of W or to the variance of the scaled injected term.","section":"Section III-A / Appendix A"},{"comment":"The central claim of substantial denoising gains is quantitatively established only under the same simulated Poisson+Gaussian noise model used for training. The Mayo2020 results use forward-projected reference images with the same noise injection, differing only in acquisition geometry. The only real-data evaluation (Tabletop) is qualitative, has no reference image, and the Discussion explicitly concedes that scatter, beam hardening, and other physical factors are omitted. Because the method's entire mechanism is to feed noise-context maps derived from the simulated model into the network, the external validity of the method on real CT noise is not yet demonstrated. I recommend adding at least one quantitative real-data experiment (e.g., a phantom with known ground truth or a validated no-reference metric) or clearly stating in the abstract and conclusion that the quantitative gains are demonstrated on simulated data.","section":"Section IV-A, Section VI, and Fig. 3"},{"comment":"The claim of 'substantial gains in ... computational efficiency' is overstated. In Table I, ENCORE consistently has higher latency than the Vanilla baseline (e.g., UNet Base at 10% dose: 10.24 ms vs 8.38 ms), and its MACs are lower only than the +COV and +NADD variants, not than Vanilla. The Discussion (Section VI) attributes this to memory-bandwidth limitations. Please clarify that the efficiency advantage is in MACs and relative to noise-aware baselines, not in end-to-end inference speed against the vanilla network.","section":"Abstract and Table I"}],"minor_comments":[{"comment":"The text states that the Gaussian term N in Eq. (4) is replaced by W in the final implementation, but the equation itself still shows N(0,P_LD). Please update the equation or add an explicit note to avoid confusion.","section":"Section III-A, Eq. (4)"},{"comment":"The underline notation for statistical significance is defined, but in several rows the best-performing ENCORE value is not underlined (e.g., UNet Base 25% SSIM). Please report the actual p-values or specify which comparisons achieved significance to support the 'most cases' statement.","section":"Table I"},{"comment":"The ablation for the skewness-corrected noise model reports only PSNR and SSIM. Since the correction targets higher-order statistics, consider reporting a metric more sensitive to texture or noise statistics, such as a local noise power spectrum error, to better demonstrate the effect.","section":"Section V-C1, Table II"},{"comment":"The autocovariance estimator with N=1 and w=5 is justified by the local-stationarity assumption [9], but the paper does not quantify the bias/variance of this estimator under simulated noise. A small synthetic experiment showing the estimator's error as a function of w and N would strengthen the default parameter choice.","section":"Section III-B, Eq. (6)"},{"comment":"Reference [24] appears to be titled 'sfrc for assessing hallucinations...'; this is likely a typo for 'SFRC' (structural frequency-domain hallucination metric). Please verify the reference title and the acronym usage.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The appendix inconsistency is the most serious technical issue; it should be fixed before acceptance. The external-validation concern is common in CT denoising papers, but it is more salient here because the method's mechanism is explicitly tied to the synthetic noise model. The authors should consider whether a real-data quantitative experiment is feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a serious methods paper, worth a real referee, but read it with the abstract's two claims separated. The quality claim is solid only inside the simulated noise model; the efficiency claim is overstated.\n\nWhat is actually new: FlyingConv, a convolution whose weights are modulated on the fly by noise autocovariance maps, plus a Cornish–Fisher correction to the Gaussian noise approximation used for Noise2Noise pair generation. The autocovariance representation is a real improvement over NADD's raw noise maps – feeding the network a stable statistic instead of a stochastic realization makes sense. The appendix derivation of the skewness-corrected W is clean and not fitted to outcomes. The ablations are thorough: window size, patch size, number of noise maps, group size, and the zero-shot d_target sweep all get tested. The statistical testing with paired Wilcoxon and Holm correction is better than what most papers in this area do.\n\nSoft spots, in order of softness. First and most important: every quantitative result in Table I is on simulated Mayo data where the noise was generated with the same model used to construct the training pairs. The real-world tabletop scan is qualitative only, with no reference image, and the paper itself admits the noise model omits scatter, beam hardening, and other physical effects. Since the whole mechanism is 'give the network accurate noise context', a mismatch between simulated and real noise could make the autocovariance maps miscalibrated and the adaptive weights misleading. That is a load-bearing external-validity gap, not a tuning detail. The paper is honest about it in Section VI, but the abstract's 'substantial gains' should be conditioned on 'in simulation'.\n\nSecond, the efficiency claim. The MACs numbers drop substantially, but measured latency for +ENCORE is about the same as +COV and sometimes higher than Vanilla. The paper acknowledges this – it is a memory-bandwidth bottleneck, not an arithmetic one – but then still sells the framework as 'computational efficiency' in the abstract and contributions. That is a mismatch between what is measured and what is claimed.\n\nThird, the absolute quality gains are modest: 0.1–0.3 dB PSNR over strong baselines, and at 25% dose on Mayo2016 the UNet +ENCORE actually has lower SSIM than Vanilla or +NADD in a few configurations. The zero-shot SSIM adjustment mostly rescues this, but it is worth noting that the default configuration is not uniformly better.\n\nWho is this for? Researchers actively working on CT denoising, noise simulation, or adaptive convolutions for medical images. It does not resolve a long-open question, but it gives a credible new architecture and a reproducible pipeline. The code is public, which counts for something.\n\nMy recommendation: send it to peer review. There is enough novelty and rigor to warrant referee time. The authors should be pushed to either provide a quantitative real-data evaluation (even with a phantom or a pre/post clinical study) or tone down the generalizability claims in the abstract. The simulation-only evidence is fine for a methods paper, but then say so.","headline":"A well-engineered CT denoising paper that convincingly shows noise-context conditioning helps in simulation, but the headline quality and efficiency claims outrun the evidence: real-world gains are only qualitative and the latency benefit is not realized.","tokens_in":16097,"tokens_out":1856,"would_cite":false,"duration_ms":20109,"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":"Telling CT denoisers about their own noise cuts cost and lifts quality.","keywords":["low-dose CT denoising","noise autocovariance context","adaptive convolution","Cornish-Fisher expansion","self-supervised denoising","Noise2Noise","zero-shot conditional denoising","computed tomography"],"falsifier":"Run ENCORE on real raw projection data from a scanner with significant scatter and beam hardening and compare it against the simpler autocovariance-concatenation baseline: if ENCORE does not beat it, the FlyingConv mechanism is not carrying the claimed advantage. Separately, acquire many repeated scans of a static phantom and compare the true measured autocovariance with the paper's single-realization $5\\times5$ estimator to test the local-stationarity premise directly.","tokens_in":15132,"feed_emoji":"🩻","tokens_out":11741,"duration_ms":87223,"temperature":0.7,"pith_summary":"This paper tries to establish that a CT denoising network performs better and runs faster when it is explicitly told what its own noise looks like, rather than when it is merely handed a noisy image or a raw random noise sample. The ENCORE framework estimates a local noise autocovariance map from a single synthesized noise realization, feeds that map into adaptive convolution layers whose weights change per image region, and trains the full pipeline under a Noise2Noise self-supervised scheme with a skewness-corrected noise model. In the reported experiments on simulated low-dose CT benchmarks and a real tabletop cone-beam scan, ENCORE improves PSNR, SSIM, and a hallucination metric over vanilla, noise-augmented, and autocovariance-concatenation baselines, with the largest gains at the unseen 10% dose level and on cross-vendor data, while reducing multiply-accumulate operations. If correct, the work shows that packaging noise statistics into a stable context is a more direct route to better CT denoising than merely enlarging the model or adding attention modules.","feed_headline":"Telling CT denoisers about their own noise cuts cost, lifts quality","feed_subtitle":"Local autocovariance guides adaptive weights, improving quality at unseen doses with zero-shot texture control.","key_machinery":"The load-bearing object is the noise autocovariance map $V$, computed by Eq. (6) as a local average of products of noise-map values at small spatial lags, with defaults $p=5$, $w=5$, $N=1$ and signed-log normalization. This map is a stable surrogate for the non-stationary, spatially correlated CT noise that generic denoisers assume away. The second load-bearing piece is FlyingConv, an adaptive convolution that predicts spatially varying kernel weights from pooled anatomical features and the autocovariance map, sharing each kernel across two channels and fusing interpolation with convolution to cut memory traffic. The supporting machine is the Cornish-Fisher skewness correction, which replaces the Gaussian noise term in training-pair synthesis with $W(0,V_{\\mathrm{target}})=\\sqrt{V_{\\mathrm{target}}}(\\alpha N(0,1)+\\beta(N(0,1)^2-1))$.","core_discovery":"The central claim is that CT noise can be turned into a stable, model-friendly input: instead of feeding raw stochastic noise realizations, the method aggregates them into local autocovariance maps that capture noise power and spatial correlation, and it makes those maps actively steer the network through FlyingConv, a convolution whose per-region kernel weights are predicted from anatomy and noise-context features. The paper reports that this combination outperforms all tested variants, including a version that simply concatenates the autocovariance map to the input, most clearly in the ultra-low-dose regime at 10% dose that the model never saw during training. It also reports that scaling the intensity of the noise context at inference changes the output texture in a zero-shot way, letting a single trained model trade denoising strength against texture preservation. A skewness-corrected noise model based on a Cornish-Fisher expansion is used during training-pair generation to better match the physics of photon starvation.","pith_inferences":["The authors demonstrate zero-shot dose control only on simulated data; it remains an open extension to show that the same scaling behaves predictably on real clinical raw data where the noise model is imperfect.","The fixed $5\\times5$ window invites a learned or anatomy-guided aggregation of noise statistics, which could improve accuracy where the local-stationarity assumption is weakest.","The same autocovariance-context idea could transfer to other tasks dominated by non-stationary noise, such as iterative reconstruction, metal artifact reduction, or low-dose super-resolution.","The gap between multiply-accumulate operations and measured latency, which the paper attributes to memory bandwidth, suggests that hardware-side optimizations for adaptive-weight convolutions are the next lever for speed."],"forward_implications":["Self-supervised CT denoising can be improved by changing what is fed into the network rather than by enlarging the model or adding perceptual losses.","A single trained model can produce outputs with different residual noise and texture levels by scaling context maps at inference, with no retraining or extra post-processing.","The reported multiply-accumulate reductions suggest the approach could fit into clinical pipelines where the full reconstruction-to-denoise chain must run on a single GPU.","Because the context is estimated per image, the method should transfer across scanners and reconstruction kernels more gracefully than a static denoiser, consistent with the cross-vendor results."],"supporting_citations":[{"why":"It supplies the Noise2Noise self-supervised paradigm that lets ENCORE train without clean reference images.","marker":"[3]"},{"why":"It provides the Gaussian-approximation arbitrary-dose pair generation that ENCORE extends with skewness correction.","marker":"[5]"},{"why":"It is the noise-augmented denoising baseline whose raw noise-map input ENCORE replaces with autocovariance context.","marker":"[7]"},{"why":"It provides the Cornish-Fisher expansion used to inject Poisson skewness into synthesized noise.","marker":"[8]"},{"why":"It establishes the local stationarity of CT noise that justifies the small-window autocovariance estimator.","marker":"[9]"},{"why":"It is the Malleable Convolution prior for adaptive weights that FlyingConv extends with noise context and fused interpolation.","marker":"[13]"},{"why":"It is the public low-dose CT grand challenge dataset used for training and quantitative evaluation.","marker":"[17]"},{"why":"It is the multi-vendor low-dose CT projection dataset used to test cross-vendor generalization.","marker":"[19]"},{"why":"It defines the hallucination metric AUHOC used to judge whether denoising removes real structures.","marker":"[24]"}],"fun_headline_variants":["Noise context maps steer CT denoising, boost quality","Efficient CT denoising with adaptive noise-aware kernels","Zero-shot texture control in CT via noise context","FlyingConv adapts to CT noise for slicker low-dose denoising","CT denoising: local noise context cuts compute, lifts quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire context pipeline rests on the claim that a single synthesized noise realization, averaged over a $5\\times5$ window, faithfully represents the true local noise correlation of real CT scans; if scatter, beam hardening, or anatomy-dependent non-stationarity makes that estimate wrong, the context misleads the network.","fun_headline_variants_meta":{"raw":{"variants":["Noise context maps steer CT denoising, boost quality","Efficient CT denoising with adaptive noise-aware kernels","Zero-shot texture control in CT via noise context","FlyingConv adapts to CT noise for slicker low-dose denoising","CT denoising: local noise context cuts compute, lifts quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1470,"prompt_tokens":925,"completion_tokens":545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":541,"tokens_out":545,"duration_ms":5333,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:22:52.815026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ENCORE on real raw projection data from a scanner with significant scatter and beam hardening and compare it against the simpler autocovariance-concatenation baseline: if ENCORE does not beat it, the FlyingConv mechanism is not carrying the claimed advantage. Separately, acquire many repeated scans of a static phantom and compare the true measured autocovariance with the paper's single-realization $5\\times5$ estimator to test the local-stationarity premise directly.","supporting_citations":[{"cited_title":"Simulating arbitrary dose levels and inde- pendent noise image pairs from a single ct scan,","cited_arxiv_id":null,"evidence_quote":"It provides the Gaussian-approximation arbitrary-dose pair generation that ENCORE extends with skewness correction."},{"cited_title":"Noise-augmented deep denois- ing: A method to boost ct image denoising networks,","cited_arxiv_id":null,"evidence_quote":"It is the noise-augmented denoising baseline whose raw noise-map input ENCORE replaces with autocovariance context."},{"cited_title":"Moments and cumulants in the speci- fication of distributions,","cited_arxiv_id":null,"evidence_quote":"It provides the Cornish-Fisher expansion used to inject Poisson skewness into synthesized noise."},{"cited_title":"The noise power spectrum in ct with direct fan beam reconstruction,","cited_arxiv_id":null,"evidence_quote":"It establishes the local stationarity of CT noise that justifies the small-window autocovariance estimator."},{"cited_title":"Fast and high quality image denoising via malleable convolution,","cited_arxiv_id":null,"evidence_quote":"It is the Malleable Convolution prior for adaptive weights that FlyingConv extends with noise context and fused interpolation."},{"cited_title":"Low-dose ct image and projec- tion dataset,","cited_arxiv_id":null,"evidence_quote":"It is the multi-vendor low-dose CT projection dataset used to test cross-vendor generalization."}],"review_version":1}