{"id":"a04fd438-56e3-4898-9f42-a811c55c661f","arxiv_id":"1908.00975","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Y-Net, a dual-encoder CNN that fuses raw photoacoustic signals with beamformed images, improves limited-view photoacoustic reconstruction over U-Net in simulation, phantom, and in-vivo tests.","lead":"This paper proposes Y-Net, a neural network that reconstructs photoacoustic images by combining raw ultrasound signals with a standard beamformed image as two separate inputs. It reports better image quality than a standard U-Net in simulation, on a chicken breast phantom, and on a human palm, suggesting a practical way to reduce artifacts in photoacoustic imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-vivo comparison is compromised by an undisclosed replacement of the DAS input to Encoder II, making the claimed Y-Net advantage non-reproducible.","rationale":"The reader's weakest assumption identifies the simulation-to-real gap and specifically flags the undisclosed 'better texture' input in Section IV.D. My stress-test confirms that this is the single most load-bearing concern: the paper's headline in-vivo validation is irreproducible without disclosure, and the comparison to U-Net is potentially unfair if the input differs. The synthetic test set (Table I) provides some support for Y-Net's architecture, but the central claim of the paper extends to in-vivo performance, and that claim rests on an ad hoc, unspecified alteration. The paper does acknowledge the gap and provides source code for the synthetic experiments, which is positive, but the in-vivo modification is not documented enough to replicate. Thus the conditional verdict stands, with the condition being full disclosure and matched-input evaluation. No change to the reader's verdict is needed.","tokens_in":13967,"tokens_out":3311,"duration_ms":29599,"concrete_test":"Require the authors to disclose the exact nature of the 'better texture reconstructed result' (algorithm, parameters, code) and rerun the in-vivo comparison in two matched conditions: (1) both Y-Net and U-Net use DAS input only; (2) both use the disclosed better input, with the networks retrained on that input if necessary. If Y-Net does not outperform U-Net under condition (1) or (2) with consistent metrics (e.g., across 10 independent trials), the in-vivo claim of superiority is unsupported. Additionally, quantify the distribution shift by measuring the input statistics (e.g., mean/std, artifact level) of the disclosed input versus training DAS images.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the undocumented alteration of Encoder II's input in the in-vivo experiment (Section IV.D): 'we alter the input of Encoder II, which is revised as a better texture reconstructed result instead of DAS.' No description, algorithm, or parameters are given, and no ablation is shown with this input. Since the network was trained on DAS inputs (Section IV.A), substituting a different input at inference creates an unquantified distribution shift. The paper claims Y-Net 'showed good performance, with no excessive associations' versus U-Net in Fig. 8, but without knowing what input U-Net received (the text says 'Likewise, the results of post-processing method are also improved'), the comparison is not reproducible and the advantage cannot be attributed to the architecture. If the 'better texture' input is, for example, a model-based reconstruction, then both networks may simply be learning to refine that better input, making the comparison with DAS-trained models meaningless. This directly undermines the central claim of in-vivo superiority and the general claim that Y-Net fills the gap between direct and post-processing methods.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Y-Net, a convolutional neural network with two encoders (one taking raw photoacoustic signals, the other taking a delay-and-sum beamformed image) and a shared decoder, intended as a non-iterative 'hybrid' reconstruction method for photoacoustic tomography. The network is trained on k-Wave simulations of segmented retinal vessels from the DRIVE dataset and evaluated on a synthetic test set, an in-vitro chicken-breast phantom, and an in-vivo human-palm experiment. The authors report that Y-Net achieves the highest SSIM, PSNR, and SNR on the synthetic test set among the compared learned and conventional methods, and qualitatively state superior performance in the experimental cases. The paper also provides an ablation study (removing either encoder path) and a point-target resolution comparison.","tokens_in":14194,"tokens_out":2168,"duration_ms":22676,"significance":"If the claims hold, the dual-encoder architecture is a reasonable and practical contribution to learned photoacoustic reconstruction, and the paper explicitly targets a real gap between direct raw-data-to-image methods and image-post-processing methods. Strengths include the use of a public vessel dataset for simulation, the inclusion of ablation studies, a point-target resolution test, and a public source-code link. The central quantitative claim on synthetic data is supported by the reported metrics, though the absolute advantage over U-Net is small. The in-vivo comparison, however, is undermined by an undocumented change to the network input in Section IV.D, which prevents attribution of the observed improvement to the architecture and makes the in-vivo claim non-reproducible as written.","major_comments":[{"comment":"The in-vivo experiment is not reproducible because the input to Encoder II is changed without specification: the text states 'we alter the input of Encoder II, which is revised as a better texture reconstructed result instead of DAS,' but no algorithm, parameters, or example of this input are given, and no ablation is shown with this input. Since the network was trained on DAS inputs (Section IV.A), substituting a different input at inference creates an unquantified distribution shift. The claimed advantage of Y-Net over U-Net in Fig. 8 cannot be attributed to the architecture unless the alternative input is fully described and either the network is retrained with it or a controlled ablation is provided. This issue is load-bearing for the abstract's claim of 'validated with experiments ... in vivo, which still performs better than other existing methods.'","section":"Section IV.D and Fig. 8"},{"comment":"The quantitative comparison on the synthetic test set reports only point estimates (SSIM 0.9119 vs 0.9002; PSNR 25.5434 vs 25.0032 dB; SNR 9.9291 vs 9.3233) without error bars, standard deviations, or significance tests across the 400 test samples. Given the small margins, the reader cannot assess whether Y-Net's advantage over U-Net is statistically meaningful or within run-to-run variability. The paper should report per-sample distributions, confidence intervals, or paired statistical tests.","section":"Table I and Section V.A"},{"comment":"The evaluation metrics are computed on a test set generated by the same k-Wave simulation pipeline used for training, so the synthetic results reflect interpolation within the training distribution rather than generalization to unseen acquisition conditions. The in-vitro and in-vivo results are qualitative and, as noted above, the in-vivo comparison is compromised. The paper should explicitly acknowledge this limitation and, if possible, include a robustness test (e.g., different acoustic speeds, noise levels, or transducer geometries) to support the claimed generalization.","section":"Section IV.A and IV.B"}],"minor_comments":[{"comment":"The hyper-parameter λ in the total loss (Eq. 19) is set to 0.5 without any sensitivity analysis; a brief statement on how the results vary with λ would help reproducibility.","section":"Section III.D"},{"comment":"Several equations appear to have rendering or typographical issues, for example Eq. (2) is missing part of the integrand and Eq. (3) has a misplaced comma; these should be carefully revised.","section":"Equations (2) and (3)"},{"comment":"The qualitative figures would benefit from consistent color bars, scale bars, and clear labels for the ROIs, as the reader cannot currently judge the absolute intensity scales.","section":"Fig. 5 and Fig. 8"},{"comment":"The point-target resolution comparison is presented qualitatively with a single profile plot; reporting a quantitative metric such as full-width at half-maximum for the reconstructed points would strengthen this comparison.","section":"Section V.A, point-target experiment"},{"comment":"Reference [31] is cited as 'unpublished' but appears to be an EMBC 2019 paper; the final citation should be updated to its published version if available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The undocumented change of Encoder II's input in the in-vivo experiment is the main obstacle; it is a reproducibility issue rather than an unfixable one, so I recommend major revision rather than rejection. The self-citations are numerous but not excessive for a methods paper. I would also ask the editor to ensure the source-code link is live and contains the exact model and preprocessing used for the reported results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:1908.00975. The paper proposes Y-Net, a U-Net variant with two encoders: one takes the raw PA signals, the other takes a DAS beamformed image, and they feed a shared decoder. That's a natural idea, and the authors implement it cleanly. They release code, show an ablation, and report small but consistent gains on a simulated test set: SSIM 0.9119 vs 0.9002 for U-Net, PSNR +0.54 dB. The phantom and in-vivo images look reasonable at a glance. If you work on learned reconstruction for photoacoustic or similar inverse problems, the dual-encoder architecture is worth knowing about.\n\nThe soft spots are real, though. The central problem is the in-vivo experiment. In Section IV.D they say they 'alter the input of Encoder II, which is revised as a better texture reconstructed result instead of DAS.' They don't say what that better input is, how it was made, or why it's fair to both methods. They also say the post-processing method (U-Net) was improved the same way, but without details the comparison in Fig. 8 is not reproducible, and you can't attribute whatever advantage Y-Net shows to the architecture. This is a load-bearing flaw in the in-vivo claim.\n\nOther issues are smaller. The simulation test set comes from the same k-Wave generator used for training, which gives the metrics a mild self-referential character. There are no error bars or significance tests, so the 1% SSIM gain over U-Net might be within run-to-run variation. They don't compare against an iterative model-based method, which matters because their framing is about filling the gap between direct and post-processing; a TV-regularized solver would be a natural baseline. None of these kill the paper, but they limit how much you can conclude.\n\nOn balance: the architecture is plausible, the code and synthetic data support the core idea, and the authors are honest about some of the domain-shift issues. But the in-vivo evaluation as reported is not a valid comparison. I'd send this to peer review with a request for major revision: disclose the Encoder II input for the in-vivo experiments and rerun both Y-Net and U-Net with identical inputs, add error bars, and include an iterative baseline. The way it stands, I wouldn't cite the in-vivo claim.","headline":"Y-Net is a plausible dual-encoder architecture with code and a clean synthetic validation, but the in-vivo comparison is compromised by an undisclosed change to Encoder II's input, so the current paper needs major revision before its claims can be trusted.","tokens_in":14720,"tokens_out":2601,"would_cite":false,"duration_ms":25947,"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":"A two-encoder CNN called Y-Net reconstructs photoacoustic images better than single-input U-Net.","keywords":["photoacoustic computed tomography","limited-view reconstruction","convolutional neural network","dual-encoder architecture","delay-and-sum beamforming","hybrid deep learning reconstruction","in vivo imaging","simulation-based training"],"falsifier":"Run the same training protocol on a matched set of real ex-vivo measurements with high-quality references, feeding both Y-Net and U-Net the identical delay-and-sum image with no ad hoc input replacement; if Y-Net's margin disappears or reverses, the reported advantage is an artifact of simulation transfer rather than the dual-input architecture.","tokens_in":13784,"feed_emoji":"🩻","tokens_out":9217,"duration_ms":81011,"temperature":0.7,"pith_summary":"The paper tries to establish that a convolutional network with two encoders—one reading raw photoacoustic signals, one reading a delay-and-sum beamformed image—reconstructs photoacoustic images better than networks using either input alone. The two inputs are meant to complement each other: the beamformed image supplies overall texture while the raw data restores fine details lost to artifacts. If true, the architecture offers a fast, non-iterative route to higher-quality photoacoustic images in real-time settings, and the same dual-input idea could transfer to other imaging modalities. The supporting experiments compare against U-Net and two ablated variants on simulated vessel data and on chicken-breast and human-palm measurements.","feed_headline":"Dual-encoder Y-Net tops U-Net in photoacoustic imaging tests","feed_subtitle":"Feeding raw signals plus a delay-and-sum image lifts SSIM to 0.91 on simulated data and cleans up in-vivo vessel images.","key_machinery":"The load-bearing architecture is the Y-shaped network: Encoder I compresses raw time-series photoacoustic signals into a small feature map, Encoder II compresses a delay-and-sum beamformed image, and a single decoder concatenates features from both arms at every level through skip connections, outputting a 128-by-128 initial-pressure image. The two arms are trained with a composite loss: a reconstruction mean-squared-error on the output plus an auxiliary loss that asks Encoder II alone to reproduce the ground truth, so the beamformed branch is explicitly responsible for texture while the raw-signal branch fills in missing detail.","core_discovery":"The central claim is that Y-Net, a U-Net-like encoder-decoder CNN with a second encoder branch for raw photoacoustic signals, outperforms single-input deep-learning reconstruction (U-Net) and conventional algorithms for photoacoustic computed tomography. On the synthetic test set it reports the best scores of all compared methods (SSIM 0.9119, PSNR 25.5434 dB, SNR 9.9291), and in chicken-breast and human-palm experiments it states that Y-Net retains clearer structure with fewer false vessel associations. All networks are trained on 4700 simulated phantoms built from segmented retinal-vessel patterns under a fixed linear-array geometry and scored on 400 held-out phantoms; the paper attributes the gain to the hybrid input, with the beamformed branch providing texture and the signal branch adding details the beamforming step discarded.","pith_inferences":["The in-vivo protocol's switch to a 'better texture reconstructed result' for Encoder II means the real-tissue comparison tests a domain-adapted variant, not the exact simulation-trained architecture; quantifying that adaptation would be needed to compare fairly with U-Net.","A direct extension would be to train with randomized acoustic parameters and source geometries; if Y-Net's edge over U-Net widens with diversity, the mechanism is the dual-input fusion, while if it shrinks, the edge may come from better-matched priors rather than the architecture.","The same two-input pattern—raw sensor time series plus fast back-projection—could transfer to other limited-view tomographies, such as ultrasound or microwave imaging, where analogous artifacts and detail loss occur."],"forward_implications":["Any fast conventional reconstruction, not only delay-and-sum, can be plugged into Encoder II, so improvements in beamforming should directly improve Y-Net's output.","At roughly 0.03 seconds per image, the network is fast enough for real-time photoacoustic computed tomography workflows, unlike iterative model-based reconstruction.","The synthetic test results place Y-Net above both of its ablated variants and above U-Net, which suggests both encoder branches contribute to the improvement.","If the hybrid input strategy is the reason for the gain, other imaging modalities that have both raw detector data and a fast approximate reconstruction can borrow the same dual-encoder design."],"supporting_citations":[{"why":"The wave-propagation simulation engine used to generate the 4700 training and 400 test photoacoustic datasets on which all compared methods are trained and scored.","marker":"[47]"},{"why":"The U-Net architecture that is the main deep-learning baseline and whose encoder-decoder skeleton the two-arm Y-Net extends.","marker":"[44]"},{"why":"The source of segmented retinal-vessel images used as initial-pressure ground truths for the simulated training and test phantoms.","marker":"[48]"},{"why":"Defines the SSIM evaluation metric reported in the quantitative comparison table.","marker":"[49]"},{"why":"Provides the stochastic optimizer used to train the network to the reported loss.","marker":"[45]"}],"fun_headline_variants":["Dual-input Y-Net sharpens photoacoustic images in vivo","Y-Net's twin encoders beat single-input U-Net for PAI","Hybrid Y-Net lifts photoacoustic clarity over U-Net","Two-branch Y-Net improves photoacoustic reconstruction","Y-Net: raw plus beamformed beats standard deep learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that vessel-like phantoms simulated under one fixed acoustic setup capture the structure of real tissue signals closely enough that a network trained only on those simulations improves real reconstructions; when this fails, the in-vivo test substitutes a better beamformed input, which masks the mismatch.","fun_headline_variants_meta":{"raw":{"variants":["Dual-input Y-Net sharpens photoacoustic images in vivo","Y-Net's twin encoders beat single-input U-Net for PAI","Hybrid Y-Net lifts photoacoustic clarity over U-Net","Two-branch Y-Net improves photoacoustic reconstruction","Y-Net: raw plus beamformed beats standard deep learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1588,"prompt_tokens":980,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":596,"tokens_out":608,"duration_ms":5191,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:33:59.840587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training protocol on a matched set of real ex-vivo measurements with high-quality references, feeding both Y-Net and U-Net the identical delay-and-sum image with no ad hoc input replacement; if Y-Net's margin disappears or reverses, the reported advantage is an artifact of simulation transfer rather than the dual-input architecture.","supporting_citations":[{"cited_title":"k -Wave: MATLAB toolbox for the simulation and reconstruction of photoacoustic wave fields,","cited_arxiv_id":null,"evidence_quote":"The wave-propagation simulation engine used to generate the 4700 training and 400 test photoacoustic datasets on which all compared methods are trained and scored."},{"cited_title":"U -net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The U-Net architecture that is the main deep-learning baseline and whose encoder-decoder skeleton the two-arm Y-Net extends."},{"cited_title":"Image quality asse ssment: from error visibility to structural similarity,","cited_arxiv_id":null,"evidence_quote":"Defines the SSIM evaluation metric reported in the quantitative comparison table."}],"review_version":1}