{"id":"7e8158d4-e938-4af4-8a44-b8ba087fe29c","arxiv_id":"1909.00240","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A consensus-equilibrium framework that fuses data-domain and image-domain conditional GANs improves limited-angle CT reconstruction on a real security dataset.","lead":"This paper combines two deep-learning networks, one that completes missing CT projection data and one that cleans image artifacts, through a consensus-based framework. On a real 90-degree limited-angle security CT dataset, the combined method beats post-processing-only and data-completion-only baselines on RMSE, PSNR, and SSIM.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Four-iteration CE is not shown to reach consensus; without a fixed-point/convergence check, the DICE advantage over MBIR may be a stopping-rule artifact rather than a CE property.","rationale":"The reader's weakest assumption correctly identifies the missing convergence analysis for the 4-iteration CE solver. The paper itself flags this limitation in Section 3.3, and the central claim is specifically that DICE works through the CE framework, not merely as a fixed-4-iteration heuristic. I frame the concrete check as a fixed-point residual plus metric-stability test because that directly tests whether the iterates have reached consensus and whether the reported margins survive a change in stopping rule. The lack of error bars in Table 1 is also a concern, but it is secondary: even with statistical significance, the CE interpretation would still require the fixed-point check. Since the required verification is absent but obtainable, the appropriate verdict remains conditional, matching the reader's assessment.","tokens_in":12101,"tokens_out":5538,"duration_ms":49789,"concrete_test":"Run Algorithm 1 on the full test set for 4, 10, 20, 50, and 100 outer iterations with the published rho=0.25, mu1=0.6, mu2=0.4, sigma^2=1e-8, recording the normalized fixed-point residual ||z_k - T(z_k)||/||z_k|| and RMSE/PSNR/SSIM at each k. If the metrics at k=4 move by more than the DICE-vs-MBIR margins (3 HU, 0.41 dB, 0.03 SSIM) relative to a stabilized plateau, or if no plateau is reached, the reported DICE numbers are stopping-rule-dependent. Additionally, rerun at k=4 with rho in {0.1, 0.25, 0.5} and mu1 in {0.5, 0.6, 0.7}; if any variant drops below \"Initialization + MBIR\" on any metric, the claimed CE-based superiority is not robust to the advertised hyperparameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 fixes the Mann iteration (Eq. 9) at 4 outer iterations, stating only that this \"was observed to be adequate\" and deferring convergence analysis. The CE solution is defined as a fixed point of T=(2F-I)(2G-I); convergence of Mann iterations requires T to be nonexpansive (or similar), but Fimage is a trained cGAN and no such property is established. Consequently, the reported DICE image is not verified to satisfy the CE equations (1)-(3). The comparison \"Initialization + MBIR\" uses the same yconsistent and initialization, so the observed margin (RMSE 73 vs 76 HU, PSNR 26.55 vs 26.14 dB, SSIM 0.81 vs 0.78) could be produced by the image cGAN applied inside the first iterations, by the particular rho, mu1, mu2, or by the early stopping rule, rather than by convergence to a CE balance between the data and image agents. Since the central claim is that DICE \"using CE framework\" outperforms, the missing convergence/fixed-point check is load-bearing: if iterates oscillate or drift, the method is a fixed-4-iteration heuristic and the advantage may not persist under a different stopping rule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DICE, a consensus equilibrium (CE) framework that integrates a data-domain deep learning (cGAN-based sinogram completion) agent with an image-domain deep learning (cGAN-based image enhancement) agent for limited-angle CT reconstruction. The data agent combines the completed projection data with a CT physics model via an MBIR-type data fidelity term; the image agent is a learned patch-based prior. The two agents are combined through the CE fixed-point equations and solved with Mann iterations, limited to four outer iterations. Experiments on a real security CT dataset with 90-degree limited-angle data show that DICE achieves the lowest RMSE, highest PSNR, and highest SSIM among the considered methods (FBP, FBP+post-processing, data completion alone and with post-processing, and MBIR initialized with the same data).","tokens_in":12367,"tokens_out":6569,"duration_ms":63363,"significance":"If the central claim holds, the paper provides a practical way to combine data-domain and image-domain deep learning within a principled multi-agent reconstruction framework, with demonstrated gains on a real, challenging limited-angle security CT problem. The use of the same initialization and completed data for the DICE and MBIR baselines is a strength of the experimental design, and the inclusion of real CT data with multiple baseline methods adds value. However, the significance of the contribution is currently limited by two load-bearing gaps: the CE iteration is not shown to converge (the algorithm fixes the number of iterations at four), and the definition of the consistent projection data used by the data-fidelity agent is internally inconsistent between the method description and the algorithm. These issues must be resolved before the claim that DICE outperforms alternatives 'using CE framework' can be accepted. The reported gains over the closest baseline are modest, and no error bars or significance tests are provided, so the statistical robustness of the claim is also open.","major_comments":[{"comment":"The CE solution is defined as a fixed point of the map T=(2F-I)(2G-I), but the Mann iteration in Eq. (9) is stopped after exactly four outer iterations, with the text stating only that this 'was observed to be adequate' and deferring convergence analysis. Since the image agent F_image is a trained cGAN, no nonexpansiveness or other contraction property of T is established. The reported DICE result is therefore not verified to be a CE solution; it may be a fixed-budget heuristic whose advantage over 'Initialization + MBIR' depends on the stopping rule, the Mann parameter rho, and the weights mu1/mu2. The authors should report the evolution of the CE residual (e.g., ||z^(k+1)-z^(k)|| or the distance to the fixed-point equations) over iterations, run more than four iterations for at least a subset of test slices, and either establish convergence conditions or explicitly frame the contribution as a fixed-budget algorithm rather than a converged CE solution.","section":"Section 3.3 and Algorithm 1"},{"comment":"There is a direct inconsistency in the definition of the consistent projection data. Section 3.1 and Eq. (4) define y_consistent = A * FBP(y_complete), but Algorithm 1 sets y_consistent = A * x^(0), where x^(0) = phi_image(FBP(y_complete)) is the image-domain-enhancement output. If the implementation follows the algorithm, then the data-fidelity agent in Eq. (5) is fitting a quantity derived from the image enhancement, not from the data-domain completion as described. This weakens the anchor to the measured limited-angle data and could make the 'Initialization + MBIR' baseline circular, since it fits the same re-projected enhanced image. The authors must clarify which definition is actually used, correct the discrepancy, and if the algorithm version is intended, justify why the data consistency term should target a reprojection of the image-enhanced initialization.","section":"Algorithm 1, step 2 vs. Section 3.1"},{"comment":"The quantitative comparison reports only average RMSE, PSNR, and SSIM over the test slices, without error bars, standard deviations, or statistical significance tests. The margins over the closest baseline (Initialization + MBIR) are small (RMSE 73 vs 76 HU, PSNR 26.55 vs 26.14 dB, SSIM 0.81 vs 0.78). To support the claim that DICE 'outperforms all the considered methods on all three metrics,' the authors should report per-slice variability, paired differences, and ideally a significance test or confidence intervals.","section":"Table 1"},{"comment":"The ground truth used for both training the image-domain cGAN and for quantitative evaluation consists of full-view MBIR reconstructions. Since the image-domain network is trained to map low-quality images to these same MBIR targets, the reported metrics partly measure how closely each method reproduces the chosen MBIR reference rather than the true object. This is a limitation that should be discussed, and the evaluation would be strengthened by additional metrics or a comparison on data where a non-MBIR ground truth is available.","section":"Section 4 and Table 1"}],"minor_comments":[{"comment":"The text repeatedly renders the angle as '900' (e.g., 'challenging 900 limited-angle problem'); this should be '90°'.","section":"Abstract and Figure 1 captions"},{"comment":"The algorithm uses '← −' in several assignments (e.g., 'z(0)← −[x(0);x(0)]', 'k← −0'), which appears to be a formatting artifact; the minus signs should be removed if not intended.","section":"Algorithm 1"},{"comment":"There are minor typos such as 'transponsed' for 'transposed' in the architecture description; these should be corrected.","section":"Section 3.1"},{"comment":"The 'Initialization + MBIR' baseline is described as using '4 iterations of MBIR', but the MBIR objective, regularization, and stopping criterion are not specified; this information should be provided for reproducibility.","section":"Section 4.1"},{"comment":"The abstract and introduction mention 'efficient computation' as a benefit of the proposed approach, but no runtime or computational cost comparison is reported; adding this would substantiate the claim.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a useful application of CE to combine data- and image-domain deep learning, and the experimental setup with real security CT data is valuable. However, the missing convergence verification and the inconsistency in the definition of the consistent projection data are significant enough that the manuscript should not be accepted in its current form. If the algorithm actually uses y_consistent = A x^(0), the data consistency is effectively anchored to the image-domain output, which would be a serious weakness; if it is a typographical error, it must be corrected and the experiments re-run or re-described. The absence of error bars is also concerning given the small reported margins. These issues are addressable and do not seem to require new conceptual machinery, so major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this paper does what it says — it integrates a data-domain cGAN (sinogram completion) and an image-domain cGAN (post-processing) through the consensus equilibrium (CE) framework, and tests the result on a real 90-degree limited-angle security CT dataset. The reported numbers beat all comparators, including MBIR run on the same completed data and same initialization. The gains are small but consistent across metrics: RMSE 73 vs 76 HU, PSNR 26.55 vs 26.14 dB, SSIM 0.81 vs 0.78.\n\nWhat's genuinely new is the specific combination. The literature has sinogram completion, image post-processing, and end-to-end learning, but I don't see another paper that fuses both DL agents through CE. The authors are straightforward about the components and the implementation is reproducible in that sense. The comparison protocol is fair: the MBIR baseline uses the same consistent completed data and the same initialization, so the DICE advantage isn't from a better initialization.\n\nThe soft spots are real but addressable. Section 3.3 fixes the Mann outer iteration at 4, with \"observed to be adequate\" and future convergence analysis. A trained cGAN is not shown to be nonexpansive, so the 4-step output is not verified to be a fixed point of the CE map. That means the reported result is a fixed-iteration heuristic; the advantage over MBIR could depend on the stopping rule, the weights, and the initialization. A fixed-point residual plot or a sensitivity analysis across iteration counts would go a long way. Second, Table 1 has no error bars or significance tests; the margins are small enough that a few slices could shift the ranking. Third, the ground truth is full-view MBIR, which is also the training target for the image-domain cGAN — not fully circular, but it biases the comparison toward MBIR-like images. Finally, no code or data release, and metallic slices are excluded.\n\nNone of this is fatal. The paper is a solid engineering demonstration for security and NDT limited-angle CT. It deserves a serious referee; with a convergence check and uncertainty quantification it would be acceptable. If you're advising a student looking for a baseline to compare against, this is worth citing.","headline":"A competent, useful applied paper that fuses data- and image-domain cGANs via consensus equilibrium for limited-angle CT; the gains are modest and the CE iteration is unverified past 4 steps, but it deserves a real review.","tokens_in":12911,"tokens_out":2539,"would_cite":false,"duration_ms":22847,"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":"A consensus-equilibrium framework fuses data-domain and image-domain deep learning for 90-degree limited-angle CT, and on real security CT slices it beats all six methods compared.","keywords":["limited-angle CT","consensus equilibrium","deep learning reconstruction","conditional generative adversarial networks","sinogram completion","image post-processing","security CT","data consistency"],"falsifier":"Re-run DICE on the same 315-slice test set with 1, 2, 4, 8, and 16 outer CE iterations, holding the networks, weights, and initialization fixed; if the RMSE/PSNR/SSIM curves do not stabilize by 4 iterations, the reported advantage is a stopping-rule artifact rather than evidence of consensus.","tokens_in":11860,"feed_emoji":"🩻","tokens_out":12762,"duration_ms":173187,"temperature":0.7,"pith_summary":"This paper sets out to show that limited-angle CT reconstruction improves when deep learning is applied on both sides of the reconstruction: a conditional GAN first completes the missing projection data, a second conditional GAN then refines the reconstructed image, and a consensus-equilibrium (CE) mechanism holds the two agents together instead of training them end to end. The authors argue that image-only post-processing cannot recover structure lost by filtered backprojection, while data-only completion leaves residual artifacts, so the two should be fused with the CT physics model. On a real 90-degree limited-angle security CT test set, the proposed DICE approach achieves the best RMSE (73 HU), PSNR (26.55 dB), and SSIM (0.81) among the six methods compared. If the approach holds, it offers a template for combining learned priors with physical forward models in other limited-data tomographic settings.","feed_headline":"Dual-domain deep learning beats all rivals on 90-degree CT scans","feed_subtitle":"Real security CT slices: fused sinogram completion plus image refinement hits 73 HU RMSE.","key_machinery":"The central object is the consensus-equilibrium (CE) fixed-point formulation, which treats the CT data-consistency step and the trained image-enhancement network as two agents that must agree on a common image $x^*$. The data agent solves $F_{\\text{data}}(v_1) = \\arg\\min_{x\\ge 0} \\frac{1}{2}\\|\\hat{y}_{\\text{consistent}} - Ax\\|_W^2 + \\frac{1}{2\\sigma^2}\\|v_1-x\\|_2^2$, where $\\hat{y}_{\\text{consistent}}$ is the completed projection data re-projected through the CT system matrix $A$; the image agent is simply $F_{\\text{image}}(v_2)=\\varphi_{\\text{image}}(v_2)$, the trained cGAN. The CE equations require $F_{\\text{data}}(v_1^*)=x^*$, $F_{\\text{image}}(v_2^*)=x^*$, and $\\mu_1 u_1^*+\\mu_2 u_2^*=0$ with $x^*=\\mu_1 v_1^*+\\mu_2 v_2^*$, so the algorithm seeks a fixed point of $T=(2F-I)(2G-I)$ via Mann iterations. The paper uses four outer iterations, $\\rho=0.25$, $\\mu_1=0.6$, $\\mu_2=0.4$, and 20 conjugate-gradient steps inside the data agent.","core_discovery":"The central claim is that a consensus-equilibrium framework can integrate a data-domain cGAN that completes limited-angle sinograms and an image-domain cGAN that enhances patch-based image quality, producing reconstructions that outperform either network used alone. The data agent does not use the cGAN output directly: completed projections are re-projected through the CT forward model to enforce consistency, and the resulting consistent data is fed into a weighted data-fidelity term; the image agent maps images toward the learned prior. CE then finds an image that balances the two agents through a fixed-point iteration. On 315 slices from a real security CT dataset, DICE reports RMSE 73 HU, PSNR 26.55 dB, and SSIM 0.81, better than FBP, FBP plus post-processing, data completion followed by FBP with or without post-processing, and MBIR initialized with the same completed data. The authors present this as a demonstration that combining data-domain and image-domain deep learning through CE can outperform existing post-processing methods on a real limited-angle security problem.","pith_inferences":["A natural stress test beyond the reported experiments is to vary the CE outer-iteration count (1, 2, 4, 8, and 16) on the same test set; if quality does not stabilize by 4 iterations, the reported margin may be a stopping-rule artifact rather than evidence of consensus.","A likely extension, which the authors flag as ongoing work, is to make the data-domain network a third CE agent instead of a fixed front-end, letting completion errors be corrected during fusion rather than locked in before reconstruction.","A testable prediction is that the fusion gain will shrink or vanish if the consistency re-projection is removed or if the completed sinogram is corrupted, since the data agent's data-fidelity term depends entirely on $\\hat{y}_{\\text{consistent}}$."],"forward_implications":["If DICE's claim holds, limited-angle scans from non-rotating security gantries can be reconstructed with much lower artifacts than FBP or image post-processing alone, recovering both missing structure and CT-number uniformity.","The same two-agent CE scaffold can be reused with different learned data and image networks, so future improvements in either cGAN should carry directly into the fused reconstruction.","Data consistency is part of the winning recipe: the completed data entering the physics agent is the re-projected, consistent estimate $\\hat{y}_{\\text{consistent}}$, not the raw cGAN output, and the authors report this improves results.","Because the image-domain network is trained against full-view MBIR reconstructions, DICE's reported quality is measured relative to that MBIR reference rather than to an absolute truth.","The authors state the same framework can be applied to other limited-data problems such as electron microscopy, non-destructive evaluation, and medical imaging; if the mechanism generalizes, those fields gain a template for fusing learned priors with physics."],"supporting_citations":[{"why":"Defines consensus equilibrium and the fixed-point/Mann-iteration machinery that the DICE fusion algorithm is built on.","marker":"[9]"},{"why":"Supplies the two-agent CE equations and the averaging operator that the paper instantiates with data and image networks.","marker":"[50]"},{"why":"Provides the conditional GAN architecture and combined adversarial-plus-L2 objective used to train both the data and image domain networks.","marker":"[14]"},{"why":"Predecessor data-domain completion method that motivates the data agent's complete-then-reproject pipeline.","marker":"[38]"},{"why":"Contributes the data-domain cGAN architecture and discriminator training schedule reused in this paper.","marker":"[53]"},{"why":"Produces the full-view MBIR ground truth images and the 'Initialization + MBIR' baseline that DICE must beat.","marker":"[8]"},{"why":"Provides the real security CT dataset (Imatron C300 scans) used for training and evaluation.","marker":"[26]"},{"why":"Basis for treating a trained network as a prior or proximal operator inside iterative reconstruction, a key idea behind the image agent.","marker":"[43]"}],"fun_headline_variants":["Consensus equilibrium fuses dual cGANs for sharper limited-angle CT","DICE: dual-domain cGANs beat single-domain on 90-degree CT","Data+image cGANs reconciled by consensus equilibrium for limited-angle CT","Real security CT: consensus equilibrium combines cGANs to beat standard methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that four steps of the consensus-equilibrium iteration land close enough to a true consensus that the reported gains come from the balanced fusion itself, rather than from the chosen stopping point, weights, or initialization.","fun_headline_variants_meta":{"raw":{"variants":["Consensus equilibrium fuses dual cGANs for sharper limited-angle CT","DICE: dual-domain cGANs beat single-domain on 90-degree CT","Data+image cGANs reconciled by consensus equilibrium for limited-angle CT","Real security CT: consensus equilibrium combines cGANs to beat standard methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001249,"raw_usage":{"total_tokens":5169,"prompt_tokens":1042,"completion_tokens":4127,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":4046}},"tokens_in":658,"tokens_out":4127,"duration_ms":54570,"temperature":1.0,"reasoning_tokens":4046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:57:18.123478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run DICE on the same 315-slice test set with 1, 2, 4, 8, and 16 outer CE iterations, holding the networks, weights, and initialization fixed; if the RMSE/PSNR/SSIM curves do not stabilize by 4 iterations, the reported advantage is a stopping-rule artifact rather than evidence of consensus.","supporting_citations":[{"cited_title":"Plug-and-play unplugged: Optimization-free reconstruc- tion using consensus equilibrium,","cited_arxiv_id":null,"evidence_quote":"Defines consensus equilibrium and the fixed-point/Mann-iteration machinery that the DICE fusion algorithm is built on."},{"cited_title":"Consensus equilibrium: A framework for model inte- gration,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-agent CE equations and the averaging operator that the paper instantiates with data and image networks."},{"cited_title":"Image-to-Image Translation with Conditional Adversarial Networks,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional GAN architecture and combined adversarial-plus-L2 objective used to train both the data and image domain networks."},{"cited_title":"Deep Learning-Based Sinogram Completion for Low-Dose CT,","cited_arxiv_id":null,"evidence_quote":"Predecessor data-domain completion method that motivates the data agent's complete-then-reproject pipeline."},{"cited_title":"Fast Enhanced CT Metal Artifact Reduction using Data Do- main Deep Learning,","cited_arxiv_id":null,"evidence_quote":"Contributes the data-domain cGAN architecture and discriminator training schedule reused in this paper."},{"cited_title":"A Model-Based Image Reconstruction Algorithm With Simultaneous Beam Hardening Correction for X-Ray CT","cited_arxiv_id":null,"evidence_quote":"Produces the full-view MBIR ground truth images and the 'Initialization + MBIR' baseline that DICE must beat."},{"cited_title":"Advances in automatic target recognition (ATR) for CT based object detection system–Final report, Dept,","cited_arxiv_id":null,"evidence_quote":"Provides the real security CT dataset (Imatron C300 scans) used for training and evaluation."}],"review_version":1}