{"id":"7c79fd95-dd55-45f6-bc25-966bfff8a9d2","arxiv_id":"2501.13961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A plug-and-play iterative reconstruction framework with a learned artifact-removal prior and automatic regularization selection achieves fast, generalizable cone-beam XCT reconstructions on large industrial volumes.","lead":"The paper presents a deep learning iterative reconstruction method for industrial cone-beam X-ray CT that alternates an artifact-removal neural network with a physics-based data-consistency step, using an automatic quality metric to set the regularization strength each iteration. It reports cleaner 3D reconstructions of dense metal parts from sparse and low-dose scans, with better robustness to unseen scanning conditions than a single-step supervised network.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative evidence is reference-dependent: the same 200 kV/0.6 s scan yields a 0.02 dB margin in the voltage table but a 0.86 dB margin in the integration-time table because different MBIR references are used.","rationale":"Read in good faith, the paper reports a practically motivated PnP algorithm with real industrial cone-beam data and plausible runtime advantages. The algorithmic design (artifact-removal CNN + adaptive beta) is coherent, and the paper includes ablation of beta selection and CG iterations. My central concern is not with the method's mechanics but with the metric-level evidence for the headline claim. The same acquisition appears in two subtables with different PSNR margins due to different references; that is an internally checkable contradiction, not a matter of taste. This is more directly load-bearing than the reader's domain-shift concern: even if the CNN were a perfect proximal operator, the claimed outperformance would still be unsubstantiated if the numbers are reference-dependent. The reader's concern about CNN behavior on intermediate iterates is legitimate and should be tested, but the current tables already give a concrete reason to require re-analysis. I therefore keep the verdict conditional (unchanged) but with a sharper, testable condition: matched-reference recomputation. If the recomputation preserves the margins, the central empirical claim survives.","tokens_in":16273,"tokens_out":11244,"duration_ms":115611,"concrete_test":"Recompute every row of Table II using a matched-voltage, high-integration-time MBIR reference (160 kV/3.6 s, 180 kV/3.6 s, 200 kV/3.6 s) and also report the 200 kV/0.6 s row with both the 180 kV and 200 kV references. If the PSNR/SSIM advantage over BHCN-DLMBIR remains at least 1 dB in all voltage rows and the 200 kV/0.6 s margin is the same under both references, the reference mismatch is not material; if margins shrink below 0.5 dB or reverse, the paper's superiority claim needs to be withdrawn or substantially revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II contains an internal inconsistency for the overlapping condition 200 kV source voltage, 0.6 s integration time. The top sub-table (variation in source voltage) reports BHCN-DLMBIR PSNR 38.89 dB and Proposed 38.91 dB (margin 0.02 dB), with reference BHCN-MBIR at 180 kV, 3.6 s. The middle sub-table (variation in integration time) reports DLMBIR 33.70 dB and Proposed 34.56 dB (margin 0.86 dB) for the same scan, with reference BHCN-MBIR at 200 kV, 3.6 s. Thus the size, and in principle even the sign, of the claimed outperformance depends on which reference image is chosen. For the voltage experiment, all rows are compared to a single 180 kV reference despite same-voltage high-integration-time MBIR references being available (Table I). Because PSNR/SSIM are computed against MBIR reconstructions rather than a true ground truth, a method that smooths away small pores (as the paper admits) can score higher without being more accurate for defect detection. The central claim that the method 'surpasses' DLMBIR on out-of-distribution voltage scans is therefore not currently supported by the reported numbers until this reference sensitivity is resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative reconstruction framework for industrial cone-beam X-ray computed tomography that combines a pre-trained artifact-removal CNN as a prior within a half-quadratic splitting (HQS) plug-and-play formulation, an adaptive regularization parameter selected by a BRISQUE-based grid search on center slices, and a conjugate-gradient data-consistency step. The framework is tested on real scans of steel, AlCe, and Inconel parts under out-of-distribution acquisition conditions (varying voltage, integration time, sparsity, and material). The authors report that the method outperforms a single-step deep-learning baseline (DLMBIR) on most OOD settings, reconstructs volumes of industrial size in a few outer iterations, and reduces runtime by an order of magnitude compared with MBIR.","tokens_in":16593,"tokens_out":5715,"duration_ms":55999,"significance":"If the reported results hold, this is a practically relevant contribution: it addresses the three-way trade-off between reconstruction quality, runtime, and memory scalability that has limited deep-learning iterative reconstruction in industrial XCT. The use of real large-scale industrial scans with realistic OOD settings, the memory-efficient 2D-patch training for 3D volumes, and the automated β-selection strategy are valuable strengths. The paper also demonstrates an order-of-magnitude runtime reduction over MBIR while maintaining competitive PSNR/SSIM. However, the quantitative evidence for the central outperformance claim is undermined by reference-dependent metrics and an internal inconsistency in Table II, so the significance hinges on a re-analysis of those numbers.","major_comments":[{"comment":"The same scan condition (ST, 200 kV, 0.6 s integration time) is evaluated against two different MBIR references and yields contradictory margins: with the 180 kV, 3.6 s reference, BHCN-DLMBIR gives 38.89 dB and the proposed method 38.91 dB (margin 0.02 dB); with the 200 kV, 3.6 s reference, the same methods give 33.70 dB and 34.56 dB (margin 0.86 dB). Since PSNR/SSIM are computed against MBIR reconstructions rather than a true ground truth, the size and even the sign of the claimed improvement depend heavily on the selected reference. The central claim that the method surpasses DLMBIR on OOD voltage scans is therefore not supported by the current numbers. The authors should recompute all metrics against a single consistent reference (or report multiple references per row) and discuss the sensitivity of the ranking to that choice.","section":"Table II (top vs. middle sub-tables)"},{"comment":"All PSNR/SSIM comparisons in Tables II and III use dense-view MBIR reconstructions as the reference, not true ground truth. The paper acknowledges in Section IV-A and Section V that the proposed method loses some smaller pores. A method that smooths away small pores can score higher on PSNR/SSIM against a noisy or structured MBIR reference without being more accurate for defect detection, which is a primary industrial use case. The authors should quantify pore-level fidelity (e.g., detection rate, size error) or validate on a phantom with known ground truth to substantiate the claim that their method outperforms DLMBIR on tasks that matter in practice.","section":"Section IV-A and Section V (metrics vs. pore preservation)"},{"comment":"The artifact-removal CNN Dθ is trained on sparse-view FDK reconstructions paired with MBIR targets, but in Algorithm 1 it is applied to intermediate iterates x_{k-1} that have already undergone one or more CG data-consistency updates. The paper does not analyze whether these iterates lie in the training distribution of Dθ, nor does it provide any stability or fixed-point analysis of the iteration. The empirical PSNR curves in Fig. 4 show no divergence for the one tested case, but this is insufficient to support the robustness claim for other OOD settings. I recommend adding a distribution-shift analysis (e.g., statistics of intermediate iterates vs. training inputs) or an ablation replacing Dθ with a Gaussian denoiser trained for the noise levels encountered in the iterates, to validate the proximal-operator assumption.","section":"Section II-B, Eq. (4)"},{"comment":"The abstract states that the method surpasses state-of-the-art supervised learning methods trained on the same data, but Table III shows that for the out-of-distribution materials ALC and IN the proposed method is slightly worse than BHCN-DLMBIR in PSNR (35.27 vs. 35.45 dB for ALC; 37.56 vs. 37.63 dB for IN) and the text says the methods perform 'at par'. The claim in the abstract overstates the material-generalization results and should be qualified to specify that the gains are on voltage, integration-time, and sparsity variations, while material transfer is at best at par with the baseline.","section":"Abstract and Section IV-D, Table III"},{"comment":"The benefit of the adaptive regularization parameter strategy over fixed β is demonstrated on a single sample (ST, 160 kV, 0.6 s). Given that the adaptive β selection is a key contribution and is used in all experiments, the authors should validate it on additional OOD conditions (e.g., varying integration time or sparsity) and, if possible, show that the BRISQUE-based selection correlates with the evaluation metrics PSNR/SSIM or with defect-preservation measures. Without such validation, the claims that the selection strategy is 'automated' and generally beneficial remain supported by only one case.","section":"Section IV-E, Fig. 4"}],"minor_comments":[{"comment":"The caption of Fig. 8 says the reference is BHCN-MBIR at 160 kV, 3.6 s, while Section IV-E states the reference is BHCN-MBIR at 180 kV, 3.6 s. One of these is inconsistent and should be corrected.","section":"Figure 8 caption vs. Section IV-E"},{"comment":"The symbol r is used both for the geometric ratio (Param. r = 0.5) and for the CNN output (Input: CNN output r). Rename one of them to avoid ambiguity, for example use ρ for the ratio.","section":"Algorithm 2"},{"comment":"The top sub-table is labeled 'Variation in Source Voltage' but includes a 200 kV row, which is the training voltage and not a variation; moreover this same 200 kV / 0.6 s condition appears with different PSNR/SSIM values in the middle sub-table. The authors should clarify how rows are grouped and ensure the 200 kV condition is either not labeled as a variation or is excluded from the voltage-variation comparison.","section":"Table II top sub-table"},{"comment":"There is a typo: 'manufactured with with out-of-distribution materials' should read 'manufactured with out-of-distribution materials'.","section":"Section IV-D"},{"comment":"The text says the proposed method requires 'three or fewer iterations (≤ 1600 seconds)', while Table IV reports 1612 s. These numbers should be reconciled.","section":"Section IV-G and Table IV"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful new thing here is the BRISQUE-driven adaptive beta selection inside the PnP/CG loop. That is a neat idea, and Figure 4 shows it beating fixed-beta choices on the one test case they report. The other piece—using an artifact-removal CNN as the proximal operator instead of a Gaussian denoiser—is less novel but reasonable for industrial XCT, and the large-volume demonstration on real steel, AlCe, and Inconel parts is genuine evidence. The runtime comparison is meaningful: MBIR takes about 7.5 hours, the proposed method about 27 minutes, on similar volumes. The paper is also honest about losing small pores, which is a real trade-off.\n\nThe main problem is the reference inconsistency in Table II. The same 200 kV / 0.6 s scan shows up in two sub-tables: with the 180 kV / 3.6 s MBIR reference, the proposed PSNR is 38.91 dB; with the 200 kV / 3.6 s MBIR reference, it is 34.56 dB. That 4 dB swing is entirely from the reference choice, not from the reconstruction. Since every PSNR/SSIM number is computed against MBIR rather than true ground truth, the claim that the method 'surpasses' supervised DL on OOD voltage scans rests on a reference-dependent comparison. At 200 kV the margin is 0.02 dB. The stress-test note is on target: this needs to be fixed before the superiority claim is credible.\n\nBeyond that, there are standard but real concerns: no error bars, no code or data released, and the CNN prior is trained on FDK inputs but applied inside the loop to iterates that have already been partially data-consistency-corrected. That domain shift is never examined. The empirical stability over four outer iterations suggests it doesn't blow up in practice, but the paper gives no analysis. I'd also want the adaptive beta selection to be tested on more than one scan setting; the current evidence is a single 160 kV / 0.6 s case.\n\nWho is this for? Anyone working on PnP reconstruction for large-scale CT, and industrial XCT people who need to get from 7 hours to 30 minutes. The contribution is engineering-solid but not deep. Worth a serious referee: the method is practical, the adaptive beta is a distinctive idea, and the runtime gain is real. The revision needs to standardize the reference images, release artifacts or at least describe the data availability, and address the prior's OOD behavior inside the loop.","headline":"A practical PnP/adaptive-beta framework with real runtime gains, but the headline outperformance is reference-dependent and needs a fix before the claim is solid.","tokens_in":17110,"tokens_out":2603,"would_cite":true,"duration_ms":24098,"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":"An iterative algorithm alternating an artifact-removal CNN with a conjugate-gradient data-consistency step reconstructs dense-metal cone-beam XCT volumes from sparse, low-dose scans in a few iterations and generalizes to unseen…","keywords":["cone-beam computed tomography","plug-and-play priors","model-based iterative reconstruction","artifact removal CNN","regularization parameter selection","out-of-distribution generalization","sparse-view CT","industrial XCT"],"falsifier":"Take a sparse-view scan with a known reference volume, run Algorithm 1 for K=4 outer iterations, and record PSNR and SSIM of each iterate against the reference; if the metrics do not improve from iteration 1 to 4, or if $D_\\theta(x_{k-1})$ has higher error on the intermediate iterates than on the FDK initialization, the CNN-as-proximal assumption is not holding.","tokens_in":16070,"feed_emoji":"🛠️","tokens_out":9774,"duration_ms":84192,"temperature":0.7,"pith_summary":"The paper aims to make model-based iterative reconstruction practical for industrial cone-beam XCT, where volumes can reach 1800-by-1800-by-1800 voxels and dense metal parts make fast scans noisy. It proposes an iterative algorithm that alternates between a CNN trained to remove artifacts from sparse-view FDK images and a conjugate-gradient step that enforces consistency with the measured projections. A key ingredient is automated regularization-parameter selection: each outer iteration reconstructs a few center slices for a geometric progression of candidate values and keeps the one with the lowest BRISQUE score. On real scans of steel, AlCe, and Inconel parts under out-of-distribution voltages, integration times, and sparsity levels, the method reports higher PSNR and SSIM than single-step DLMBIR, with reconstruction times near 27 minutes versus about 7.5 hours for MBIR.","feed_headline":"Fast iterative CT method cleans dense-metal scans in minutes","feed_subtitle":"CNN plus adaptive regularization beats single-step deep learning on unseen scan settings.","key_machinery":"The central object is the alternating loop of Algorithm 1: a pre-trained artifact-removal CNN $D_\\theta$ takes the current iterate $x_{k-1}$ and outputs $z_k$; a regularization-selection subroutine reconstructs a few center slices for 14 candidate $\\beta$ values on a geometric progression and chooses the one with the lowest BRISQUE score; and ten conjugate-gradient iterations solve the quadratic data-consistency subproblem $(A^\\top A + \\beta_k I)x = A^\\top y + \\beta_k z_k$. The CNN is a 2D residual U-Net trained on paired 256-by-256 patches extracted from sparse-view FDK reconstructions and dense-view MBIR volumes, which keeps training memory independent of volume size.","core_discovery":"The paper's central claim is that a CNN trained only to remove artifacts from FDK reconstructions can act as the prior inside an iterative reconstruction loop, provided the regularization parameter is re-selected at every iteration; the result is clean reconstructions in three or fewer outer iterations, with better out-of-distribution generalization than single-step supervised deep learning. Concretely, Algorithm 1 initializes with FDK, applies the artifact-removal CNN, chooses $\\beta$ by a BRISQUE-guided grid search on center slices, and solves the quadratic data-consistency subproblem with ten conjugate-gradient iterations; repeating this loop suppresses streaks, rings, and noise that single-step DLMBIR leaves behind. The experimental evidence is on real 3D-printed steel at 160–200 kV and 0.6–3.6 s integration times, on steel at twice the training sparsity, and on AlCe and Inconel parts, with PSNR gains of roughly 1–3 dB over DLMBIR in out-of-distribution settings and essentially tied performance on in-distribution settings.","pith_inferences":["The paper does not examine how the artifact-removal CNN behaves on intermediate iterates that are no longer FDK-like; a direct test would be to measure the CNN's output error on $x_{k-1}$ across outer iterations, since an unstable prior there could make the loop diverge even if each step looks plausible.","The BRISQUE-based $\\beta$ selection assumes center slices represent the full volume; for parts with strong axial material or geometry variation, the optimal parameter might differ by slice, which is a testable extension.","The runtime advantage rests on stopping after three outer iterations; if harder out-of-distribution conditions require more outer loops, the gap to single-step deep learning narrows, so deployment studies should report quality versus iteration count.","Because the data-consistency step and $\\beta$ selection are not specific to cone-beam geometry, the same alternating framework could be retrained for other artifact types such as motion or metal artifacts in medical CT."],"forward_implications":["Industrial XCT can move toward in-line inspection: the proposed algorithm reconstructs roughly 1500-by-1800-by-1800 volumes in about 27 minutes on four GPUs, compared with about 7.5 hours for MBIR.","The method keeps image quality on acquisition settings the CNN never saw, including lower source voltage, shorter integration time, and higher view sparsity, where single-step DLMBIR leaves residual rings and streaks.","Training on 2D patches and storing a single network keeps memory complexity at $O(1)$ rather than $O(K)$ for unrolled networks, so the approach can scale to very large volumes.","The adaptive $\\beta$ selection outperforms every fixed regularization parameter tested in the paper and preserves small pores better than fixed choices.","The alternating structure is presented as a general recipe for large-scale inverse imaging problems beyond cone-beam XCT."],"supporting_citations":[{"why":"Defines the FDK cone-beam reconstruction used for initialization and for producing the low-quality training inputs.","marker":"[6]"},{"why":"Fast MBIR algorithm that provides the reference reconstructions used as ground truth and the runtime baseline.","marker":"[8]"},{"why":"Single-step DLMBIR baseline, the beam-hardening correction network, and the source of the paired FDK-to-MBIR training data.","marker":"[12]"},{"why":"U-Net architecture used as the artifact-removal CNN prior.","marker":"[21]"},{"why":"Introduces the plug-and-play priors framework that justifies replacing the proximal map with a pre-trained network.","marker":"[29]"},{"why":"Consensus-equilibrium rationale for why fixed regularization is standard, which the adaptive selection strategy departs from.","marker":"[44]"},{"why":"BRISQUE no-reference image quality metric that drives the regularization parameter selection.","marker":"[45]"},{"why":"Supplies the GPU-accelerated cone-beam forward and adjoint operators used in the conjugate-gradient data-consistency step.","marker":"[49]–[51]"}],"fun_headline_variants":["Iterative deep CT cleanup tames dense-metal scans in minutes","CNN-in-the-loop CT reconstruction beats single-step on unseen scans","Adaptive regularized iterative CT cleans dense metal in few iterations","Deep iterative XCT for dense metal generalizes beyond training data","Artifact-removing CNN prior with auto-tuned beta for fast dense-metal CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the artifact-removal CNN, trained on FDK reconstructions, also acts as a stable and valid prior on the intermediate iterates that have already been partially corrected by the data-consistency step; the paper gives no analysis of that domain shift.","fun_headline_variants_meta":{"raw":{"variants":["Iterative deep CT cleanup tames dense-metal scans in minutes","CNN-in-the-loop CT reconstruction beats single-step on unseen scans","Adaptive regularized iterative CT cleans dense metal in few iterations","Deep iterative XCT for dense metal generalizes beyond training data","Artifact-removing CNN prior with auto-tuned beta for fast dense-metal CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4156,"prompt_tokens":898,"completion_tokens":3258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":3168}},"tokens_in":514,"tokens_out":3258,"duration_ms":23787,"temperature":1.0,"reasoning_tokens":3168,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:09:20.202529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sparse-view scan with a known reference volume, run Algorithm 1 for K=4 outer iterations, and record PSNR and SSIM of each iterate against the reference; if the metrics do not improve from iteration 1 to 4, or if $D_\\theta(x_{k-1})$ has higher error on the intermediate iterates than on the FDK initialization, the CNN-as-proximal assumption is not holding.","supporting_citations":[{"cited_title":"Fast model-based X-ray CT reconstruction using spatially nonhomogeneous ICD optimization,","cited_arxiv_id":null,"evidence_quote":"Fast MBIR algorithm that provides the reference reconstructions used as ground truth and the runtime baseline."},{"cited_title":"Deep learning based workflow for accelerated industrial X-ray Computed Tomography,","cited_arxiv_id":null,"evidence_quote":"Single-step DLMBIR baseline, the beam-hardening correction network, and the source of the paired FDK-to-MBIR training data."},{"cited_title":"Deep generalization of structured low-rank algorithms (Deep-SLR),","cited_arxiv_id":null,"evidence_quote":"Introduces the plug-and-play priors framework that justifies replacing the proximal map with a pre-trained network."},{"cited_title":"RARE: Image reconstruction using deep priors learned without groundtruth,","cited_arxiv_id":null,"evidence_quote":"Consensus-equilibrium rationale for why fixed regularization is standard, which the adaptive selection strategy departs from."},{"cited_title":"Plug- and-play unplugged: Optimization-free reconstruction using consensus equilibrium,","cited_arxiv_id":null,"evidence_quote":"BRISQUE no-reference image quality metric that drives the regularization parameter selection."}],"review_version":1}