{"id":"7337dc06-d4fb-41b1-8174-be674173fc9f","arxiv_id":"2411.09798","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A noise simulation pipeline and three causal deep-learning baselines that outperform standard video denoisers on fluorescence-guided surgery data by modeling and removing laser leakage light.","lead":"This paper builds a realistic computer noise model for fluorescence-guided surgery video that includes stray laser light, and tests three new neural-network denoisers against it. If the approach holds, surgeons could use dimmer, more specific contrast agents, because the denoisers remove a bias noise term that previous methods ignored.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'accurate noise simulation' claim is not established because Eq. (1) synthesizes LLL from the same LLL-PN used to train the denoisers, and real-data validation is qualitative only; LLL-PN misses specular/3D-dependent LLL, so the closed loop may bake in prediction errors.","rationale":"","tokens_in":21816,"tokens_out":6216,"duration_ms":67835,"concrete_test":"Quantitative simulator-to-real validation on held-out real LLL data with known zero fluorescence: use OL-LLL test frames (or a new capture of a non-fluorescent scene on the same FGS system) as real observations. Generate synthetic frames with Eq. (1) from the same RVs using the LLL-PN and the calibrated read-noise sampler at Sm=0, then compare (i) pixel-intensity distributions between real and synthetic frames (e.g., Wasserstein distance or patch statistics) and (ii) residual energy after subtracting the LLL-PN prediction, specifically in hand-shadow and specular-reflection regions identified by a simple detector. Then run the trained BL-RNN and NafNet32 on both real and synthetic frames and compare output residuals in those regions against the known zero signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is that the paper's central claim of an 'accurate noise simulation pipeline that includes LLL' (Abstract; Sec. III) is untested at the point where it matters. Eq. (1) generates noisy frames using \\hat{LLL} = fLLL(Rv_t), and the same predicted LLL is used for both training targets and simulated test inputs. The denoisers therefore learn to invert the LLL-PN's predictions, not to remove real LLL. Sec. III-A states that fLLL accounts for only 40% of the total L2 energy of noisy LLL frames and explicitly fails on specular reflections and 3D-dependent shadows (Fig. 4). Those structures are absent from the synthetic noise, so the trained models have never been asked to remove them; a biased LLL-PN would produce simulated data that looks plausible while systematically omitting the hardest real LLL components. The OL-Real evaluation has no ground truth, so qualitative agreement cannot show whether these components are handled. The paper's own caveat in Sec. III-A, that the RV-to-LLL prediction assumption 'will be important to check in human data,' concedes exactly this gap. A secondary concern is the implausibly low PSNR/SSIM for FastDVDnetC and OFDVDnetC in Table I (6.1/0.020 and 2.5/0.000), which suggests an implementation or training failure rather than an inherent property of temporal denoisers; this should be explained, though it does not by itself overturn the proposed baselines versus NafNet32 and BasicVSR++C.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses video denoising for fluorescence-guided surgery (FGS), where the fluorescence video is corrupted by shot noise, read noise, and a spatially varying laser leakage light (LLL) term. The authors propose a noise simulation pipeline that predicts LLL from a co-located reference video using a learned LLL-PN, generates synthetic noisy FV frames via Eq. (1), and trains denoising networks on these data. They introduce OL-2024, expanding existing FGS datasets, and propose three causal baseline models (BL-SW, BL-A&M, BL-RNN) built on a NAFNet backbone. They report that an image denoiser (NafNet32) outperforms adapted causal video denoisers (BasicVSR++, FastDVDnet, OFDVDnet) on their simulated FGS test set, and they analyze LLL robustness, ablations, and temporal behavior.","tokens_in":22100,"tokens_out":5572,"duration_ms":51605,"significance":"If the claims hold, the paper provides a valuable simulate-and-train pipeline for FGS, a new public dataset, and strong baselines for a clinically relevant problem. The surprising result that a single-image denoiser outperforms temporal video denoisers on LLL-dominated noise, with a Davis control showing the opposite on conventional noise, is an important empirical finding. The paper also includes useful robustness analyses (mlll, LLL-PN switching) and makes data/code available. However, the significance is contingent on whether the simulated LLL distribution matches real deployment conditions, which the current validation does not establish quantitatively.","major_comments":[{"comment":"The 'accurate noise simulation' claim (Abstract; Sec. III) is not validated at the point that matters. Eq. (1) generates both training and test noise using the same LLL-PN, fLLL(Rv_t), so the test set is in-distribution by construction; this does not demonstrate that the simulated noise matches real LLL. Sec. III-A reports that fLLL accounts for 40% of the L2 energy of noisy LLL frames and that specular reflections and 3D-dependent shadows (Fig. 4) are not predicted. Those structures are absent from the synthetic noise, so the trained models are never asked to remove them, while the OL-Real evaluation (Sec. V-E) has no ground truth and cannot show whether they are handled. The authors' own caveat in Sec. III-A that the RV-to-LLL assumption 'will be important to check in human data' concedes exactly this gap. I request a quantitative comparison of synthetic and real noise (e.g., patch statistics, noise power spectra, or detection rates on held-out OL-LLL frames with known LLL) and either simulation of the missing LLL structures or real-data evaluation with ground truth LLL.","section":"Sec. III / Eq. (1) / Sec. III-A / Fig. 4"},{"comment":"The reported PSNR/SSIM values for FastDVDnetC (6.1/0.020) and OFDVDnetC (2.5/0.000) are below what a trivial baseline (e.g., outputting the noisy input) would achieve, strongly suggesting an implementation or training failure in the adaptations rather than a genuine property of temporal denoisers. As written, the conclusion that 'SOTA video denoisers are unsuitable' (Sec. IV-A2) rests on a comparison in which the baseline models may not be functioning correctly. The authors should verify that their causal adaptations train correctly, for example by reporting training curves, comparing against the original non-causal models on the same data, or checking performance on clean/unnoisy inputs. Without this, the headline comparison of Table I is not interpretable.","section":"Table I / Sec. IV-A2"},{"comment":"The LLL robustness test with a second LLL-PN f2 changes only the training data split of the same OL-LLL set. It therefore measures sensitivity to LLL-PN weights, but not to the structural failure modes the paper identifies in Sec. III-A: specular reflections and shadows that depend on 3D scene geometry. The mlll slope measure is useful for quantifying sensitivity to Lm, but it cannot test whether models handle those missing components. To support the claim that the proposed baselines are robust to LLL in practice, the authors need a test that includes those structures, e.g., by injecting synthetic specular/shadow LLL into the simulation or by collecting real LLL frames with paired RV and evaluating where the denoisers fail.","section":"Sec. V-B / Table III"}],"minor_comments":[{"comment":"There are typographical errors: 'We simulate simulate noisy frames' in Sec. IV and 'LLL rmeoval' in Sec. V-C should be corrected.","section":"Sec. IV"},{"comment":"The notation '1/KS m' is ambiguous; please write it as a single fraction with K and Sm clearly separated, and state explicitly whether Quant is applied before or after the multiplication by 1/(K Sm).","section":"Eq. (1)"},{"comment":"The statement that fLLL 'accounts for 40% of the total energy' is ambiguous because it does not specify what fraction of the remaining energy is shot/read noise versus unmodeled LLL structure; please provide a decomposition of the residual energy.","section":"Sec. III-A"},{"comment":"The red polygon indicating 'realistic noise scenarios' is not defined in the text; please provide the criterion used to draw it.","section":"Fig. 7"},{"comment":"The training budget of 'maximum 2 weeks' is a hardware-specific statement; report iterations/epochs and early-stopping criteria instead of wall-clock time as the primary convergence measure.","section":"Sec. IV-C"},{"comment":"The dataset reference [12] is listed as 'Forthcoming 2024'; if the dataset and code are now available, update the citation to the final DOI.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a solid engineering contribution, but the review hinges on the shortage of quantitative validation of the LLL simulator. The same-generator circularity (Sec. III, Eq. 1) and the implausibly low PSNR of two comparison methods (Table I) are the main correctness risks. If the authors can add a distribution-level comparison of simulated vs real noise and fix the SOTA training, I would support publication. The paper's scope (medical imaging / computational imaging) fits the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read. The genuinely new piece is the LLL prediction network: using the reference video to synthesize laser leakage light as a spatially varying bias in a physically motivated noise model. That is a real step beyond OFDVDnet, which ignores LLL entirely. The authors also double the public dataset (OL-2024) and add calibration sets, and they ship three causal baselines with careful ablations. The Davis control is the right check: BasicVSR++C beats NafNet32 on conventional video denoising, which shows the FGS finding is not just a model failure. The result that a single-image denoiser can beat temporal video denoisers under LLL is well supported by the simulated-data experiments and is worth taking seriously.\n\nThe soft spots are real but not fatal. The main one is the circularity: training and test noise are generated with the same LLL-PN, so the models learn to invert that network's predictions, not necessarily real LLL. The paper partially addresses this with a second LLL-PN robustness test, and the authors are honest that the RV-to-LLL assumption needs checking in human data. Still, the abstract's 'accurate noise simulation' claim overreaches: the LLL-PN explains only 40% of the LLL energy and misses specular reflections and 3D shadows. The real-data evaluation is qualitative only, so we do not actually know how the closed loop performs on those omitted components. The implausibly low PSNR/SSIM for FastDVDnetC and OFDVDnetC (6.1/0.020 and 2.5/0.000) also suggests something went wrong in adaptation or training, not an inherent limit of those architectures; that needs a sentence of explanation. And there are no error bars anywhere, which matters when the claimed margins are a few dB.\n\nNone of that sinks the central argument. The relative ranking of the proposed baselines over the adapted SOTA methods is consistent across the noise sweep, and the ablation shows the RV and channel attention contribute meaningfully at high Lm. For anyone working on FGS or medical video denoising, this is a useful contribution with a clear path forward. It should go to peer review: a good referee can push for quantitative simulator-to-real validation (e.g., using the OL-Real frames with known zero-fluorescence regions as a proxy), error bars across seeds, and an explanation for the catastrophic PSNR values. I'd bring it to reading group.","headline":"A solid FGS denoising paper with a genuinely useful simulate-then-train pipeline and a surprising single-image-over-video result, though the 'accurate' noise simulation claim is only partially validated and a few numbers need explanation.","tokens_in":22691,"tokens_out":1702,"would_cite":true,"duration_ms":18365,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that fluorescence guided surgery video can be denoised by simulating laser leakage light from a co-located reference video, and that a simple recurrent network built on a single-image denoiser outperforms adapted…","keywords":["fluorescence guided surgery","video denoising","laser leakage light","noise simulation","reference video","recurrent neural network","Poisson shot noise","medical imaging"],"falsifier":"Acquire human FGS frames with the excitation laser on and no fluorophore present, as in the paper's OL-LLL protocol, and measure how much of the true LLL energy a LLL-PN trained on chicken-thigh data can predict; if the predicted energy share falls well below the 40% observed in OL-LLL, or if a denoiser trained on LLL-PN-simulated data fails to remove a hand or tool shadow on real human video, the transfer claim collapses.","tokens_in":21542,"feed_emoji":"🏥","tokens_out":5400,"duration_ms":47193,"temperature":0.7,"pith_summary":"Fluorescence guided surgery (FGS) produces dim videos corrupted not only by shot and read noise but also by laser leakage light (LLL), an additive bias whose brightness can rival the fluorescence signal itself. The paper tries to establish that LLL can be predicted from a co-located reference video and inserted into a noise simulator, so denoising networks can be trained on realistic synthetic data. It reports that a single-image denoiser, NafNet32, outperforms adapted state-of-the-art video denoisers on this task, and that a recurrent baseline built on it, BL-RNN, gives the best perceptual quality in the noise regime where FGS systems operate. If correct, the work gives the field a validated simulate-and-train pipeline, strong baselines, and a way to evaluate hardware choices over the signal/LLL tradeoff space.","feed_headline":"Single-frame denoiser beats video rivals in surgical fluorescence","feed_subtitle":"Reference video predicts laser leakage light, and a recurrent NafNet baseline leads state-of-the-art adaptations by 4.3 dB.","key_machinery":"Two objects carry the argument. The LLL-PN (a lightweight NafNet32 trained on OL-LLL with a median-seeking L1 loss) predicts the laser leakage frame from the reference video and supplies the bias term for Eq. (1); it is load-bearing because without it the simulated LLL distribution would not match the real one. The second is the noise simulator of Eq. (1), which combines clean fluorescence frames, predicted LLL, Poisson shot noise with calibrated camera gain K, and real dark frames for read noise. The third is BL-RNN, a recurrent network that feeds the previous denoised frame back into a NafNet32 U-net with ReLU activations, giving temporal integration without the unstable propagation of the BasicVSR++ style recurrence the paper tried first.","core_discovery":"The central claim is that FGS video denoising should be treated as a noise-formation problem with three terms—Poisson shot noise, sampled read noise, and spatially varying LLL—and that the LLL term can be approximated by a network fLLL that maps each reference-video frame Rv_t to a predicted LLL frame. The noisy fluorescence frame is modeled as F̃ = Quant(K Pois(Sm S + Lm LLL_pred) + R) / (K Sm), with calibrated gain K = 1/1764 and dark frames sampled from OL-Dark. Trained on OL-LLL in a Noise2Noise fashion with L1 loss, the LLL-PN accounts for 40% of the energy in noisy LLL frames, capturing large structure while missing specular reflections and 3D-dependent shadows. On simulated test data, NafNet32 reaches 34.8 dB PSNR versus 30.5 dB for causal BasicVSR++, while FastDVDnetC and OFDVDnetC fall below 7 dB; the paper attributes this to the bias and causality violations in FGS, and shows on the Davis dataset that the advantage reverses in conventional video denoising. The paper's proposed baselines combine NafNet blocks with sliding-window, align-and-merge, and recurrent temporal strategies, with BL-RNN giving the best LPIPS in the middle noise region and being the only model that removes LLL from a real no-fluorescence scene.","pith_inferences":["Editorial inference: if the RV-to-LLL correlation holds in human tissue, the same simulator could model autofluorescence, which is also correlated with the reference view; the paper notes this possibility.","Editorial inference: the repeated-frame experiment suggests current recurrent models underuse temporal information, so an architecture that explicitly forces long averages, like the align-and-merge baseline, may be necessary as signal drops.","Editorial inference: the qualitative success on the no-fluorescence real scene implies BL-RNN has learned a physical prior about LLL structure rather than memorizing simulated noise; a quantitative real-data test with injected fluorescence of known concentration would settle how far this extends.","Editorial inference: the gain calibration and dark-frame sampling procedures are transferable to other FGS cameras, so the simulate-and-train pipeline could be reproduced without the authors' hardware."],"forward_implications":["Simulated training data that includes LLL should let denoisers trained on mock-surgery videos transfer to real FGS systems, as long as the reference video remains informative.","Causal FGS denoising is feasible: the proposed baselines use only past and current frames, which fits the clinical workflow.","Conventional video denoisers designed for zero-mean, non-causal noise will need structural changes—especially access to the reference video and robustness to biased noise—before they can serve FGS.","The two-dimensional (Sm, Lm) noise space provides a cost landscape that could rank candidate hardware configurations by denoiser performance.","Dim fluorescent contrast agents become more clinically viable if post-capture denoising can recover signal near the noise floor."],"supporting_citations":[{"why":"Supplies the OL-2023 FGS dataset with paired reference and fluorescence video, and the OFDVDnet align-and-merge baseline used for comparison.","marker":"[8]"},{"why":"Supplies the NafNet architecture and the single-image denoiser that the paper finds outperforms video denoisers on FGS data.","marker":"[10]"},{"why":"Supplies BasicVSR++, the recurrent video denoiser that is adapted to causality and used as the main video baseline.","marker":"[9]"},{"why":"Provides the Noise2Noise training paradigm used to train the LLL-PN from noisy LLL frames without clean ground truth.","marker":"[22]"},{"why":"Motivates sampling real contiguous dark frames rather than Gaussian noise for realistic read noise simulation.","marker":"[21]"},{"why":"Provides the Quel calibration phantom used to estimate the camera gain K from the Poisson mean-variance relationship.","marker":"[13]"},{"why":"Supplies FastDVDnet, the sliding-window video denoiser adapted and retrained as a comparison baseline.","marker":"[33]"},{"why":"Provides the Davis dataset used as a control experiment showing the FGS-specific performance reversal between image and video denoisers.","marker":"[60]"}],"fun_headline_variants":["Fluorescence denoiser predicts laser leakage to cut noise","Single-frame net beats video rivals in surgical fluorescence","Reference video predicts bias noise to denoise surgery","Deep baselines tackle laser leakage in fluorescence surgery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reference video can predict the laser leakage light strongly enough that the simulated LLL distribution matches reality; the authors find this true in their mock-surgery data but flag that it must be checked on human tissue, and their LLL-PN only explains 40% of the energy in noisy LLL frames, missing specular reflections and 3D-dependent shadows.","fun_headline_variants_meta":{"raw":{"variants":["Fluorescence denoiser predicts laser leakage to cut noise","Single-frame net beats video rivals in surgical fluorescence","Reference video predicts bias noise to denoise surgery","Deep baselines tackle laser leakage in fluorescence surgery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000344,"raw_usage":{"total_tokens":1926,"prompt_tokens":1022,"completion_tokens":904,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":843}},"tokens_in":638,"tokens_out":904,"duration_ms":9136,"temperature":1.0,"reasoning_tokens":843,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:18:46.534041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire human FGS frames with the excitation laser on and no fluorophore present, as in the paper's OL-LLL protocol, and measure how much of the true LLL energy a LLL-PN trained on chicken-thigh data can predict; if the predicted energy share falls well below the 40% observed in OL-LLL, or if a denoiser trained on LLL-PN-simulated data fails to remove a hand or tool shadow on real human video, the transfer claim collapses.","supporting_citations":[{"cited_title":"Ofdvdnet: A sensor fusion approach for video denoising in fluorescence-guided surgery,","cited_arxiv_id":null,"evidence_quote":"Supplies the OL-2023 FGS dataset with paired reference and fluorescence video, and the OFDVDnet align-and-merge baseline used for comparison."},{"cited_title":"Rethinking noise synthesis and modeling in raw denoising,","cited_arxiv_id":null,"evidence_quote":"Motivates sampling real contiguous dark frames rather than Gaussian noise for realistic read noise simulation."},{"cited_title":"Indocyanine green matching phantom for fluorescence-guided surgery imaging system characterization and performance assessment,","cited_arxiv_id":null,"evidence_quote":"Provides the Quel calibration phantom used to estimate the camera gain K from the Poisson mean-variance relationship."},{"cited_title":"Bp-evd: Forward block-output propagation for efficient video denoising,","cited_arxiv_id":null,"evidence_quote":"Supplies FastDVDnet, the sliding-window video denoiser adapted and retrained as a comparison baseline."}],"review_version":1}