{"id":"3fee599f-8a95-4c10-9139-a94d6a0315e3","arxiv_id":"1908.00593","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Learned weight factors in the universal backprojection formula reduce average reconstruction error by about half for sparse and limited-view photoacoustic tomography in simulated tests.","lead":"Researchers trained a computer to adjust the weighting inside a standard sound-based medical image reconstruction formula, and in simulations this roughly halved the average image error when measurements were incomplete. The method is a lightweight change to an algorithm many imaging systems already use, potentially improving clarity of photoacoustic images without adding extra detectors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported factor-of-two gain may be specific to the training distribution: weights are fit and evaluated on the same synthetic phantom family and directivity model, and transfer to unseen source classes or real detector characteristics is untested.","rationale":"The reader identifies generalization to real data as the weakest assumption; I agree and sharpen it. The core issue is not merely the absence of experimental data, but that the evaluation is performed entirely within the training distribution, using the same phantom generator, directivity model, and noise-free simulation. With roughly two million learned parameters and only 1800 training examples, the network can fit distribution-specific structure. The test set is drawn from the same random process, so Table 1 measures in-distribution performance, not transfer. This makes the central claim's scope ambiguous: as a statement about the simulated Shepp-Logan family with the specified cos^2 directivity, the result is plausible and supported; as a statement about improving UBP for sparse and limited-view photoacoustic tomography more generally, it is not yet established. The reader's CONDITIONAL verdict is appropriate, and my analysis does not move that verdict. I agree with the reader's weakest_assumption and would keep the verdict unchanged, while emphasizing that a distribution-shift experiment is the decisive missing check.","tokens_in":6979,"tokens_out":4295,"duration_ms":53541,"concrete_test":"Build a second test set from a genuinely different source model, for example random vascular or polygonal phantoms, and add realistic detector noise, while keeping the already trained weights fixed. Recompute the relative L2 errors for UBP and weighted UBP for scenarios A, B, and C. If the weighted-UBP advantage over UBP falls below roughly a factor of two, or if weighted UBP becomes worse than UBP on the new source class, the claimed general improvement does not survive distribution shift. As a complementary control, also retrain on the first phantom class and evaluate on the second class, to separate memorization of the training distribution from a transferable operator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Table 1) is that learned weight factors in Eq. (7) roughly halve the relative L2 error relative to standard UBP for limited-view, sparse, and combined scenarios. The load-bearing assumption is that these learned weights constitute a general improvement to UBP rather than a fit to the particular simulation distribution. In Section 3, M=1800 training phantoms are generated by a randomized Shepp-Logan procedure, and the 200 test phantoms are drawn from the same procedure. The weight tensor W has approximately 2 million free parameters, while the training set is only 1800 samples; the model is therefore high-capacity relative to the data. The held-out test set is independent in the sense of not being used for training, but it is not independent in distribution: it uses the same phantom generator, the same cos^2 directivity model (Eq. 6), the same geometry, and no noise. The reported error reduction may thus reflect learned compensation for the specific statistics of this phantom family and for the particular directivity kernel, rather than a general correction for incomplete-data artifacts. If the learned weights are tuned to the training distribution, the factor-of-two gain will not transfer to other source classes, other noise levels, or real measurement characteristics, and the practical value of the method is substantially weaker. This is the most load-bearing concern because the paper's title and conclusion generalize beyond the single simulated setup, yet no cross-distribution evidence is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a learned version of the universal backprojection (UBP) formula for two-dimensional photoacoustic tomography. Instead of the constant weight w≡1 in Eq. (7), the authors introduce a position- and detector-dependent weight tensor W and train it on simulated incomplete-data/ground-truth pairs for limited-view, sparse-sampling, and combined scenarios. The optimization is carried out with stochastic gradient descent in a Keras/TensorFlow layer that performs the time filtering and weighted backprojection. For a randomized Shepp-Logan phantom family, a cos^2 detector directivity model, and no noise, the authors report that the learned UBP roughly halves the relative l2 error compared with standard UBP on 200 held-out phantoms (Table 1), and they show learned weight images and example reconstructions.","tokens_in":7210,"tokens_out":5577,"duration_ms":53033,"significance":"The idea is simple, elegant, and computationally appealing: it keeps the fast analytic backprojection structure and merely replaces a constant by learned weights, so inference remains a single weighted backprojection pass. The evaluation is honest in that the 200 test phantoms were not used for training, and the reported gain is consistent across all three scenarios. I also credit the authors for explicitly addressing detector directivity, which standard UBP ignores. The main limitation is that the experiments stay entirely within one simulated distribution, so the general significance is conditional on transfer to other source classes, noise levels, and real detector characteristics. If such transfer is demonstrated, or if the claims are appropriately narrowed, the method is a useful addition to learned reconstruction for PAT.","major_comments":[{"comment":"The claim that learned weights generally improve reconstruction quality in PAT is not established, because both training and test sets are generated by the same randomized Shepp-Logan procedure, with the same cos^2 directivity model (Eq. (6)), the same circular geometry, and no measurement noise. With roughly 2 million free parameters in W and only 1800 training examples, the held-out 200 phantoms test independence of the particular samples but not independence of the distribution; the learned weights could be compensating for the specific statistics of this phantom family and directivity kernel. Please add out-of-distribution tests (a different phantom class, noisy data, a different directivity model, or experimental measurements), or restrict the conclusions to the simulated setting.","section":"Section 3 and Table 1"},{"comment":"The quantitative evaluation compares weighted UBP only to unweighted UBP. To support the practical significance claimed in the introduction, the method should be compared with existing iterative limited-view and sparse-data reconstructions (e.g., refs. [7,13]) and with the adaptive weighted FBP of refs. [18,21], as well as with a standard learned post-processing CNN. Without such baselines, the factor-of-two improvement over plain UBP does not tell the reader whether the method is competitive with current practice.","section":"Section 3.4"},{"comment":"The reported averages in Table 1 are not accompanied by standard deviations, confidence intervals, or the number of independent training runs, and the optimization setup is underspecified (no learning rate, no regularization, no details of the discretization of Eq. (7), no policy for random seeds). Since the training is stochastic and the high-capacity weight tensor may be sensitive to initialization, the quantitative claim needs error bars and enough implementation detail to be reproducible.","section":"Section 3 and 3.4"}],"minor_comments":[{"comment":"Typo: 'in this papers' should be 'in this paper'.","section":"Abstract"},{"comment":"Typo: 'Discetizing' should be 'Discretizing'.","section":"Section 2.2"},{"comment":"The phrase 'given on a subset of the boundary Γ ⊂ Ω' should probably read 'Γ ⊂ ∂Ω', since Γ is a subset of the detection surface.","section":"Section 2.2"},{"comment":"There is an empty citation in the sentence '...have been developed []'; the missing reference should be supplied.","section":"Introduction"},{"comment":"The caption says 'The third row displays the learned weights...', but the figure appears to have two rows of images; the row reference should be corrected.","section":"Figure 4 caption"},{"comment":"The authors do not mention code or data availability; providing the trained weights or the data-generation code would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is more a proof-of-concept than a full journal paper; the editor may wish to consider whether its length and experimental depth meet the journal's standards. The main risk is distribution overfitting, which the major comments ask the authors to address with out-of-distribution tests and comparisons to iterative baselines. I do not see a load-bearing logical error: the held-out evaluation is legitimate, so the work is not circular in the samples sense."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on photoacoustic or limited-view tomography. The new thing here is simple: instead of hand-designing angle weights for the universal backprojection formula, the authors make the full weight field w(x,s) learnable and fit it on simulated incomplete-data/ground-truth pairs. That direct optimization of UBP weights is not in the cited PAT literature, which uses hand-crafted weights or post-processing networks. The idea is a natural composition of known pieces, but it is a legitimate new application and the reported effect is large: roughly half the relative L2 error on a held-out set of 200 simulated phantoms across limited-view, sparse, and combined scenarios. The paper is clearly written, the setup is reproducible in principle, and the claims are stated without overreach. Credit where due: the held-out test set is independent of training, so the headline result is not circular.\n\nThe soft spots are real but mostly about scope, not about a flawed core. The model has about 2 million free parameters trained on 1800 phantoms; the test phantoms come from the same generator, the same cos^2 directivity model, the same geometry, and no noise. That means the factor-of-two gain may be partly a fit to the statistics of this phantom family rather than a general correction for incomplete-data artifacts. The paper does not test transfer to other source classes, other detector models, or real data, and it does not compare against iterative baselines or time-reversal, which are strong competitors for limited-view PAT. There are also no error bars on the Table 1 numbers, and no code or data released, so the practical value, while plausible, is not yet pinned down.\n\nThe central argument, however, holds up on its own terms: within the stated simulation setup, learned UBP beats standard UBP. The main limitation is that the paper generalizes beyond that setup in its conclusions without evidence. That is a standard weakness in early deep-learning-for-inverse-problems papers, and it is not a fatal flaw.\n\nWho is this for? Someone building fast reconstruction methods for PAT or other tomographic modalities where you have a good analytic backprojection and want a cheap learned correction. It is a useful building block, not a breakthrough. I would send it to peer review, but I would ask the authors for code, error bars, and at least one experiment with noise or a different phantom class. If those are added, the result becomes solid.\n\nRecommendation: engage with it; assign a referee familiar with PAT reconstruction and with the pitfalls of learning on simulated data.","headline":"A modest, clearly written methods paper: learning the weight field inside UBP roughly halves simulated reconstruction error on held-out phantoms, but the evaluation is narrow and the generalization claim is untested.","tokens_in":7760,"tokens_out":648,"would_cite":false,"duration_ms":7958,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C55","65R32","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding learned weight factors to the universal backprojection formula roughly halves the reconstruction error in limited-view and sparse photoacoustic tomography.","keywords":["photoacoustic tomography","universal backprojection","filtered backprojection","limited view","sparse sampling","detector directivity","learned reconstruction","inverse problems"],"falsifier":"Apply the same learned weights to experimentally recorded photoacoustic data from a tissue phantom under the same half-circle geometry and compare the average relative squared error over a held-out set. If the error is not close to half that of plain UBP—or if plain UBP matches it—the learned weights are tied to the simulation rather than to the imaging problem.","tokens_in":6728,"feed_emoji":"🩻","tokens_out":16065,"duration_ms":141175,"temperature":0.7,"pith_summary":"The paper argues that the standard filtered-backprojection formula used in photoacoustic tomography can be materially improved for incomplete data by making its angular weight factors learnable rather than leaving them constant. The authors train those weights on 1,800 simulated source–data pairs that include a direction-dependent detector sensitivity, then use the learned weights in the same backprojection formula. Across three incomplete-data scenarios, the learned weighted UBP cuts the average relative squared reconstruction error roughly in half (from 0.20/0.35/0.35 to 0.09/0.18/0.16). If the approach transfers beyond simulation, it matters because a small, interpretable modification of an exact analytic inversion formula gives a fast reconstruction method that also handles detector directivity, which standard UBP ignores.","feed_headline":"Learned weights halve photoacoustic reconstruction error","feed_subtitle":"Adding trained weight factors to universal backprojection cuts error roughly in half for limited-view and sparse data.","key_machinery":"The central object is the weighted universal backprojection operator $$\\Psi(w,g)(x)=-\\frac{2}{\\pi}\\int_\\Gamma w(x,s)^2 \\langle \\nu_s, x-s\\rangle \\int_{\\|x-s\\|}^\\infty \\frac{\\partial_t($t^{{-1}}$g(s,t))}{\\sqrt{$t^{2}$-\\|x-s\\|^2}}\\,dt\\,ds,$$ which maps pressure data $g$ to a reconstruction by integrating over detector positions on the measurement surface. The standard UBP is $w\\equiv 1$; replacing that constant with a learned weight tensor $W\\in\\mathbb{R}^{N_x\\times N_x\\times N_s}$ is the entire trainable modification. The paper generates training data with a directivity model $\\phi(s,\\omega)=\\cos^2(\\alpha)$ for $|\\alpha|<\\pi/2$, so the simulations already contain the angular sensitivity the learned weights must counteract. The weights are optimized by stochastic gradient descent on the mean squared error objective over 1,800 phantom–data pairs; in the solutions, the learned weights show angle-dependent structure tied to the detector arc and the source–detector angle, which is the mechanism that suppresses artifacts from missing views and coarse sampling.","core_discovery":"The paper's central claim is that the universal backprojection formula, normally used with constant weight $w\\equiv 1$, can be turned into a learned reconstruction operator by optimizing a weight for every pair of reconstruction point and detector position. The training objective is the mean squared error between the weighted backprojection and the true source on a simulated dataset with 1,800 pairs, and the evaluation is on 200 held-out phantoms. Quantitatively, the learned UBP reduces the average relative squared $\\ell^2$ error from 0.2002 to 0.0912 on half-circle limited-view data, from 0.3461 to 0.1806 on 20-detector sparse data, and from 0.3546 to 0.1649 when both limitations are combined. The paper also argues the recipe is not tied to a specific geometry: the weights can be optimized for any measurement surface, and the learned backprojection can serve as the first layer of a deeper convolutional network.","pith_inferences":["Testable extension: train the same weighted UBP with additive measurement noise and compare the optimized weight fields; if the weights shift substantially, the correction is learning noise statistics rather than geometry alone.","A consequence the authors leave implicit is that if the learned weights are smooth in the detector angle, the improvement could be compressed into a small analytic correction factor, letting existing filtered-backprojection codes adopt it without a training stage.","The weight tensor holds one scalar per reconstruction pixel per detector position, so extending the recipe to three-dimensional point-detector data would need much larger tensors; weight sharing across detector positions or radial symmetry would be required to keep training feasible."],"forward_implications":["For a fixed detector geometry, reconstruction with the learned weights remains a single backprojection pass, so the speed of standard UBP is retained while accuracy improves.","In the reported experiments, held-out phantoms show relative squared error drops from 0.2002 to 0.0912 (limited view), 0.3461 to 0.1806 (sparse), and 0.3546 to 0.1649 (combined).","Because the weight factors are optimized rather than derived, the same training recipe applies to measurement surfaces where no exact inversion formula is known.","The learned backprojection can serve as the first layer of a deep convolutional network, composing model-based reconstruction with learned artifact removal."],"supporting_citations":[{"why":"Derives the two-dimensional UBP for integrating line detectors, the inversion rule this paper modifies.","marker":"[4]"},{"why":"Defines the standard universal backprojection algorithm against which the learned version is compared.","marker":"[25]"},{"why":"Introduces heuristic angle-dependent weights for limited-angle photoacoustic reconstruction, the idea the learned weights replace.","marker":"[21]"},{"why":"Proposes iterative updates of angle-dependent weights in filtered backprojection, a direct predecessor of gradient-based weight learning.","marker":"[18]"},{"why":"Supplies the direction-dependent data model used to simulate detector directivity in the training data.","marker":"[27]"},{"why":"Shows deep learning can reconstruct photoacoustic images from sparse data, establishing the training-data paradigm.","marker":"[2]"},{"why":"Demonstrates deep-learning artifact reduction for photoacoustic projection imaging, the network into which the learned backprojection could be plugged.","marker":"[24]"}],"fun_headline_variants":["Learned backprojection halves photoacoustic error","AI-weighted backprojection cuts photoacoustic artefacts","Trained weights sharpen limited-view photoacoustics","Learned FBP improves sparse photoacoustic scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that weights trained on noiseless simulated data from one family of synthetic phantoms and one model of detector sensitivity will transfer to real photoacoustic measurements; if that transfer fails, the factor-of-two improvement may disappear outside the simulation.","fun_headline_variants_meta":{"raw":{"variants":["Learned backprojection halves photoacoustic error","AI-weighted backprojection cuts photoacoustic artefacts","Trained weights sharpen limited-view photoacoustics","Learned FBP improves sparse photoacoustic scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1359,"prompt_tokens":898,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":398}},"tokens_in":514,"tokens_out":461,"duration_ms":4097,"temperature":1.0,"reasoning_tokens":398,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:44:40.219915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same learned weights to experimentally recorded photoacoustic data from a tissue phantom under the same half-circle geometry and compare the average relative squared error over a held-out set. If the error is not close to half that of plain UBP—or if plain UBP matches it—the learned weights are tied to the simulation rather than to the imaging problem.","supporting_citations":[{"cited_title":"Temporal back-projection algorithms for photoacous- tic tomography with integrating line detectors","cited_arxiv_id":null,"evidence_quote":"Derives the two-dimensional UBP for integrating line detectors, the inversion rule this paper modifies."},{"cited_title":"Universal back-projection algo- rithm for photoacoustic computed tomography","cited_arxiv_id":null,"evidence_quote":"Defines the standard universal backprojection algorithm against which the learned version is compared."},{"cited_title":"Weight factors for limited angle photoacoustic tomography","cited_arxiv_id":null,"evidence_quote":"Introduces heuristic angle-dependent weights for limited-angle photoacoustic reconstruction, the idea the learned weights replace."},{"cited_title":"Limited-view photoacoustic imaging based on an iterative adaptive weighted filtered backprojection approach.Applied optics, 52(15):3477–3483, 2013","cited_arxiv_id":null,"evidence_quote":"Proposes iterative updates of angle-dependent weights in filtered backprojection, a direct predecessor of gradient-based weight learning."},{"cited_title":"Photoa- coustic tomography with direction dependent data: An exact series re- construction approach","cited_arxiv_id":null,"evidence_quote":"Supplies the direction-dependent data model used to simulate detector directivity in the training data."},{"cited_title":"Deep learning for photoacoustic tomography from sparse data.Inverse Prob- lems in Science and Engineering , pages 1–19, 2018","cited_arxiv_id":null,"evidence_quote":"Shows deep learning can reconstruct photoacoustic images from sparse data, establishing the training-data paradigm."}],"review_version":1}