{"id":"0c70e14c-2866-47b9-a9da-b51975228889","arxiv_id":"1908.06792","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data consistent artifact reduction method that couples a U-Net prior with iterative reconstruction reduces RMSE by over 10% (noise-free) and over 24% (noisy) in simulated 120 degree cone-beam limited angle tomography.","lead":"The authors combine a deep learning network with a traditional iterative reconstruction method for limited angle CT scans, forcing the final image to match the measured X-ray data in the scanned range. Because the network's corrections can disagree with the actual measurements, this data consistency step repairs those errors and improves image quality by 10 to 24 percent over a pure deep learning approach.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The >10%/>24% RMSE gains rest on an untested e2 choice and no direct evaluation of the U-Net prior's accuracy in the missing angular range; a sensitivity sweep and projection-domain residual check would settle this.","rationale":"The paper's empirical design supports a genuine improvement: leave-one-out experiments show consistent RMSE reductions for all 17 patients in both regimes, and the example slices show visible artifact correction. I do not see a fatal flaw. The most load-bearing gap is that the key parameter e2 and the prior's accuracy in the unmeasured angular range are not interrogated. The reader identified the prior-accuracy assumption; I refine it by pointing out that the paper's own e2=0.5 choice could either mask or amplify that assumption, and no sensitivity study exists. A second issue is the single noise realization, which matters for the noisy-case claim. Neither issue invalidates the method; they make the headline percentages conditional on an untested setting. The proposed sweep and residual check would settle whether the gain is robust. Therefore the reader's CONDITIONAL verdict stands, and no further adjustment is needed.","tokens_in":9139,"tokens_out":12377,"duration_ms":130384,"concrete_test":"Run a leave-one-out fold with e2 swept over {0.01, 0.05, 0.1, 0.5, 1.0, 2.0, 5.0} in Eq. (3), holding all other parameters fixed, and record the mean RMSE for both the noise-free and noisy cases. Also compute the projection-domain residual ||Au f_U-Net - Au f_reference|| over the unmeasured angular range for each test patient. The concern is resolved if the reported >10% and >24% improvements persist across at least a factor of 5 in e2, and if DCAR remains beneficial even when the prior's missing-range error is large.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DCAR reduces RMSE over the U-Net baseline by more than 10% in the noise-free case and more than 24% in the noisy case. The mechanism depends on Eq. (3), where the unmeasured projections are replaced by the forward projection of the U-Net prior and the tolerance e2 controls how much the final image may deviate from that prior. The paper sets e2=0.5 empirically with no sensitivity analysis and never measures how accurate the prior actually is in the unmeasured angular range. If the prior's missing-angle projections are substantially wrong and e2 is tighter than that error, the final reconstruction inherits the U-Net's dark holes and smoothed structures. If e2 is very loose, the unmeasured constraint is effectively inactive and the method reduces to SART+wTV initialized with f_U-Net, so the attribution of the gain to the proposed constraint is unclear. In the noisy case, only one Poisson noise realization is used and no error bars are reported, so the 24% figure could be specific to one noise draw. These issues are addressable, but until e2 sensitivity and the missing-range residual are reported, the quantified claim is not yet robust.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data consistent artifact reduction (DCAR) method for limited angle cone-beam CT. A U-Net first predicts an artifact-reduced prior image from an FBP reconstruction; then an iterative SART+wTV reconstruction is run using the U-Net output as initialization, with data consistency enforced in the measured angular range and an additional soft-thresholded constraint that keeps the forward projection of the reconstruction close to that of the U-Net prior in the unmeasured angular range. The method is validated on 17 patients from the AAPM Low-Dose CT Grand Challenge with 120° limited angle data, using leave-one-out cross-validation. The reported mean RMSE improvements over the U-Net baseline are more than 10% in the noise-free case and more than 24% in the case with simulated Poisson noise.","tokens_in":9363,"tokens_out":5388,"duration_ms":50048,"significance":"If the results hold, DCAR is a practically valuable and conceptually clean way to combine a deep learning prior with analytic data consistency constraints, directly addressing the known robustness limitations of pure deep learning reconstructions. The paper has clear strengths: it uses a public dataset, a patient-level leave-one-out protocol, and a straightforward formulation that should be reproducible. The central idea of constraining the reconstruction in the unmeasured projection domain is plausible and interesting. However, the quantitative claims currently rest on an empirically chosen tolerance parameter without a sensitivity analysis, the noisy experiments use a single noise realization with no error bars or significance tests, and no ablation isolates the contribution of the proposed unmeasured-range constraint from the effect of better initialization. These issues are addressable and do not undermine the core concept, but they must be resolved before the stated improvement percentages can be considered robust.","major_comments":[{"comment":"The central quantitative claim (more than 10% RMSE reduction noise-free and more than 24% in the noisy case) depends critically on the tolerance e2 in Eq. (3), which controls how much the final image may deviate from the forward projection of the U-Net prior in the unmeasured angular range. The manuscript states that e2=0.5 is chosen empirically, but it provides no sensitivity analysis and no direct evaluation of how accurately Au f_U-Net approximates the true unmeasured projections. If the prior is inaccurate in the missing angular range and e2 is too tight, the reconstruction inherits those errors; if e2 is too loose, the constraint is inactive and the method reduces to SART+wTV initialized with f_U-Net. Please report a sweep over e2 (including e2 near zero and e2 effectively infinite), a projection-domain residual analysis in the unmeasured angular range, and the aggregate mean RMSE improvement across the 17 patients.","section":"Section 2.2, Eq. (3) and Section 3"},{"comment":"The experimental comparison does not isolate the contribution of the proposed unmeasured-range constraint. The paper compares DCAR (50 iterations, initialized with f_U-Net, with the Au constraint) against SART+wTV from zero initialization (100 iterations), but not against SART+wTV initialized with f_U-Net and without the Au constraint. Without this ablation, the reported improvement cannot be attributed specifically to the constraint in Eq. (3); it may be explained by the improved initialization alone combined with wTV regularization. Please add an ablation that uses f_U-Net initialization and the same 50-iteration SART+wTV protocol but omits or loosens the e2 constraint.","section":"Section 2.3 and Eq. (6)"},{"comment":"The noisy-case evaluation is based on a single Poisson noise realization per patient, and no standard deviations, confidence intervals, or significance tests are reported. The claim of robustness to Poisson noise would be substantially strengthened by repeating the noisy experiment with multiple independent noise realizations and reporting mean ± standard deviation per patient or a paired statistical test (e.g., Wilcoxon signed-rank test) across the 17 patients. The current figures show only point estimates of mean RMSE per patient, which is not sufficient to support the 'significant' improvement claimed in the abstract.","section":"Section 3, Figs. 4 and 5"}],"minor_comments":[{"comment":"Discussion states 'only 13 slices are chosen from each patient' but Section 2.3 says 25 slices per patient are used for training; please correct this inconsistency.","section":"Section 4"},{"comment":"The soft-thresholding operator S_tau is used in the SART update but is never explicitly defined; please state its definition and clarify the units of e1 and e2 (e.g., HU in projection domain) so the chosen values are meaningful to readers.","section":"Eq. (6)"},{"comment":"In the caption of Fig. 3, 'DACR' should be 'DCAR'; also consider reporting the average relative improvement across all patients in the text in addition to the per-patient labels.","section":"Figure 3 caption"},{"comment":"There is a typo in the Introduction: 'consitent' should be 'consistent'.","section":"Section 1"},{"comment":"Please justify the different iteration counts for wTV (100 iterations) and DCAR (50 iterations) in the comparison; if the iteration count changes the outcome materially, report results at matched iteration numbers or explain why the chosen counts are appropriate.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid MICCAI-style contribution with a genuinely useful idea, and the leave-one-out protocol on a public dataset is a real strength. The main concerns are the lack of sensitivity analysis for the key tolerance parameter, the absence of statistical confidence measures, and the missing ablation that would attribute the gain to the proposed constraint. These are all fixable with additional experiments and should be requested. I would not reject on the basis of the current evidence, but the quantitative claims should not be accepted as stated until the robustness questions are answered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes DCAR, a way to combine a U-Net prior with a conventional iterative reconstruction for limited angle CT. The new piece is using the forward projection of the U-Net image onto the unmeasured angular range as a soft constraint (Eq. 3), while keeping consistency with measured projections. That is a sensible and not-obvious design, and it goes beyond the metal-artifact work it cites.\n\nWhat the paper does well: leave-one-out evaluation on 17 patients from the AAPM challenge. Every held-out patient shows RMSE reduction over the U-Net baseline, 10.9–24.7% noise-free and 24–32% with Poisson noise. The visual examples support the numbers. This is real evidence, not a single cherry-picked slice.\n\nSoft spots: The headline percentages rest on e2=0.5, picked empirically, with no sensitivity analysis. If e2 were much larger, the method would reduce to SART+wTV initialized with the U-Net, and the gain attribution would blur. Also, the paper never checks how accurate the U-Net prior actually is in the missing angular range. The mechanism depends on that. The noisy experiment uses one noise realization and no error bars, so the 24% figure is not yet robust. Standard deviations across patients would cost little to add. Code is not released.\n\nNone of these are fatal. The central claim—DCAR consistently improves on a U-Net prior alone—is credible from the figures. The missing experiments are easy to specify and would settle the main concern.\n\nWho this is for: researchers working on learned+model-based CT reconstruction, especially limited angle. It deserves a serious referee; with a sensitivity sweep, a missing-range residual check, and per-patient variance, the quantified claims would be much stronger.\n\nSend it to peer review.","headline":"DCAR is a credible, well-evaluated combination of a U-Net prior with data consistency for limited-angle CT; the headline RMSE gains are plausible, but the e2 choice and missing-range prior accuracy need a sensitivity check.","tokens_in":9938,"tokens_out":1516,"would_cite":true,"duration_ms":14780,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65J22","92C55"],"pacs":["87.57.Q-","87.57.nf"],"model":"deepseek-v4-flash","headline":"Hybrid reconstruction cuts limited-angle CT error by forcing data consistency","keywords":["limited angle tomography","data consistency","deep learning prior","U-Net","artifact reduction","Poisson noise","iterative reconstruction","cone-beam CT"],"falsifier":"Run a limited-angle reconstruction with a deliberately corrupted U-Net prior (for example, a prior containing a false dark hole in the unmeasured angular range) and check whether DCAR reproduces the false structure; if it does, the unmeasured-range constraint is not actually preventing the prior's bad influence, contradicting the paper's claim that inaccuracies lie only in unmeasured areas.","tokens_in":1600,"feed_emoji":"🩻","tokens_out":3480,"duration_ms":44221,"temperature":0.7,"pith_summary":"This paper argues that purely deep-learning reconstructions for limited-angle CT are unreliable because they ignore the measured projection data and are sensitive to noise. To fix this, the authors propose DCAR, a method that uses a U-Net's output only as a prior image and then applies an iterative reconstruction that enforces agreement with the measured projections while gently steering the unmeasured angular range toward the U-Net prior. In 120-degree cone-beam experiments on patient data, DCAR reduces root-mean-square error by more than 10 percent in noise-free cases and more than 24 percent in Poisson-noise cases compared with the U-Net alone. The central point is that deep learning should complement, not replace, the physics-based reconstruction constraint.","feed_headline":"Cut limited-angle CT error by forcing data consistency","feed_subtitle":"A U-Net prior plus measured-ray constraints beats pure deep learning by 10-24% RMSE.","key_machinery":"The central object is the constrained optimization problem with two data-fidelity terms: one for the measured angular range (using $A_m$ and $p_m$) and one for the unmeasured angular range (using $A_u$ and the forward projection of the U-Net prior $f_{\\text{U-Net}}$). The mechanism carries the argument by allowing the final image to be initialized as the U-Net output and then iteratively corrected so that it agrees with measured data; a soft-thresholding operator $S_\\tau$ handles error tolerance, and iterative reweighted total variation regularizes the result. The work this does is to confine the deep learning prior's influence to the unmeasured angles, where no true data exist, while measured rays directly correct any inconsistency the U-Net introduced.","core_discovery":"The paper establishes that a deep-learning reconstruction that is not consistent with the measured projection data can be substantially improved by treating it as a prior inside a constrained iterative reconstruction. The authors define a reconstruction $f$ that must satisfy $\\|A_m f - p_m\\| < e_1$ for measured projections and $\\|A_u f - A_u f_{\\text{U-Net}}\\| < e_2$ for the unmeasured angular range, initialized with $f_{\\text{U-Net}}$. Solving this with SART plus iterative reweighted total variation yields images that have both the anatomical plausibility learned by the U-Net and the fidelity to actual measurements that the U-Net lacks. On leave-one-out cross-validation with 17 patient datasets, the mean RMSE improvement over the U-Net is more than 10 percent in the noise-free setting and more than 24 percent with Poisson noise, with the largest gains occurring where the U-Net produces dark holes or smoothed-out cavities.","pith_inferences":["The paper's stated improvements are likely conservative: the U-Net is trained on only 400 slices, so a stronger prior could push DCAR gains even higher while keeping measured-data fidelity.","A natural testable extension not explored here is replacing the U-Net prior with an ensemble or a network trained on a larger dataset, then measuring whether the $e_2$ tolerance can be reduced to tighten the unmeasured-range constraint further.","The choice of $e_2=0.5$ is empirical; if the U-Net prior were known to be accurate in the missing angular range, a smaller $e_2$ would theoretically improve resolution, implying that adaptive tuning of $e_2$ per projection could be worthwhile.","DCAR's reliance on the U-Net prior to estimate unmeasured projections means its ceiling is bounded by the prior's accuracy; a principled uncertainty estimate for the prior could allow the method to automatically relax $e_2$ where the prior is unreliable."],"forward_implications":["DCAR-type hybrid reconstruction could make deep learning reconstructions viable in clinical limited-angle CT settings where data consistency is mandatory.","The method's noise robustness suggests it can be used with lower X-ray doses, since Poisson noise is suppressed by the iterative data-consistency framework more effectively than by the U-Net alone.","Because the U-Net prior only fills unmeasured angles, the approach should generalize better to unseen anatomies than direct image-to-image prediction, as the measured data remain the dominant constraint.","The same principle of using a learned prior inside a constrained reconstruction could be transferred to other ill-posed inverse problems, such as sparse-view CT or metal artifact reduction, wherever a measured subset of the data is reliable."],"supporting_citations":[{"why":"Establishes that the U-Net method is not robust to Poisson noise, motivating the need for data consistency.","marker":"[19]"},{"why":"Supplies the U-Net architecture that is modified and used to generate the prior image $f_{\\text{U-Net}}$.","marker":"[22]"},{"why":"Supplies the SART + iterative reweighted total variation algorithm used to solve the constrained optimization.","marker":"[15]"},{"why":"Provides the 17-patient Low-Dose CT Grand Challenge dataset used for training and leave-one-out validation.","marker":"[32]"},{"why":"Demonstrates the multi-scale wavelet U-Net approach for limited-angle CT, which is the state-of-the-art baseline the paper compares against.","marker":"[18]"},{"why":"Shows the precedent of using deep learning predictions as prior images for iterative metal artifact reduction, the category DCAR extends.","marker":"[10]"},{"why":"Presents a hybrid deep learning-shearlet framework that also separates visible and invisible singularities, contextualizing DCAR's hybrid strategy.","marker":"[21]"}],"fun_headline_variants":["Data-consistent prior cuts CT RMSE by up to 24%","U-Net prior plus ray consistency beats pure deep learning","Limited-angle CT: Deep prior plus data fidelity wins","Enforce measured rays to improve deep learning CT","Hybrid method reduces limited-angle CT error 24%"],"cache_read_input_tokens":12032,"weakest_assumption_plain":"The method assumes that the U-Net's forward projection onto the unmeasured angular range is a reliable estimate of the true missing projections; if the U-Net hallucinates or smooths away structures in those angles, the final DCAR image inherits those errors because no measured data can correct them.","fun_headline_variants_meta":{"raw":{"variants":["Data-consistent prior cuts CT RMSE by up to 24%","U-Net prior plus ray consistency beats pure deep learning","Limited-angle CT: Deep prior plus data fidelity wins","Enforce measured rays to improve deep learning CT","Hybrid method reduces limited-angle CT error 24%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00034,"raw_usage":{"total_tokens":1877,"prompt_tokens":952,"completion_tokens":925,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":844}},"tokens_in":568,"tokens_out":925,"duration_ms":10647,"temperature":1.0,"reasoning_tokens":844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:34:35.404307+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a limited-angle reconstruction with a deliberately corrupted U-Net prior (for example, a prior containing a false dark hole in the unmeasured angular range) and check whether DCAR reproduces the false structure; if it does, the unmeasured-range constraint is not actually preventing the prior's bad influence, contradicting the paper's claim that inaccuracies lie only in unmeasured areas.","supporting_citations":[{"cited_title":"Some investigations on robustness of deep learning in limited angle tomography,","cited_arxiv_id":null,"evidence_quote":"Establishes that the U-Net method is not robust to Poisson noise, motivating the need for data consistency."},{"cited_title":"U-Net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net architecture that is modified and used to generate the prior image $f_{\\text{U-Net}}$."},{"cited_title":"Scale- space anisotropic total variation for limited angle tomography,","cited_arxiv_id":null,"evidence_quote":"Supplies the SART + iterative reweighted total variation algorithm used to solve the constrained optimization."},{"cited_title":"Low-dose CT for the detection and classification of metastatic liver lesions: Results of the 2016 low dose CT grand challenge,","cited_arxiv_id":null,"evidence_quote":"Provides the 17-patient Low-Dose CT Grand Challenge dataset used for training and leave-one-out validation."},{"cited_title":"Multi-scale wavelet domain residual learning for limited-angle CT reconstruction,","cited_arxiv_id":null,"evidence_quote":"Demonstrates the multi-scale wavelet U-Net approach for limited-angle CT, which is the state-of-the-art baseline the paper compares against."},{"cited_title":"Convolutional neural network based metal artifact reduction in X-ray computed tomography,","cited_arxiv_id":null,"evidence_quote":"Shows the precedent of using deep learning predictions as prior images for iterative metal artifact reduction, the category DCAR extends."},{"cited_title":"Learning the invisible: a hybrid deep learning-shearlet framework for lim- ited angle computed tomography,","cited_arxiv_id":null,"evidence_quote":"Presents a hybrid deep learning-shearlet framework that also separates visible and invisible singularities, contextualizing DCAR's hybrid strategy."}],"review_version":1}