{"id":"877ad005-7a76-4db0-b020-9cc3ab2c6461","arxiv_id":"2501.09845","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A U-Net predicts spatially adaptive Total Variation weights from a filtered backprojection, yielding stable few-view CT reconstructions, though the supporting stability proof contains an error.","lead":"This paper proposes a way to choose pixel-wise Total Variation weights in few-view CT: a neural network converts a fast filtered backprojection into an approximate image, and those weights are then fixed inside a weighted TV reconstruction. The authors add a theoretical stability analysis and show improved results on synthetic and real chest CT images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's uniqueness proof rests on Lemma 3, which mischaracterizes the subdifferential of isotropic TV at zero-gradient pixels; with the correct subdifferential the strict inequality in Eq. (3.15) is not guaranteed, so the claimed well-posedness is unproven.","rationale":"The reader's verdict is REJECT with the weakest assumption identified as Lemma 3's subdifferential formula. My independent reading confirms that this is the most load-bearing concern. The paper's theoretical contribution is the well-posedness of the Psi-W l1 scheme, and every stability theorem and corollary depends on the uniqueness of the minimizer. Lemma 3 is not a minor typo: it changes the set of admissible subgradients at flat pixels. The correct subdifferential of the Euclidean norm at the origin is the whole unit disk, so the proof's use of a single scalar c with |c|<1 to force a strict inequality in Eq. (3.15) is invalid. Even if Theorem 2 could be repaired by adding assumptions that force all such subgradients to have norm strictly less than 1, those assumptions are absent from the paper. The empirical section also tunes lambda, p, and eta on the test image and reports metrics for that same image, which weakens the numerical claims, but the mathematical flaw alone justifies a reject verdict as written. There is no formal verification or parameter-free derivation that would independently support the central claim. Since my concern matches the reader's weakest assumption and does not change the verdict, I mark agreement as 'agree' and the verdict as unchanged.","tokens_in":23048,"tokens_out":4823,"duration_ms":50914,"concrete_test":"Independently re-derive Lemma 3 for the single-pixel functional TV(x) = sqrt(x1^2 + x2^2). At x = 0 the subdifferential is {(u,v) : u^2 + v^2 <= 1}, which contradicts Eq. (3.6)-(3.7) as soon as a non-diagonal unit vector such as (1,0) is considered. Then rerun the Bregman argument of Theorem 2 using the corrected subdifferential: construct a minimizer x* with a zero-gradient pixel and a subgradient of Euclidean norm 1 at that pixel satisfying the optimality condition (3.9); if the strict inequality in (3.15) fails, the uniqueness conclusion does not follow. This directly settles whether the load-bearing premise of Theorem 2 holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the Psi-W l1 model is a well-posed regularization method. The uniqueness theorem (Theorem 2) is the load-bearing result: Theorem 3 and Theorem 4 both identify limit points with the unique minimizer of the limit problem, and Corollary 2 inherits that uniqueness. Theorem 2's proof depends on Lemma 3, which asserts that the subdifferential of the isotropic TV term at a pixel with zero gradient is restricted to (c,c) with c in [-1,1] (Eq. 3.6-3.7). This is incorrect. For the Euclidean norm at 0, the subdifferential is the full unit disk: any vector (u,v) with u^2+v^2 <= 1 is admissible. The diagonal set (c,c) with c in [-1,1] is neither a subset of the unit disk (for |c|>1/sqrt(2) its norm exceeds 1) nor does it contain all valid subgradients (e.g., (1,0) is valid but not diagonal). Consequently, Eq. (3.7) does not characterize the subdifferential. The contradiction in Eq. (3.15) relies on a strict inequality at pixels where x' has nonzero gradient but x* has zero gradient, obtained from the assumption |c|<1. With the correct subdifferential, an optimality subgradient at such a pixel can have Euclidean norm exactly 1, so the strict inequality is not guaranteed. The paper does not show that a subgradient satisfying the optimality condition (3.9) can always be chosen with norm strictly less than 1 at all such pixels. Therefore the uniqueness proof is unsupported, and the derived noise stability, reconstructor stability, and the learning-related corollary all inherit this gap. This is a mathematical flaw in the argument, not a disagreement with an external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatially adaptive weighted total variation model, denoted Ψ-Wℓ1, for few-view CT reconstruction. The weights are computed once from an intermediate image produced by a neural network (or from the magnitude of its gradient), avoiding iterative reweighting. The authors state and prove four theoretical properties—existence, uniqueness, noise stability, and reconstructor stability—and derive Corollary 2 linking the accuracy of the learned reconstructor to the accuracy of the final reconstruction. Numerical experiments on a synthetic COULE image and on Mayo Clinic chest data compare the method with global TV and with two iterative reweighted ℓ1 strategies.","tokens_in":23373,"tokens_out":6825,"duration_ms":69853,"significance":"Adaptive weighted TV is a practically important topic, and the idea of fixing the weights in advance from a learned intermediate reconstruction is computationally attractive. If the theoretical results were correct, the paper would provide a useful bridge between learned priors and classical regularization theory. The authors are also transparent in pointing out that the assumptions of Corollary 2 coincide with the training losses. The out-of-distribution noise experiment is a valuable addition. However, the central theoretical contribution is undermined by an incorrect subdifferential characterization, and the numerical evaluation contains a test-set tuning issue. As a result, the main claims are not presently supported.","major_comments":[{"comment":"The claimed characterization of the subdifferential of the isotropic TV term is incorrect. For the term ∥U_i Dx∥_2 with U_iDx = ((D_h x)_i, (D_v x)_i), the subdifferential of the Euclidean norm at the origin is the full unit disk {v ∈ R^2 : ∥v∥_2 ≤ 1}, not the diagonal set {(c,c) : c ∈ [-1,1]}. For example, (1,0) is an admissible subgradient at a zero-gradient pixel but is not of the form (c,c), while (c,c) with |c| > 1/√2 has norm larger than 1 and is not admissible. Consequently Eq. (3.7) does not describe ∂TV(x), and all subsequent uses of this formula are unsupported.","section":"Appendix A.2, Lemma 3 (Eqs. (3.6)–(3.7))"},{"comment":"Theorem 2 is load-bearing for Theorems 3 and 4 and Corollary 2, since those results identify limit points with the unique minimizer of the limit problem. The uniqueness proof depends on the strict inequality in Eq. (3.15), which uses the assumption that at a zero-gradient pixel the subgradient has the form (c,c) with |c| < 1. With the correct subdifferential, a subgradient at such a pixel can have Euclidean norm exactly one, for instance (1,0), so the strict inequality R(x') < R(x') is not guaranteed. The paper does not prove that an optimality subgradient satisfying Eq. (3.9) can always be chosen with norm strictly less than one at every pixel where x' has nonzero gradient and x* has zero gradient. Thus the uniqueness claim, and consequently the noise stability and reconstructor stability theorems, are unproven.","section":"Theorem 2 (Section 3), Eq. (3.15)"},{"comment":"All model and algorithmic parameters, including λ, p, η, and the comparison methods' parameters, are tuned to minimize the RE metric on the final solutions, and the same test image is then used to report the RE values in Table 2. This means the reported improvements over global TV and the IRℓ1 baselines are partially the result of test-set optimization. No independent validation set or cross-validation procedure is described. The empirical claim that the proposed method outperforms competing approaches is therefore weakened, independently of the theoretical issues.","section":"Section 4.1"},{"comment":"The paragraph immediately following Corollary 2 states that the quantities minimized in assumptions 1 and 2 are the same, up to the norm, as those minimized by the training losses (2.12) and (2.13). This makes the corollary a consistency statement about the training objective rather than an independent predictive guarantee for a trained network. Moreover, assumptions 1 and 2 use the ℓ1 norm while the losses use squared ℓ2 norms, so the equivalence is only qualitative. This limitation should be stated explicitly; the current framing overstates what the theorem adds beyond the training loss itself.","section":"Corollary 2 (Section 3)"}],"minor_comments":[{"comment":"The text says the relative error plots are 'consistent with the behavior discussed in Table 3', but the synthetic experiment is reported in Table 1; Table 3 concerns the Mayo dataset.","section":"Section 2.1"},{"comment":"The final paragraph contains the incomplete sentence 'FBP-GNet-W ℓ1 outperforms FBP-Net-W ℓ1, fewer noise artifacts'; it should read '... with fewer noise artifacts' or be rephrased.","section":"Section 4.2"},{"comment":"The letter M is used both for the set of minimizers in Eq. (3.4) and for the constant bounding the objective in the proofs of Theorems 3 and 4; using different symbols would avoid confusion.","section":"Section 3, Eqs. (3.4) and (3.20)"},{"comment":"In the proof of Lemma 3, the notation w is reused for the vector U_i^T v, although w already denotes the adaptive weight vector from Section 2; a different symbol, such as g^{(i)}, would be clearer.","section":"Appendix A.2"}],"recommendation":"reject","confidential_remarks":"The incorrect subdifferential formula in Lemma 3 is not a cosmetic issue: it invalidates the uniqueness proof and, through Theorems 3 and 4, the claimed well-posedness of the method. The test-set tuning in Section 4.1 further weakens the empirical evidence. Because the central theoretical argument would need to be rebuilt with the correct subdifferential and likely additional assumptions, I do not see a local fix within the scope of this manuscript, and I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I think the reader's basic verdict is right: the paper has a neat computational idea, but the theoretical proof collapses at Lemma 3. I checked the proof in Appendix A.2. The subdifferential of the Euclidean norm at zero is the whole unit disk, so restricting the subgradient to (c,c) is false. More importantly, that diagonal restriction is what makes the strict inequality in Eq. (3.15) work. Without it, the Bregman-distance argument in Theorem 2 does not go through, and since uniqueness feeds directly into Theorems 3 and 4 and Corollary 2, the well-posedness claims are not established. This is a load-bearing error, not a cosmetic one.\n\nThat said, the underlying idea is worth attention. Fixing the weights from a learned intermediate image, rather than iteratively reweighting, is a simple and sensible way to get adaptivity in few-view CT. The experiments on COULE and Mayo show consistent gains over global TV and the two IR baselines, and the out-of-distribution noise test is a reasonable robustness check. The authors also deserve credit for being explicit that the gradient loss can work even when the intermediate image itself is poor.\n\nThe soft spots beyond the theory: in Section 4.1 they tune lambda, p, and eta on the test image and then report results on that same image. That makes the gains look stronger than they are. And Corollary 2 is mostly a consistency check: the training losses in (2.12) and (2.13) minimize exactly the quantities the corollary assumes go to zero, so the empirical agreement is expected, not an independent prediction.\n\nWho is this for? People working on learned regularizers for inverse problems, especially CT. The empirical framework is useful and the writing is clear. The theory is currently wrong, though. I would send this to peer review — a good convex-analysis referee will catch the subdifferential error immediately — but my recommendation to the editor would be major revision at best. If the authors can fix Lemma 3, or scale back the claims to avoid uniqueness, the paper becomes a solid applied contribution.","headline":"Neat learned-weights TV scheme for few-view CT, but the central uniqueness proof leans on a mischaracterization of the subdifferential of isotropic TV at zero-gradient pixels.","tokens_in":23955,"tokens_out":3436,"would_cite":false,"duration_ms":33387,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J22","68T07","94A08","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-net weight guess gives few-view CT a provably stable TV model","keywords":["Weighted Total Variation","Spatially Adaptive Regularization","Few-view Tomography","Neural Networks","Image reconstruction","Well-posedness","Total Variation regularization","Inverse problems"],"falsifier":"Compute the subdifferential of the isotropic total variation term at an image with at least one zero-gradient pixel and check whether every vector in the unit disk is allowed; finding a subgradient outside the diagonal set $(c,c)$ would show the formula in equations (3.6)-(3.7) is false as stated, so the uniqueness theorem would need a different proof.","tokens_in":22795,"feed_emoji":"🩻","tokens_out":12887,"duration_ms":128821,"temperature":0.7,"pith_summary":"The paper proposes a way to make total-variation regularization spatially adaptive in few-view CT without recomputing weights at every iteration: a neural network turns a fast coarse reconstruction (filtered back projection) into an estimate of the true image or of its gradient magnitude, and that estimate fixes pixel-wise weights exactly once. The paper then proves that the resulting model $\\Psi$-W $\\ell^1$ is a well-posed regularization method: existence and uniqueness of the minimizer, noise stability as $\\delta\\to 0$, and stability with respect to the reconstructor $\\Psi$, under the standard condition $\\ker(K)\\cap\\ker(D)=\\{0\\}$. It further proves that if the network output converges to the ground truth in image norm or gradient-magnitude norm, the final solution converges to the solution obtained by using the ground truth itself to set the weights. A reader should care because the guarantee attaches the reliability of variational regularization to a learned intermediate image: the network only proposes weights, and the convex model supplies the stable reconstruction.","feed_headline":"A neural-net weight guess gives few-view CT a provably stable TV model","feed_subtitle":"One network pass fixes per-pixel weights; theory shows existence, uniqueness, and noise stability.","key_machinery":"The engine of the method is the fixed-weight functional with weights $w(\\tilde{x})_i=(\\eta^p/(\\eta^2+|D\\tilde{x}|_i^2))^{1-p}$ for $0<p<1$, computed once from a reconstructor $\\tilde{x}=\\Psi(y_\\delta)$. The weights are high (equal to 1) in flat regions and smaller near edges, so the weighted total-variation term $\\|w(\\Psi(y_\\delta))\\odot |Dx|\\|_1$ penalizes detail-rich pixels less. On the theoretical side, the proof machinery combines the primal-dual algorithm used to compute solutions with compactness of the functional's sublevel sets for stability and Bregman-distance comparisons (a convex-analytic separation measure) for uniqueness; Theorem 4 and Corollaries 1-2 make the map from the intermediate estimate to the final minimizer continuous, which is what converts network approximation error into reconstruction error.","core_discovery":"On its own terms, the paper's central claim is that the $\\Psi$-W $\\ell^1$ model (2.7) is a well-posed regularization method whose solution tracks the ideal ground-truth-weighted solution when the intermediate estimate is accurate. For every Lipschitz reconstructor $\\Psi$, every $\\lambda>0$, and every noise level $\\delta\\ge 0$, the objective admits at least one minimizer; under the condition $\\ker(K)\\cap S(x^*_{\\Psi,\\delta})=\\{0\\}$ and the subgradient condition $|c|<1$ stated in Theorem 2, the minimizer is unique. The solution depends continuously on the noise level and on the reconstructor, and Corollary 2 transfers convergence of $\\Psi(y_\\delta)$ to $x^{\\mathrm{GT}}$ (in image or gradient-magnitude norm) into convergence of $x^*_{\\Psi,\\delta}$ to the oracle solution $x^*_{\\mathrm{GT},\\delta}$. The consequence the authors emphasize is that neural-network outputs become interpretable: the network never forms the final image, it only fixes the regularization weights, and the variational stage provides stability and convergence.","pith_inferences":["The theory does not require the intermediate estimator to be trained with paired ground truths; it only needs a Lipschitz estimator of the image or its gradient magnitude, so unsupervised or classical edge-preserving preprocessors should inherit the same stability guarantees if they can control the Lipschitz error.","The stability estimates suggest a quantitative calibration experiment: measure the gradient-magnitude error of a network on held-out noise levels and check whether the final reconstruction error scales linearly with that error, as Corollary 2 would predict.","The same fixed-weight trick could be applied to wavelet or shearlet sparsity regularizers, provided the network estimates the transform-domain magnitudes; the convexity and Lipschitz-continuity properties used in the proofs are not specific to the gradient operator."],"forward_implications":["Because the global isotropic TV model is the special case with all weights equal to one, the uniqueness theorem also supplies an extension of uniqueness results from anisotropic to isotropic TV under the stated kernel conditions.","Practitioners can fix weights after a single network pass and still receive a reconstruction that is stable as measurement noise tends to zero, which is exactly the property that iterative reweighting schemes often sacrifice for speed.","The final image inherits variational stability even when the neural network is fed out-of-distribution noisy sinograms, because the network is not the terminal reconstructor.","Using the gradient-magnitude loss is at least as well justified as using the image loss, since Corollaries 1 and 2 require only closeness of gradient magnitudes to transfer accuracy.","If a reconstructor approximates the ground truth ever better, the resulting solutions converge to the oracle-weighted solution, giving a formal sense in which learned intermediate images can be trusted."],"supporting_citations":[{"why":"Defines the well-posed regularization criteria and supplies the existence-and-stability proof scheme that Theorems 1-4 follow.","marker":"[24]"},{"why":"Provides the iterative reweighted sparsity strategy whose per-iteration weight update motivates the fixed weight formula (2.6).","marker":"[21, 22]"},{"why":"Supplies the primal-dual algorithm used to compute the minimizers in all numerical experiments.","marker":"[26]"},{"why":"Gives the CT-specific iterative reweighted total variation baseline and the practical solver setup for few-view CT.","marker":"[2]"},{"why":"Introduces the reconstructor notion and the set of admissible reconstructors used in the model statement.","marker":"[25]"},{"why":"Analyzes a data-dependent regularizer analogous to the one studied here, framing the noise-stability argument.","marker":"[34]"},{"why":"Supplies the lemmas about minimizer sets and kernel alignment that the uniqueness proof adapts.","marker":"[33]"},{"why":"Defines the residual encoder-decoder network architecture used as the neural reconstructor.","marker":"[30]"}],"fun_headline_variants":["Neural net sets TV weights; theory proves stable few-view CT","One network pass, provable TV regularization for few-view CT","Adaptive TV weights from a neural net: well-posed and stable","Neural network guides TV, guarantees unique stable CT solutions","Few-view CT: neural-net weights with proven convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the paper's formula for the subdifferential of isotropic total variation at pixels where the image gradient is zero; if the true set of subgradients there is larger than the diagonal set $(c,c)$ used in equation (3.6), the Bregman-distance contradiction that proves uniqueness does not go through.","fun_headline_variants_meta":{"raw":{"variants":["Neural net sets TV weights; theory proves stable few-view CT","One network pass, provable TV regularization for few-view CT","Adaptive TV weights from a neural net: well-posed and stable","Neural network guides TV, guarantees unique stable CT solutions","Few-view CT: neural-net weights with proven convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2312,"prompt_tokens":890,"completion_tokens":1422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1336}},"tokens_in":506,"tokens_out":1422,"duration_ms":11632,"temperature":1.0,"reasoning_tokens":1336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:37:50.290476+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the subdifferential of the isotropic total variation term at an image with at least one zero-gradient pixel and check whether every vector in the unit disk is allowed; finding a subgradient outside the diagonal set $(c,c)$ would show the formula in equations (3.6)-(3.7) is false as stated, so the uniqueness theorem would need a different proof.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the well-posed regularization criteria and supplies the existence-and-stability proof scheme that Theorems 1-4 follow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the primal-dual algorithm used to compute the minimizers in all numerical experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the CT-specific iterative reweighted total variation baseline and the practical solver setup for few-view CT."},{"cited_title":"To be or not to be stable, that is the question: understanding neural networks for inverse problems","cited_arxiv_id":"2211.13692","evidence_quote":"Introduces the reconstructor notion and the set of admissible reconstructors used in the model statement."},{"cited_title":"A data-dependent regularization method based on the graph Laplacian","cited_arxiv_id":"2312.16936","evidence_quote":"Analyzes a data-dependent regularizer analogous to the one studied here, framing the noise-stability argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the lemmas about minimizer sets and kernel alignment that the uniqueness proof adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the residual encoder-decoder network architecture used as the neural reconstructor."}],"review_version":1}