{"id":"06461b46-5dac-4b9c-a656-52a0d21e2743","arxiv_id":"2508.02025","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Frequency-domain lock-in filtering lifts fluorescence signal from noise, doubling penetration depth, cutting dye dose by 95%, and enabling 600 Hz artery/vein imaging in mice.","lead":"This paper describes a frequency-domain denoising method for fluorescence imaging that extracts the signal at a known modulation frequency, sharply reducing noise. It reports large contrast gains in mice, longer tumor visibility with the FDA-approved dye ICG, and real-time video at 600 Hz.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing concern is the unvalidated 'internal modulation' assumption: for non-periodic in vivo signals, the Fourier component at f0 = 1/T is dominated by windowing and drift rather than biology, so the Fig. 5-6 phase/amplitude gains and the 600 Hz temporal-resolution claim may be inflated.","rationale":"The reader's weakest_assumption already identifies the risk that tissue motion, autofluorescence, or physiological oscillations leak into the extracted Fourier bin, with special attention to the acquisition-window fundamental for non-periodic signals. My analysis agrees with that broad concern but locates the decisive failure mode more specifically: the internal-modulation variant, as described, projects a non-periodic signal onto the lowest Fourier mode of the finite window. That projection is not a denoised instantaneous measurement; it is a trend/window estimator. For the 600 Hz video, the growing-window construction means each output frame is a low-frequency coefficient over all previous frames, so the stated frame rate does not establish temporal resolution. This is a model-validation gap, not an internal contradiction: the external-modulation lock-in experiments in Fig. 2-3 are more defensible because a true periodic reference exists, and the static ICG tumor results in Fig. 4 do not depend on the internal-modulation assumption. Because the internal-modulation issue is correctable with a control experiment or a reanalysis using a sliding-window transform, the verdict stays CONDITIONAL rather than moving to REJECT or UNVERDICTED. The abstract's 2500/300-fold claim versus the in-text 734/98-fold value remains a separate reporting inconsistency, but it is secondary to the conceptual gap in the dynamic imaging branch.","tokens_in":13422,"tokens_out":8193,"duration_ms":103251,"concrete_test":"Recompute the internal-modulation FDD from the raw pixel time courses underlying Fig. 5d and Fig. 6e for three window lengths T, T/2, and 2T, and after adding a synthetic linear drift of amplitude comparable to the measured signal to one background ROI. If the artery-vein phase difference or the SBR/SNR enhancements shift by more than ~20%, or if drift alone produces a nonzero FDD 'vessel' image, then the f0 projection is driven by windowing/drift rather than by the biological dynamic, and the dynamic-imaging claims require reanalysis with a sliding-window or model-based method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing and least secure element is the 'internal modulation' variant used in Fig. 5e, Fig. 6c-6e, and the 600 Hz video. In this mode there is no external modulation; each pixel's time series over the full acquisition window T is Fourier-transformed and the component at f0 = 1/T is treated as the denoised signal. For non-periodic in vivo dynamics, such as ICG hepatic clearance or a one-time contrast-agent bolus, this f0 coefficient is not an instantaneous denoised measurement. It is a windowed projection of the total trend: for a monotonic rise or decay, its amplitude is set by the start and end values and its phase by the window endpoints. Any linear drift, respiratory or peristaltic motion, or baseline shift falls into the same f0 bin and is indistinguishable from the intended biology. The paper provides no no-injection or static control and no estimate of motion leakage at f0. Consequently, the reported gains in Fig. 5f-5g, the vessel-over-liver separation, and especially the artery/vein phase split in Fig. 6d-6e may be windowing artifacts rather than physiological signal. The '600 Hz real-time video' also conflates output frame rate with temporal resolution: each FDD output is computed from all previously acquired frames, so the integration window grows over time; frame k is a low-frequency coefficient over k frames, not a 1.67-ms instantaneous sample. The claimed 5-fold temporal-resolution improvement over prior 100 Hz work is therefore unsupported unless effective resolution is defined by the evolving window, not the display rate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents frequency-domain denoising (FDD) as a post-processing technique for in vivo NIR-II fluorescence imaging. Excitation light is periodically modulated, a time series of frames is acquired, and each pixel's Fourier coefficient at the modulation frequency is used as the denoised signal, with the phase optionally used for additional contrast. The authors report large SBR/SNR enhancements in intralipid phantoms, 95% reductions in contrast-agent dosage or excitation intensity for mouse vascular imaging, improved tumor-margin and vessel visibility with ICG in the 1,400–1,700 nm band, a 600 Hz 'real-time' video of contrast-agent diffusion with artery/vein differentiation, and motion-blur suppression via an increased frame rate evaluated by SSIM. The central idea is straightforward and the external-modulation phantom experiments are the most controlled part of the study.","tokens_in":13744,"tokens_out":4214,"duration_ms":44970,"significance":"If the claims hold, FDD would be an inexpensive, broadly applicable computational enhancement for NIR-II imaging, potentially making FDA-approved ICG usable in a spectral region where no approved NIR-II agents exist. The external-modulation phantom depth data (Fig. 2) provide a clear, controlled demonstration of the core mechanism, and the proposed extension to ICG-based surgical navigation is clinically motivated. However, the most clinically interesting claims—ICG vessel-over-liver separation, artery/vein phase discrimination, and the 600 Hz temporal-resolution gain—rest on an 'internal modulation' variant that is not validated against the windowing and motion artifacts inherent to Fourier analysis of non-periodic, drift-dominated signals. The abstract's headline enhancement figures (2,500-fold SBR, 300-fold SNR) are also not supported by the numbers reported in the text (734-fold and 98-fold at the 6 mm penetration limit). The technique is potentially significant, but the load-bearing validation is currently incomplete.","major_comments":[{"comment":"The abstract claims that FDD improves SBR and SNR by more than 2,500-fold and 300-fold, respectively, but the text reports only 734-fold and 98-fold improvements at the 6 mm penetration limit of the original image (Fig. 2e, 2f, right axes). No other measurement in the manuscript supports the larger figures. The abstract should either cite the specific experimental condition that yields 2,500-fold/300-fold or be corrected to match the reported data.","section":"Abstract; Results, Fig. 2e-f"},{"comment":"The SSIM evaluation for motion-blur suppression is self-referential: the reference image used for SSIM calculation is itself an FDD image with the same parameters and a constant spatial position. Comparing both the original and FDD images against an FDD reference measures similarity to the FDD output, not fidelity to a true ground truth. The resulting 21-fold SSIM enhancement is therefore inflated by construction. A motion-free conventional image (or an independent ground truth, e.g., a sharp image of a static phantom) should be used as the reference.","section":"Extended Data Fig. 1d; 'Real-time imaging of spatially varying signals'"},{"comment":"For the 'internal modulation' variant (used for ICG hepatic clearance, the 600 Hz video, and the artery/vein phase images), there is no external modulation: the Fourier component at f0 = 1/T of the full acquisition window is treated as the denoised signal. For non-periodic dynamics such as a one-time bolus or monotonic hepatic clearance, this coefficient is a windowed projection of the total trend, not an instantaneous measurement; linear drift, respiratory or peristaltic motion, and baseline shifts all fall into the same f0 bin and are indistinguishable from the intended biology. The manuscript provides no no-injection control, no static-phantom control, and no estimate of motion leakage at f0. This undermines the reported >100-fold SBR/SNR gains for liver-overlapping vessels (Fig. 5f, 5g), the vessel-over-liver separation, and the artery/vein phase split in Fig. 6d-6e, which can arise from window endpoints rather than physiology. The authors should validate the method on known synthetic dynamics and on a static control, and quantify spectral leakage at f0.","section":"Results, 'Video contrast and frame rate improvement of temporally varying signals', Fig. 5e, Fig."},{"comment":"The phrase '600 Hz real-time video' conflates output frame rate with temporal resolution. The Methods state that after each new frame an FDD calculation is performed based on all previously collected frames, so the integration window grows over time; frame k is a low-frequency coefficient over k frames, not a 1.67-ms instantaneous sample. The claimed 5-fold temporal-resolution improvement over prior 100 Hz work is therefore unsupported unless the effective temporal bandwidth of the FDD output is characterized. The 'phase' images derived from the same windowed transform inherit the same ambiguity.","section":"Methods, 'Real-time imaging of temporally varying signals'; Results, 'Video contrast and frame rate improvement of…"}],"minor_comments":[{"comment":"Equation (2) defines SNR as (Signal - Bckground) / standard deviation of Background; 'Bckground' should be 'Background'.","section":"Results, Eq. (2)"},{"comment":"The duration of distinguishable tumor imaging is inconsistently stated: the Introduction says 'seven-fold increase', Results say 'four times those of the original images', and the Discussion says 'three-fold extension of tumor margin duration'. These numbers should be harmonized.","section":"Introduction and Discussion"},{"comment":"In the caption for Fig. 5, 'local magnifications of the live' should presumably be 'liver'.","section":"Fig. 5 caption"},{"comment":"In the Methods, 'the total FDD acquisition times of tumor and vascular imaing were 320 s and 480 s' contains a typo ('imaing' should be 'imaging').","section":"Methods"},{"comment":"The figure captions use 'Florescence imaging' instead of 'Fluorescence imaging'.","section":"Fig. 3, Fig. 4 captions"},{"comment":"The abstract states 'we achieved a SBR far exceeded the Rose criterion', but the text (Results) reports SNRs of 45 and 65 as exceeding the Rose criterion; SBR and SNR should be stated consistently.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from a careful reconciliation of abstract claims with reported numbers and from independent validation of the internal-modulation variant. The external-modulation core is plausible, but the paper's most impactful claims (ICG vessel/liver separation, artery/vein phase, 600 Hz temporal resolution) rest on a method that currently lacks the controls needed to rule out windowing and motion artifacts. These issues are addressable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new: FDD is the optical lock-in idea from refs 35/36, applied to in vivo NIR-II fluorescence, with an internal-modulation variant for non-periodic dye kinetics, phase-based arterial/venous segmentation, and FPGA processing at 600 Hz. The external-modulation experiments are the strongest part: intralipid phantom depth curves show 734-fold SBR and 98-fold SNR at the 6 mm limit, about 2x penetration, and those numbers are computed from a defined background ROI, so they are independently checkable. The ICG tumor-margin duration extension and the 95% dose/intensity reduction are clinically meaningful if they hold up. The citation pattern is honest; the frequency-domain roots are attributed.\n\nNow the soft spots. The abstract's headline claims (≥ 2,500-fold, ≥ 300-fold) do not match the in-text values (734, 98). That is a correction, not a fatal flaw, but it must be fixed. The SSIM comparison in Extended Data Fig. 1d uses an FDD image as its reference, so the 21-fold gain partly measures self-agreement; use an unprocessed high-SNR static frame or a clear external reference. The load-bearing issue is the internal-modulation mode (Fig. 5e, Fig. 6, the 600 Hz video). Taking the f0 = 1/T Fourier bin of a non-periodic bolus response makes that coefficient a windowed projection of the whole trend, not an instantaneous measurement. Respiratory motion, drift, and peristalsis can land in the same bin. The liver-overlap separation and artery/vein phase split could still be real, but without a no-injection control, a static-tissue control, or a motion estimate at f0, the phase contrast is not yet quantitative. And the 600 Hz output display rate is not temporal resolution; each frame is a coefficient over an expanding window. The claimed 5x improvement over prior 100 Hz work needs a definition of effective resolution, not frame count.\n\nWho is it for: people working on NIR-II imaging, surgical navigation, or lock-in fluorescence microscopy will get value. The engineering is serious and the external-modulation data are solid enough to referee. My recommendation: send it to review with a request for the internal-modulation controls, an abstract reconciled with the text, and a clearer SSIM baseline. It should not be desk-rejected.","headline":"The external-modulation FDD results are plausible and worth a real look, but the headline gains are inflated relative to in-text numbers and the 'internal modulation' variant needs controls before it carries the 600 Hz claim.","tokens_in":14346,"tokens_out":2595,"would_cite":false,"duration_ms":29423,"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 frequency-domain denoising step lifts in vivo NIR-II fluorescence contrast by more than 2,500-fold and makes FDA-approved ICG a practical deep-tissue imaging agent.","keywords":["frequency-domain denoising","NIR-II fluorescence imaging","indocyanine green","signal-to-background ratio","penetration depth","real-time vascular imaging","Fourier transform","in vivo imaging"],"falsifier":"Record a time series from a dye-free or dead mouse while the excitation is modulated and the subject is breathing or being moved; a persistent peak at the modulation frequency in regions with no contrast agent would show that tissue background or motion has a coherent component in that bin, inflating the reported SBR/SNR gains.","tokens_in":13178,"feed_emoji":"🔬","tokens_out":6557,"duration_ms":71454,"temperature":0.7,"pith_summary":"FDD fluorescence imaging tries to solve a practical bottleneck: near-infrared-II fluorescence is faint in living tissue because background and detector noise drown the signal, and no FDA-approved agent emits strongly there. The paper's proposal is to modulate the excitation laser periodically, record many cycles of frames, and keep only the Fourier component at the modulation frequency for each pixel. This single processing step is claimed to lift signal-to-background ratio by more than 2,500-fold and signal-to-noise ratio by more than 300-fold, which would double penetration depth, cut required dye dose or laser power by 95%, and bring the FDA-approved dye ICG into the NIR-II window for tumor-margin and vascular imaging. It also enables a 600 Hz video of contrast-agent diffusion and artery/vein differentiation. If true, a pure computational method could accelerate clinical NIR-II surgery navigation without developing new contrast agents.","feed_headline":"A Fourier filter lifts NIR-II fluorescence contrast 2,500-fold","feed_subtitle":"Frequency-domain denoising doubles penetration depth and makes FDA-approved ICG practical for deep-tissue imaging.","key_machinery":"The central object is the per-pixel Fourier coefficient at the modulation frequency $\\omega_0$ (or, in the 'internal modulation' mode, at the fundamental frequency $f_0 = 1/T$ of the acquisition window). For each pixel, the amplitude of this coefficient forms the denoised image, and the phase forms a complementary image that distinguishes structures by the timing of contrast-agent arrival. The operation is optical lock-in detection translated to in vivo imaging: the periodic signal adds coherently across many frames while random broadband noise does not, so a single Fourier bin isolates the signal.","core_discovery":"On the paper's own terms, the discovery is that frequency-domain denoising makes in vivo fluorescence imaging dramatically more sensitive by exploiting the temporal signature of the signal rather than its brightness. Fluorescence is excited by an intensity-modulated source, so each pixel's time trace is periodic, while the in vivo background and detector noise are non-periodic and spread across the spectrum. Taking the Fourier transform and reading only the bin at the modulation frequency (or, for dynamic signals, at the fundamental of the acquisition window) suppresses the broadband noise and recovers the signal, and the phase of that Fourier component carries arrival-time information. The paper reports that this yields SBR/SNR improvements exceeding 2,500-fold and 300-fold, doubles intralipid penetration depth, permits a 95% reduction in contrast-agent dose or excitation intensity, makes ICG usable in the 1,400–1,700 nm window with SNRs above the Rose criterion, and produces real-time 600 Hz videos in which arteries and veins are separated by phase.","pith_inferences":["If the broadband-noise assumption holds, the same processing could be transferred to other periodic-excitation imaging modalities with stationary scenes, such as photoacoustic imaging or conventional NIR-I fluorescence, without changing the contrast agent.","The phase image is effectively a map of local arrival-time delays, so it could be processed further to estimate blood-flow velocities or perfusion gradients across the field, beyond the paper's artery/vein demonstration.","The technique sets a fair benchmark for evaluating new NIR-II agents: if an agent's advantage over ICG disappears after FDD processing, its real contribution may be brightness or photostability rather than noise suppression.","A practical caveat follows from the assumption: any motion, heartbeat-synchronous tissue movement, or physiological oscillation that locks to the modulation frequency would enter the extracted bin and inflate apparent contrast, so motion correction or gating would be needed in less-anesthetized or freely moving subjects."],"forward_implications":["Because FDD needs no photoswitchable fluorophore, it should work with any contrast agent that emits fluorescence, including agents already approved for humans.","The FDD technique extends the time window over which ICG tumor margins remain distinguishable to hours, which should reduce repeat injections during surgical navigation.","A 95% reduction in contrast-agent dosage or excitation intensity would lower toxicity and cost, addressing a main barrier to NIR-II agent approval.","The 600 Hz frame rate and phase-based artery/vein separation offer a route to motion-blur-free, functional vascular imaging during surgery.","FDD provides a greater-than-10-fold contrast increase when total acquisition time is extended threefold, giving more tolerance for low-quantum-yield contrast agents."],"supporting_citations":[{"why":"Supplies DIPT-ICF, the AIE nano contrast agent used for intralipid penetration, whole-body vascular, and 600-Hz diffusion experiments.","marker":"[21]"},{"why":"Establishes optical lock-in detection imaging microscopy, the frequency-domain signal-recovery principle FDD adapts to in vivo imaging.","marker":"[35]"},{"why":"Demonstrates selective fluorescence signal recovery via modulated fluorophores, another precursor of the FDD approach.","marker":"[36]"},{"why":"Provides the Rose criterion (an SNR threshold of 4) used to argue that FDD-processed ICG images are clear enough for navigation.","marker":"[39]"},{"why":"Shows that ICG, an FDA-approved NIR-I dye, has a shortwave-infrared emission tail, the basis for pushing ICG into NIR-II with FDD.","marker":"[42]"},{"why":"Supplies the previous 100-Hz NIR-II dynamic-imaging benchmark and the phase/principal-component style separation that FDD's 600-Hz video improves upon.","marker":"[48]"},{"why":"Defines SSIM, used to quantify the motion-blur improvement in spatially varying FDD imaging.","marker":"[50]"}],"fun_headline_variants":["FDD imaging: 2,500-fold contrast boost, 600 Hz video","Frequency denoising makes NIR-II fluorescence 2,500x clearer","Temporal filtering lifts NIR-II fluorescence 2,500-fold","Fourier trick doubles penetration and cuts dye dose 95%","Fluorescence denoising: 2,500x signal gain, real-time video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that in vivo background and detector noise are effectively random and broadband, so essentially nothing except the modulated fluorescence occupies the extracted Fourier bin.","fun_headline_variants_meta":{"raw":{"variants":["FDD imaging: 2,500-fold contrast boost, 600 Hz video","Frequency denoising makes NIR-II fluorescence 2,500x clearer","Temporal filtering lifts NIR-II fluorescence 2,500-fold","Fourier trick doubles penetration and cuts dye dose 95%","Fluorescence denoising: 2,500x signal gain, real-time video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3259,"prompt_tokens":982,"completion_tokens":2277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2177}},"tokens_in":598,"tokens_out":2277,"duration_ms":19358,"temperature":1.0,"reasoning_tokens":2177,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:14:22.178208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a time series from a dye-free or dead mouse while the excitation is modulated and the subject is breathing or being moved; a persistent peak at the modulation frequency in regions with no contrast agent would show that tissue background or motion has a coherent component in that bin, inflating the reported SBR/SNR gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies DIPT-ICF, the AIE nano contrast agent used for intralipid penetration, whole-body vascular, and 600-Hz diffusion experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes optical lock-in detection imaging microscopy, the frequency-domain signal-recovery principle FDD adapts to in vivo imaging."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates selective fluorescence signal recovery via modulated fluorophores, another precursor of the FDD approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Rose criterion (an SNR threshold of 4) used to argue that FDD-processed ICG images are clear enough for navigation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that ICG, an FDA-approved NIR-I dye, has a shortwave-infrared emission tail, the basis for pushing ICG into NIR-II with FDD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the previous 100-Hz NIR-II dynamic-imaging benchmark and the phase/principal-component style separation that FDD's 600-Hz video improves upon."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Defines SSIM, used to quantify the motion-blur improvement in spatially varying FDD imaging."}],"review_version":1}