{"id":"521786f8-1828-47e5-8e3f-aef90fc08397","arxiv_id":"1908.02825","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A2TV regularization, with a structure-tensor-adaptive anisotropy field, improves optoacoustic reconstruction of vessel-like structures compared with TV-L1 in simulations and a phantom experiment.","lead":"This paper applies an adaptive anisotropic total-variation regularizer to optoacoustic tomography, aiming to reconstruct blood-vessel images while preserving complex, non-convex boundaries. On numerical simulations and one experimental phantom, the method shows stronger contrast for weak structures than a TV-L1 baseline, at the cost of more tuning parameters and no convergence proof.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparative claim rests on manually selected parameters and a single un-repeated phantom with a single normalized 1D slice; with four tunable parameters versus two for TV-L1, the reported 4x contrast advantage may be overfitting rather than a robust property of A2TV.","rationale":"The reader identified the alternating-optimization convergence as the weakest assumption. I agree that is a genuine limitation, and the paper acknowledges it in Section 3.2. However, I do not think it is the most load-bearing for the central claim. The central claim is empirical: A2TV outperforms TV-L1 in preserving non-convex morphology and weak-structure contrast. Even if the algorithm converged perfectly, the current evidence would still be insufficient, because the experimental demonstration uses one phantom, manually chosen parameters, and a single normalized slice; and in the sparse-projection simulation the quantitative MAD favors TV-L1. Conversely, if the comparison were made with an objective, pre-registered parameter-selection rule and repeated measurements, the central claim could be supported even without a convergence proof, since heuristic nonconvex optimization is often acceptable when validated empirically. Thus the load-bearing weak point is the uncontrolled comparison, not the missing proof. The proposed test would directly check whether the 4x advantage is robust to parameter selection and measurement noise. If it is not, the paper's claim would need to be downgraded to a proof-of-concept with no demonstrated advantage over TV-L1; the current CONDITIONAL verdict is appropriate, possibly with an explicit request for the pre-registered comparison before acceptance.","tokens_in":14669,"tokens_out":9970,"duration_ms":109185,"concrete_test":"Design a pre-registered comparison on the numerical phantom: select all parameters for both methods by an objective rule (e.g., L-curve or cross-validation on a validation split of the projection data, without using the ground-truth image for test); freeze those parameters; acquire the experimental hair phantom at least five times under identical conditions; compute the weak-hair contrast as a region-based peak-to-peak or signal-to-noise statistic over the lower part of the image, with bootstrap confidence intervals. If A2TV's contrast ratio over TV-L1 is not significantly above 1 across repeats, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that A2TV is more capable than TV-L1 at preserving non-convex boundaries and enhancing weak-structure contrast is not yet established because the comparison is not controlled. In the noisy-data simulation, parameters are selected by lowest MAD (Section 4), an oracle choice using the true image; in the sparse-projection simulation, TV-L1 actually achieves a lower MAD (Fig. 11), yet the paper still claims superiority based on visual inspection. The experimental support (Section 5) is a single hair phantom: the reported 4x peak-to-peak figure is read from one 1D slice (Fig. 15g) after normalizing slices by their maximum, with no repeated measurements, no error bars, and no stated selection rule for choosing Figs. 13e/14e from the parameter grid. A2TV has four free parameters (sigma, rho, lambda, k) versus two for TV-L1; with manual selection on the same phantom, the extra flexibility can absorb measurement-specific noise and make the method look superior. The absence of a comparison to the adaptive-TV baseline [25] also prevents attributing the effect to A2TV rather than to the added adaptive degrees of freedom. Thus, as presented, the evidence leaves open that the advantage is an artifact of tuning rather than a robust property of the regularizer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive anisotropic total-variation (A2TV) regularizer for model-based optoacoustic tomography. The tensor A(x) is constructed from an initial estimate using structure-tensor eigenvalue analysis, and the reconstruction minimizes the A2TV energy together with a quadratic data-fidelity term, solved by a Chambolle-Pock-type algorithm in which the tensor is updated within the iterations. The method is evaluated on simulated mouse-retina vasculature images with additive Gaussian noise and with 32 projections, and on an experimental phantom of four intersecting hairs. The authors claim that A2TV preserves non-convex blood-vessel boundaries and enhances weak-structure contrast better than TV-L1 regularization.","tokens_in":14920,"tokens_out":5111,"duration_ms":55262,"significance":"If the claims are established, A2TV would provide a practical regularization framework that avoids the boundary-rounding artifact of conventional TV while enabling stronger denoising, with natural extensions to 3D and 4D optoacoustic imaging. The paper contains a clear forward-model formulation, an explicit algorithm description, and both numerical and experimental demonstrations. However, the central comparative claims currently rest on manual parameter selection, a single experimental phantom without repeated measurements, a visual morphology assessment in the case where the quantitative metric favors TV-L1, and an unproved convergence of the adaptive alternating scheme. These issues must be resolved before the claimed advantage over TV-L1 can be considered robust.","major_comments":[{"comment":"The manuscript explicitly states in Section 3.2 that there is no mathematical proof of convergence of the alternating minimization, and Algorithm 1 updates the tensor A(x) inside the Chambolle-Pock iterations (step 12), which changes the operator K during the primal-dual update. Because the reconstructions and the comparison in Sections 4 and 5 depend on the fixed point of this procedure, the lack of a convergence guarantee is load-bearing. Please provide empirical convergence evidence (e.g., residual norms of u and A versus iteration count, or a fixed-point distance measure) for the noisy, sparse, and experimental cases, or give a proof under reasonable assumptions.","section":"Section 3.2 and Algorithm 1"},{"comment":"In the sparse-projection case the paper reports that TV-L1 achieves a lower MAD than A2TV (text near Figs. 9-11) and then claims superiority of A2TV based on visual inspection of vessel morphology. A quantitative metric such as boundary localization error, Dice overlap of a segmented vessel mask, or structural similarity must be reported to support the claim that A2TV preserves non-convex structures better; otherwise the central comparative conclusion is contradicted by the stated MAD values.","section":"Section 4, Figs. 9-11"},{"comment":"The experimental support is a single agar phantom, with no repeated measurements and no error bars, and the reported 4x peak-to-peak improvement in weak-hair contrast is read from one normalized 1D slice (Fig. 15g). The choice of the displayed reconstructions (Figs. 13e and 14e) from the parameter grid is not justified by any stated selection rule. To support the contrast-enhancement claim, the authors should report statistics over multiple slices and repeated measurements and define the parameter-selection criterion before evaluating the reconstructions.","section":"Section 5, Fig. 15"},{"comment":"The adaptive directional TV method of Wang et al. [25] is cited in the Introduction but is never compared against numerically or experimentally. Since A2TV's additional adaptive degrees of freedom may be responsible for the reported gains, including [25] as a baseline (or explicitly arguing why it is not applicable to the settings considered) is necessary to attribute the improvement specifically to A2TV rather than to the added adaptivity.","section":"Section 5 and comparison to [25]"}],"minor_comments":[{"comment":"The displayed update contains the string 'xxxxxx' between the two proximal arguments; this appears to be a typographical artifact and should be corrected.","section":"Algorithm 1, line 7"},{"comment":"Please state explicitly whether the MAD-based parameter selection used the true image as an oracle; if so, discuss the implications for the comparison and whether a validation-set rule would change the conclusions.","section":"Section 4"},{"comment":"The text does not specify the number of experimental measurements or the protocol for identifying the weak hair structure; please add these details for reproducibility.","section":"Section 5"},{"comment":"The stopping criterion for 'numerical convergence' is not quantified; please report the iteration count at which the reconstructions stabilize in terms of the change in u and A.","section":"Section 4 and Algorithm 1"},{"comment":"Reference [26] is cited as an arXiv preprint; if it has since been published in a peer-reviewed venue, please update the citation.","section":"Reference [26]"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem in optoacoustic imaging and the A2TV idea is interesting, but the evidence supporting the strong comparative claims is not yet at the level required. The revision should add quantitative boundary-preservation metrics, empirical convergence checks, and a more thorough experimental validation with repeated measurements. I would not reject the paper, but I would require these additions before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a clean application of the authors' A2TV regularizer to model-based optoacoustic inversion, not a new theory paper. The functional itself is from their prior work; what's new is the structure-tensor construction, the Chambolle-Pock scheme, and tests on simulated vessel images and one hair phantom. On the positive side, the forward model is standard, the algorithm is clearly stated, and the qualitative effect—less boundary rounding than TV-L1 under strong regularization—shows up consistently in the noisy-data simulations and in the phantom. The 1D slice in Fig. 15g does show a weak hair with substantially higher peak-to-peak amplitude in A2TV than in TV-L1. I believe the effect is probably real, in the sense that adaptive anisotropic TV can preserve non-convex structures better than isotropic TV.\n\nThe soft spots are exactly where the reader's report puts them. There is no convergence proof for the alternating update of u and A(x); the paper says so in Section 3.2. That limits the theoretical framing but not necessarily the practical value. The experiment is one phantom, no repeats, no error bars, and the 4x figure comes from one normalized slice. With four tuning parameters (sigma, rho, lambda, k) versus two for TV-L1, manual selection on the same data leaves room for overfitting. The comparison also omits the closest baseline, Wang et al.'s directional TV [25], so we cannot tell whether the gain comes from the A2TV functional or just from added adaptivity. One thing the stress-test note gets half-wrong: the sparse-projection simulation is reported honestly—they state TV-L1 achieved the lower MAD and attribute it to better texture preservation. They still lean on visual superiority for A2TV, which is weaker evidence, but they did not hide the quantitative result. The citation pattern is fine: [26] is their own theory paper and [25] is properly credited.\n\nWho is this for: people working on regularization for optoacoustic or other linear inverse problems with vessel-like structure. It is not a breakthrough and not a definitive validation, but it is a serious, well-written engineering contribution with an honest limitation section. It deserves peer review, and a referee should ask for a controlled parameter-selection protocol and the missing comparison to [25] before publication.","headline":"A clean, honest application of the authors' A2TV regularizer to optoacoustic inversion; the core idea is sound but the headline 4x contrast claim rests on thin experimental support.","tokens_in":15486,"tokens_out":1964,"would_cite":true,"duration_ms":22415,"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":"Adaptive anisotropic total-variation regularization preserves non-convex blood-vessel boundaries in optoacoustic tomography.","keywords":["optoacoustic tomography","adaptive anisotropic total variation","total variation regularization","model-based reconstruction","blood vessel imaging","non-convex boundary preservation","structure tensor","primal-dual optimization"],"falsifier":"A reader could settle the claim by imaging a phantom with a known ground-truth vessel whose boundary is deliberately non-convex with high curvature, scanning the regularization parameters for both A2TV and TV-$L_1$, and measuring the distance between the reconstructed boundary and the truth, for example the mean absolute error or Hausdorff distance at matched regularization strength. If TV-$L_1$ preserves the boundary as well as A2TV, or if A2TV's advantage disappears when the ground truth is known, the central claim fails; likewise, if the alternating optimization stalls and the final image depends strongly on initialization, the practical claim collapses.","tokens_in":14424,"feed_emoji":"🩸","tokens_out":8503,"duration_ms":86217,"temperature":0.7,"pith_summary":"Optoacoustic tomography reconstructs images from acoustic signals, often from noisy or incomplete data; the paper proposes a new regularization scheme that is designed for the images this modality most often produces, networks of thin, curved blood vessels. The scheme, adaptive anisotropic total-variation (A2TV), replaces the isotropic total-variation penalty with one whose direction is tuned locally to the vessel geometry, so it can smooth noise along vessel walls without rounding their non-convex bends. In numerical simulations on a vascular image and in an experimental phantom of intersecting hairs, the authors show that A2TV preserves vessel morphology under strong regularization better than the standard TV-$L_1$ scheme, and in the experiment a weak hair structure appears with a peak-to-peak signal more than four times higher. A sympathetic reader would take this as evidence that A2TV is a practical tool for vasculature imaging, where keeping fine curved boundaries is more important than keeping intra-vessel texture.","feed_headline":"A2TV preserves curved blood vessels in optoacoustic scans","feed_subtitle":"On a hair phantom, a weak structure shows over four times the contrast of TV-L1 regularization.","key_machinery":"The central object is the A2TV functional $J_{\\mathrm{A2TV}}(u)=\\int_\\Omega \\|A(x)\\nabla u(x)\\|_2\\,dx$, where $A(x)\\in\\mathbb{R}^{2\\times2}$ is a spatially adaptive tensor that defines an adaptive gradient. The tensor is built from the eigenvalue decomposition of a smoothed structure tensor $J_\\rho(\\nabla u_{0;\\sigma})$ computed from an initial reconstruction: in flat regions its eigenvalues are left equal so the cost reduces to standard isotropic TV, while near edges the eigenvalue along the gradient direction is suppressed so regularization flows along the tangent, preserving boundary curvature. The anisotropy parameter $k$ controls how much of the image is treated anisotropically and thereby the degree of non-convexity that can be preserved. Reconstruction is posed as a generalized ROF model, $u^*=\\arg\\min_u J_{\\mathrm{A2TV}}(u)+\\frac{\\lambda}{2}\\|Mu-p\\|_2^2$, and solved with a first-order primal-dual algorithm that alternates between updating the image and updating the tensor $A(x)$.","core_discovery":"The paper's central discovery is that a regularizer which adapts its anisotropy to the local image structure, A2TV, can be embedded in model-based optoacoustic inversion and outperforms TV-$L_1$ when the reconstruction target has complex, non-convex boundaries. Concretely, the authors demonstrate on a mouse-retina vasculature image that under strong regularization TV-$L_1$ smears the vessels, while A2TV removes noise between vessels and preserves their curvature; in the sparse-projection (32-projection) simulation TV-$L_1$ attains a lower mean absolute distance because it preserves intra-vessel texture, yet A2TV still better retains fine vessel morphology. On the experimental hair phantom, the A2TV reconstruction achieves the highest image quality, with the bottom weak hair reaching over four times the peak-to-peak signal of the TV-$L_1$ reconstruction. The underlying claim is that A2TV shifts the trade-off frontier: stronger denoising no longer costs the vessel boundary shape.","pith_inferences":["If A2TV's boundary preservation transfers to other ill-posed linear inversions whose targets are thin curved structures, such as ultrasound or x-ray imaging of vessels, the same tensor-adaptation mechanism could replace isotropic TV in those settings; the paper does not test this.","The four-parameter tuning ($\\sigma$, $\\rho$, $\\lambda$, $k$) is the main practical cost; a natural extension would be to make the structure-tensor scales adaptive to local noise or resolution, which could widen the range of $\\lambda$ over which A2TV beats TV-$L_1$.","Because A2TV deliberately removes texture inside vessels, it would be unsuitable for imaging tasks where that texture carries information, such as speckle-based flow or red-blood-cell distribution; the paper acknowledges this trade-off but does not quantify it.","The non-convexity of the joint energy suggests a testable diagnostic: recording the full energy or the dual residual over iterations could reveal whether the heuristic convergence seen in the examples is reliable, and whether restarting the tensor update from different initializations changes the reconstruction."],"forward_implications":["In optoacoustic images dominated by blood vessels, A2TV permits stronger denoising (lower $\\lambda$) without smearing vessel morphology, making weak and small vessels visible that TV-$L_1$ would erase.","Under-sampled data: both A2TV and TV-$L_1$ remove the streak artifacts of unregularized inversion; A2TV sacrifices intra-vessel texture, which the paper treats as an acceptable price for cleaner morphology.","Because the structure-tensor construction is dimension-agnostic, the same A2TV framework extends to 3D and to 4D optoacoustic reconstruction (three spatial dimensions plus time), where TV regularization has already been used.","In applications where vascular morphology is the diagnostic feature, A2TV offers a different operating point on the regularization frontier than TV-$L_1$: equal denoising with less boundary rounding."],"supporting_citations":[{"why":"Supplies the theoretical analysis showing A2TV admits non-convex, high-curvature stable structures, which motivates the regularization.","marker":"[26]"},{"why":"Defines the TV-$L_1$ sparsity-plus-total-variation baseline that A2TV is compared against.","marker":"[15]"},{"why":"Earlier directional-TV reconstruction for optoacoustic imaging that A2TV extends by using structure-tensor eigenvalue decomposition.","marker":"[25]"},{"why":"Provides the anisotropic-diffusion tensor construction and parameter values used to build the adaptive tensor $A(x)$.","marker":"[34]"},{"why":"Supplies the first-order primal-dual algorithm used to solve the generalized ROF model in the inversion.","marker":"[37]"},{"why":"Supplies the linear-interpolation forward model used to construct the model matrix $M$.","marker":"[32]"},{"why":"Provides the LSQR solver used for the unregularized baseline reconstruction.","marker":"[33]"}],"fun_headline_variants":["A2TV sharpens optoacoustic images with adaptive edge preservation","Adaptive regularization boosts contrast in optoacoustic scans","Optoacoustic imaging: adaptive TV preserves vessel curves","New regularization solves optoacoustic edge blur","A2TV: stronger denoising without losing vessel shape"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The image and the adaptive tensor are updated alternately, and although the energy is convex for a fixed tensor, it is non-convex in the joint variables; the paper states there is no mathematical proof of convergence, so the claimed boundary preservation depends on this alternation actually reaching a good fixed point.","fun_headline_variants_meta":{"raw":{"variants":["A2TV sharpens optoacoustic images with adaptive edge preservation","Adaptive regularization boosts contrast in optoacoustic scans","Optoacoustic imaging: adaptive TV preserves vessel curves","New regularization solves optoacoustic edge blur","A2TV: stronger denoising without losing vessel shape"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1246,"prompt_tokens":880,"completion_tokens":366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":496,"tokens_out":366,"duration_ms":4280,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:32:48.750105+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by imaging a phantom with a known ground-truth vessel whose boundary is deliberately non-convex with high curvature, scanning the regularization parameters for both A2TV and TV-$L_1$, and measuring the distance between the reconstructed boundary and the truth, for example the mean absolute error or Hausdorff distance at matched regularization strength. If TV-$L_1$ preserves the boundary as well as A2TV, or if A2TV's advantage disappears when the ground truth is known, the central claim fails; likewise, if the alternating optimization stalls and the final image depends strongly on initialization, the practical claim collapses.","supporting_citations":[{"cited_title":"Adaptive Anisotropic Total Variation - A Nonlinear Spectral Analysis","cited_arxiv_id":"1811.11281","evidence_quote":"Supplies the theoretical analysis showing A2TV admits non-convex, high-curvature stable structures, which motivates the regularization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the TV-$L_1$ sparsity-plus-total-variation baseline that A2TV is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier directional-TV reconstruction for optoacoustic imaging that A2TV extends by using structure-tensor eigenvalue decomposition."},{"cited_title":"Chambolle, T","cited_arxiv_id":null,"evidence_quote":"Supplies the first-order primal-dual algorithm used to solve the generalized ROF model in the inversion."},{"cited_title":"Rosenthal, D","cited_arxiv_id":null,"evidence_quote":"Supplies the linear-interpolation forward model used to construct the model matrix $M$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LSQR solver used for the unregularized baseline reconstruction."}],"review_version":1}