{"id":"721c8fd3-9b21-4ae1-89ff-8d85b3a27afa","arxiv_id":"2509.06082","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"MIP RO, a mixed-integer post-processing step driven by a DNN that approximates the Sobel edge filter, sharpens edges and homogenizes intensities in CSHM tomographic reconstructions, at the cost of longer runtime and worse raw-data fit.","lead":"This paper trains a small neural network to detect edges and embeds it into a mixed-integer optimization program that post-processes tomographic images, yielding sharper interfaces and more uniform material intensities. The method is tested on a simulated phantom and on real electron and X-ray nano-tomography data, with clear quantitative gains on the phantom but only qualitative gains on the real data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the DNN's contribution: MIP RO's gains may come from the hand-crafted binarization prior, not learned edge information.","rationale":"The reader's weakest assumption focused on DNN approximation accuracy and threshold transferability. While those are relevant, the more fundamental issue is that the paper never tests whether the DNN component is necessary. The method's objective (14) bundles three terms: the DNN edge term, a hand-crafted binary-pushing term (12), and a fidelity term (13). The phantom results show large RME improvements, but the Dev term alone could plausibly produce most of the binarization and noise suppression. Without an ablation, the central claim that 'integrating neural networks' is what drives the improvement is not established. This is a standard control that should be present. The reader's concern about DNN accuracy is secondary because even an inaccurate DNN could be overridden by the Dev/L terms; conversely, an accurate DNN may be irrelevant. Therefore, I disagree that the weakest assumption is DNN accuracy; it is the missing attribution of the improvement to the learned component. This does not change the verdict: the paper remains CONDITIONAL, pending an ablation. The code is publicly available, so such a test is feasible. I give no weight to ad hominem and acknowledge the phantom experiments as real quantitative support, but the missing control is a genuine gap.","tokens_in":23761,"tokens_out":5047,"duration_ms":54317,"concrete_test":"Run MIP RO on the simulated phantom with the Figure 9 setting (512² pixels, 20 projections, T=800, α=β=1/50), but remove the DNN term from objective (14), solving fa = argmax −α Dev(fa) − β L(fa) with the same subregion spacing and averaging as Algorithm 1. Compute RME and BMS. If the no-DNN variant yields RME within ~0.002 of the full method's 0.008 and BMS above ~0.99, the DNN term is not essential. If RME degrades substantially (e.g., >0.02), the learned edge term is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MIP RO improves CSHM reconstructions by integrating a DNN that approximates Sobel edge intensities (Sections 3.4–3.6). However, the experiments in Section 5 compare full MIP RO only against external baselines (CSHM, SIRT, CS, TVR-DART). No ablation removes the DNN term (MIPa, Eq. 14) while keeping the same optimization framework. The objective (14) comprises three terms: the DNN-based edge term, the hand-crafted material deviation term (Dev, Eq. 12), and the pixel fidelity term (L, Eq. 13). Dev alone pushes each subregion toward binary values (0 or ω), and L keeps the solution near the CSHM reconstruction. It is plausible that a large part of the RME reduction (phantom: 0.042→0.017 in Fig. 8; 0.060→0.008 in Fig. 9) and BMS increase stems from the hand-crafted binarization prior, not from learned edge information. If a no-DNN variant (optimizing only -α Dev - β L) achieves comparable RME/BMS, then the paper's claim that 'integrating neural networks' drives the improvement is unsupported; the method would reduce to non-learning post-processing. This is load-bearing because the novelty and title rest on the learned edge model.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-processing reconstruction framework for tomographic images of homogeneous materials with sharp edges. A ReLU neural network is trained to approximate the Sobel edge intensity on 3×3 subregions, and the trained network is encoded as mixed-integer linear constraints following Fischetti and Jo (2018). This DNN-MIP block is then combined with the CSHM reconstruction model in two ways: a joint integrated model and a cheaper per-subregion 'MIP RO' re-optimization that balances the learned Sobel prediction, a hand-crafted binary material-prior term, and a pixel-loss term relative to the CSHM solution. Experiments compare MIP RO against SIRT, CS, TVR-DART, and CSHM on a simulated phantom (with and without missing wedge), a zeolite ET dataset, and a copper nano-CT dataset. The paper reports lower RME and higher BMS on the phantom, and sharper interfaces/higher BMS on the experimental datasets, at the cost of longer runtimes and worse raw-data coverage.","tokens_in":24113,"tokens_out":8734,"duration_ms":99175,"significance":"If the central claim is valid, the approach is a useful contribution to model-based/learning hybrid tomography: it embeds a learned edge operator into an exact MIP framework, and it is openly implemented with code and data made public. The phantom results are self-consistent and show a substantial reduction in RME (0.042→0.017 at 5 projections, 0.060→0.008 at 20 projections) over CSHM. However, the novelty claim specifically rests on the learned DNN edge term, and the current evidence does not isolate that term from the hand-crafted binarization prior. Moreover, the ReLU encoding shown in Eqs. (7)–(8) is incomplete, and the experimental quality metrics for the no-ground-truth datasets are partly circular. These issues are load-bearing, so the manuscript needs substantial revision before the central claim can be accepted.","major_comments":[{"comment":"The claimed ReLU MIP encoding is incomplete. The implications z=1⇒x≥0 and z=0⇒s≥0 are already consequences of x,s≥0 and do not enforce the complementarity x·s=0 that defines ReLU(y)=x for y=x−s. Without additional big-M bounds such as x≤U(1−z) and s≤Uz, the feasible set contains non-ReLU solutions, so (DNN) does not represent the trained network. Since every subsequent model inherits (DNN), the formulation must be corrected or the implicit finite bounds must be stated.","section":"Section 3.4, Eqs. (7)–(8)"},{"comment":"The simplification of αDev+βL for α=β=1 uses the maximum DNN/Sobel value \\bar u from Section 3.7.1 instead of the material density ω from Eq. (12). With Dev(f)=(ω−f)f, the affine coefficient is (ω−2f*), not (\\bar u−2f*). This appears to be a typo, but since the linearity claim for the α=β case is used to explain runtimes and parameter behavior, the displayed identity should be corrected.","section":"Section 3.7.2, Eq. (14)"},{"comment":"No ablation isolates the learned DNN contribution. The MIP RO objective combines the DNN edge term MIPa, the hand-crafted binarization term Dev(a), and the pixel-fidelity term L(a). A variant optimizing only −αDev(a)−βL(a) should be compared against full MIP RO; otherwise the reported RME/BMS gains could be entirely due to the hand-crafted binary prior rather than to 'integrating neural networks' as claimed in the abstract and title. This is load-bearing for the paper's central claim.","section":"Section 5, Eq. (14)"},{"comment":"For the experimental datasets there is no ground truth, and the principal quantitative metric BMS (Eq. 18) is directly encouraged by the Dev term (Eq. 12), making the sharpness/homogeneity gains partly circular. Meanwhile RDC for MIP RO is worse than CSHM in both experimental cases (zeolite: 0.090 vs 0.028; Cu: 0.126 vs 0.058). Please add an independent metric (e.g., edge localization error, segmentation agreement, or comparison against a high-projection reference) to support the claim of superior experimental reconstructions.","section":"Sections 5.4–5.5"},{"comment":"The method assumes the DNN approximates the Sobel operator sufficiently well on tomography subregions and that the threshold T=800 transfers across datasets. No test-set approximation error of the DNN against the true Sobel operator is reported, and T is tuned on the zeolite dataset (Figure 12) but applied to the phantom and Cu datasets. Please report DNN accuracy on a held-out tomography-like set and provide per-dataset threshold sensitivity.","section":"Sections 4.3 and 3.7.1"}],"minor_comments":[{"comment":"'642 pixels' should read '64² pixels'.","section":"Figure 7 caption"},{"comment":"The caption lists results as (a)–(d) for the simulated dataset and (e)–(h) for the zeolite dataset, but the figure contains panels (a)–(j). The caption needs to be aligned with the figure panels.","section":"Figure 17 caption"},{"comment":"The discussion of threshold T is internally inconsistent: low T is first said to make subregions more likely to be classified as edges, but then said to 'suppress weaker edges', and higher T is said to detect weaker edges. This is likely a wording error and should be clarified.","section":"Section 5.4.1"},{"comment":"The subregion spacing s∈{1,3} is never reported in the computational experiments. The results may depend on whether overlapping or non-overlapping subregions are used; please state the value of s for every experiment.","section":"Algorithm 1"},{"comment":"The integrated model is stopped at 15% optimality gap, but the caption does not state whether the displayed image corresponds to the incumbent or the best proven solution; this should be clarified.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal, and the phantom RME results are encouraging. However, the title and abstract claim that neural-network integration drives the improvement, while the experiments do not ablate the DNN term. If the authors show that the DNN contribution matters beyond Dev+L, the paper could be a solid contribution; otherwise the novelty claim should be substantially revised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The practical result is real: MIP RO post-processing measurably improves on CSHM for the phantom (RME roughly halved) and produces noticeably cleaner interfaces on two experimental datasets. The code and data are public. That is the honest core, and it is worth taking seriously.\n\nWhat is actually new: the specific architecture—training a tiny ReLU net to mimic Sobel responses on 3x3 patches, embedding it as MIP constraints via Fischetti–Jo, then re-optimizing each window of a CSHM reconstruction—is not in the cited literature, as far as I can tell. The sliding-window design with averaging of overlaps is sensible, and the linear-objective special case alpha=beta is a nice practical touch that gives a real speedup.\n\nThe biggest soft spot is exactly what the stress-test note says: no ablation. The objective (14) bundles the DNN edge term with the hand-crafted Dev term that pushes pixels to 0 or omega. On the phantom, with alpha=beta, the Dev/L terms alone may explain most of the RME drop and BMS increase. Without removing only the DNN term, the paper's title claim—that neural-network integration drives the improvement—is not established. This is load-bearing and easily fixable, but the current experiments do not rule out non-learning post-processing.\n\nSecond, the math has a few warts. The ReLU MIP encoding (7)–(8) is incomplete without explicit big-M bounds linking x, s, z; the implication constraints need finite bounds to be an exact encoding. Section 3.7.2's derivation writes bar-u (the max Sobel value) where it should be omega (the material density); the algebra happens to be affine either way, but as written it is wrong and confusing. Also, the subregion spacing s in the experiments is never stated—that is a reproducibility gap.\n\nMinor: the DNN is trained on one flower photo and T=800 is selected on the zeolite dataset, so the transfer claims are thin. To the paper's credit, it acknowledges the RDC tradeoff openly and discusses when grayscale variation is preferable to forced binarization.\n\nWho this is for: experimentalists in nano/micro-CT looking for a sharpening post-processor, and MIP folks curious about a clean embedding of a learned operator into a solver. The paper deserves a serious referee, but it needs an ablation and some mathematical cleanup before the central claim is convincing. My recommendation: send it to peer review with expectations of major revision—add a no-DNN ablation, fix the MIP bounds and the derivation, report the spacing. If the ablation shows the learned edge term contributes little, the paper is still publishable, but as a post-processing study rather than a 'neural network integration' method.","headline":"Solid engineering story with real phantom gains, but the paper never isolates the learned edge model from the hand-crafted binarization prior—an ablation is needed before the central claim holds.","tokens_in":24633,"tokens_out":2394,"would_cite":false,"duration_ms":26693,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","90C11","68T07","65R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural edge prior halves sparse-scan tomography error","keywords":["tomography","mixed-integer programming","neural networks","edge detection","Sobel operator","compressed sensing","homogeneous materials","image reconstruction"],"falsifier":"Measure the trained DNN's output against the exact Sobel response on 3×3 patches sampled from the real zeolite and copper projection data; if the mean absolute error is of the same order as the difference between the threshold T=800 and the Sobel values it separates, the edge decisions are not trustworthy. Then re-run MIP RO with T and ω re-estimated on each dataset: if the RME/BMS superiority over CSHM disappears, much of the reported gain is parameter transfer rather than the method itself.","tokens_in":23644,"feed_emoji":"🔬","tokens_out":6451,"duration_ms":65178,"temperature":0.7,"pith_summary":"This paper tries to show that a learned edge detector can be turned into exact mathematical constraints and used to sharpen tomographic reconstructions of homogeneous materials. The authors train a small neural network to reproduce Sobel edge intensity on 3x3 subimages, then encode the trained network as a mixed-integer program using a standard ReLU-binary-variable construction. That MIP is added to a compressed-sensing reconstruction designed for homogeneous materials (CSHM), first as an integrated model and then as a cheaper per-window re-optimization called MIP RO. On a simulated phantom with only 5 projections, MIP RO lowers relative mean error from 0.042 (CSHM) to 0.017 and raises the bimodal contrast score from 0.974 to 0.998; on real zeolite and copper datasets it visibly sharpens interfaces and homogenizes material intensity. The point is not speed — runtimes grow — but fidelity: for homogeneous, sharp-edged specimens, a data-driven edge prior can be embedded in an exact optimizer to outperform heuristic and compressed-sensing baselines.","feed_headline":"Neural edge prior halves sparse-scan tomography error","feed_subtitle":"A learned Sobel detector becomes mixed-integer constraints that sharpen reconstructions of homogeneous materials.","key_machinery":"The load-bearing object is the DNN-as-MIP block: a trained ReLU network that approximates the Sobel edge response on a 3x3 patch, converted into linear constraints plus binary variables by the standard Fischetti–Jo formulation, so the optimizer can reason about edges exactly. Around it sits the MIP RO window objective: for each patch, a binary edge variable e_a selects max(y_a, T - y_a), with α and β weights balancing material-density deviation against fidelity to the base CSHM image. Setting α=β makes the objective affine in the pixels, so each small window is a binary linear program solvable to global optimality; the sliding-window algorithm averages overlapping patch solutions.","core_discovery":"The central claim is that MIP RO — a DNN-guided mixed-integer re-optimization of an already good CSHM reconstruction — systematically improves reconstruction fidelity for homogeneous materials with sharp edges. Each 3x3 subregion is re-solved independently: a binary variable decides whether the region is an edge or a uniform area, with threshold T controlling the decision; the objective rewards approximated Sobel values either above T or below T, while a quadratic loss keeps the new pixels close to the rescaled CSHM image and a deviation term pushes values toward 0 or the known material density ω. Because the DNN is encoded as linear MIP constraints, the whole decision can be solved to globa","pith_inferences":["The most direct test the paper leaves open is measuring the trained network's approximation error against the exact Sobel operator on the actual tomography patches; if the DNN is not faithful where it matters, the threshold T=800 may be silently misclassifying edges.","The same encoding could be applied to other imaging priors — Hessian, structure tensor, learned segmentation scores — turning them into constraints for any discrete ill-posed inverse problem, not just homogeneous-material tomography.","Because the threshold T was tuned on the zeolite data and reused elsewhere, the method's apparent generality may be optimistic; re-estimating T per dataset would show how much of the gain depends on that hand-set value.","The RME improvements on the phantom are large enough that end-to-end learning methods could benchmark against this exact-optimization hybrid; conversely, replacing Sobel with a learned edge detector trained on tomographic data might transfer better to multi-material specimens."],"forward_implications":["For simulated sparse-view data, MIP RO beats SIRT, CS, TVR-DART, and CSHM on relative mean error and bimodal contrast in every tested setting, including a 60° missing wedge.","On experimental zeolite and copper datasets, the method yields sharper interfaces and more uniform material intensity than all benchmarks, at the price of higher raw-data-coverage values and longer runtimes.","Choosing α=β makes each subproblem a linear binary program, restoring computational tractability and a nearly binary image; choosing α<β preserves grayscale variation needed when raw data contain inconsistencies.","Because each 3×3 subproblem is independent, the entire re-optimization can be parallelized, so the practical runtime gap versus CSHM could shrink substantially.","The scheme is not limited to tomography: any data-learned knowledge expressible as a ReLU network can be inserted into a mathematical optimization model this way."],"supporting_citations":[{"why":"Supplies the CSHM baseline model, the experimental zeolite and copper datasets, and the base reconstructions that MIP RO post-processes.","marker":"Kreuz et al. (2024)"},{"why":"Provides the mixed-integer linear formulation of ReLU networks used to encode the trained edge-detection DNN as exact constraints.","marker":"Fischetti and Jo (2018)"},{"why":"Justifies the Sobel operator as the edge-intensity target the DNN is trained to approximate.","marker":"Gonzalez and Woods (2017)"},{"why":"Gives the compressed-sensing approach that, together with its homogeneous-material extension, is the benchmark and starting point for the optimization models.","marker":"Candès et al. (2006)"},{"why":"Supplies the TVR-DART discrete-tomography benchmark against which the new method is compared.","marker":"Zhuge et al. (2015)"},{"why":"Supplies the SIRT iterative reconstruction benchmark used as another comparison baseline.","marker":"Gilbert (1972)"}],"fun_headline_variants":["Neural edge detection sharpens tomography reconstructions","Blend of DNN and optimization eliminates edge blur","Learned edges guide math solver for crisper scans","MIP-encoded neural priors improve nano-CT image quality"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole method assumes that a small DNN trained on one flower photograph approximates the Sobel edge response well enough on tomography subregions, and that the threshold T=800 and the known density ω carry over from one dataset to all others; if the network mislabels edges on real tomographic patches, the sharper interfaces it produces are artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Neural edge detection sharpens tomography reconstructions","Blend of DNN and optimization eliminates edge blur","Learned edges guide math solver for crisper scans","MIP-encoded neural priors improve nano-CT image quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1537,"prompt_tokens":670,"completion_tokens":867,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":803}},"tokens_in":414,"tokens_out":867,"duration_ms":10148,"temperature":1.0,"reasoning_tokens":803,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:30:06.969757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the trained DNN's output against the exact Sobel response on 3×3 patches sampled from the real zeolite and copper projection data; if the mean absolute error is of the same order as the difference between the threshold T=800 and the Sobel values it separates, the edge decisions are not trustworthy. Then re-run MIP RO with T and ω re-estimated on each dataset: if the RME/BMS superiority over CSHM disappears, much of the reported gain is parameter transfer rather than the method itself.","supporting_citations":[{"cited_title":", author Apeleo Zubiri, B","cited_arxiv_id":null,"evidence_quote":"Supplies the CSHM baseline model, the experimental zeolite and copper datasets, and the base reconstructions that MIP RO post-processes."},{"cited_title":", author Woods, R","cited_arxiv_id":null,"evidence_quote":"Justifies the Sobel operator as the edge-intensity target the DNN is trained to approximate."},{"cited_title":", author Palenstijn, W.J","cited_arxiv_id":null,"evidence_quote":"Supplies the TVR-DART discrete-tomography benchmark against which the new method is compared."},{"cited_title":", year 1972","cited_arxiv_id":null,"evidence_quote":"Supplies the SIRT iterative reconstruction benchmark used as another comparison baseline."}],"review_version":1}