{"id":"28b34725-e983-441b-a228-05fe45e3971c","arxiv_id":"2607.05824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A deterministic diffusion prior (ResShift-ODE) inverts a known numerical-aperture band-limiting operator to transfer low-NA holotomography volumes to high-NA-equivalent RI volumes in five denoiser evaluations.","lead":"This paper presents a computational method that upgrades low-resolution 3D microscope images to high-resolution equivalents using a deterministic diffusion model, without changing the microscope hardware. A smart generalist might read it because it could enable high-throughput biological screening with high-resolution quality at low cost.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The evaluation is circular: training and test pairs come from the same analytic operator, headline metrics are dominated by the preserved passband, and the cone-preservation 'physical check' is guaranteed by construction — so no quantitative evidence currently tests real NA-transfer performance.","rationale":"The reader's verdict of CONDITIONAL with MODERATE confidence is correct. The identified concern — that the emulation operator may not faithfully represent physical low-NA acquisition — is the single most load-bearing premise, and the paper itself acknowledges it explicitly in §4. My additional observation is that the current quantitative evidence is weaker than it appears at first reading: the headline voxel-error statistics measure passband preservation (the easy part), and the cone-preservation behavior is guaranteed by construction rather than being an independent physical check. However, the paper is unusually transparent about these limitations, the mathematical formulation is sound, and the method represents a legitimate application of established diffusion-inverse-problem techniques to a specific microscopy problem. The novelty claim is appropriately scoped — the paper claims the operator-plus-modality formulation as new, not the diffusion sampler itself. No mathematical errors or internal inconsistencies were identified in the derivation of Eqs. (1)–(3) or the equivalence to flow matching. The three test volumes (one per cell type) are insufficient for population-level statistics, but the paper does not claim population-level results. The conflict of interest (Tomocube employees/founders) is disclosed and does not affect the mechanical assessment. The verdict should remain CONDITIONAL: the method is sound in principle but requires (1) quantitative out-of-band recovery metrics even on emulated data, and (2) registered physical low-NA/high-NA validation to move toward ACCEPT.","tokens_in":14274,"tokens_out":3614,"duration_ms":256113,"concrete_test":"Compute a quantitative out-of-band recovery metric on the three held-out test volumes: specifically, the ratio of spectral power in the lateral annulus K_H \\ K_L between the model output and the high-NA target, normalized by the same ratio for the low-NA input. If this ratio for the output is not substantially above the input ratio (e.g., <2× improvement), the method is not meaningfully recovering the missing band even under the emulated operator, which would weaken the claim independent of physical validation. This requires only reprocessing the existing test volumes and can be done without new data acquisition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies emulation fidelity as the load-bearing premise. I agree and want to sharpen the specific mechanism. The paper trains on pairs generated by operator O (Eq. 1: Fourier-support restriction + nonlinear TV reconstruction) and tests on held-out pairs generated by the same O. The headline metric — >99% of voxels within RI error 0.002–0.003 — is, as the paper itself states, 'dominated by the preserved low-NA passband, where recovery reduces to passing through measured frequencies.' This means the headline numbers measure how well the model preserves frequencies it was given, not how well it recovers the annihilated annulus (K_H \\ K_L) — which is the actual task. Meanwhile, the Fourier-domain 'physical check' in §3.3 (axial missing cone not filled) is not an independent validation: the paper acknowledges it is guaranteed by construction, since 'every pair is produced by the same operator that annihilates the axial cone.' So the two pieces of evidence that look like validation — the voxel-level error statistics and the cone-preservation behavior — respectively measure the easy part of the task and a property built into the training data. The only evidence for out-of-band recovery is the qualitative Fourier projection in Fig. 3d, with no quantitative metric for spectral power recovery in the lateral annulus. This does not invalidate the method, which is sound in principle, but it means the current evidence supports only 'the model can invert a known analytic operator while preserving passband content,' not the broader claim of NA transfer performance. The paper is transparent about all of this, which is why CONDITIONAL is appropriate rather than REJECT.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The manuscript presents ResShift-ODE, a deterministic diffusion-prior framework for transferring low-NA refractive-index (RI) tomograms to high-NA-equivalent volumes in holotomography (HT). The method extends residual-shifting diffusion to 3D RI data, reformulates the reverse process as a probability-flow ODE for reproducible five-step inference, and is evaluated on emulated low-NA/high-NA pairs generated by an analytic Fourier-support restriction operator (Eq. 1). The paper reports RI errors of 0.002–0.003 for >99% of voxels on three held-out test volumes (one per cell type), demonstrates that the axial missing cone is not artificially filled, and shows substantial speedup over a 1000-step DDPM baseline. The mathematical formulation is internally consistent, the Fourier-domain analysis of missing-cone preservation is a sound physical check, and the paper is notably transparent about the scope and limitations of its evaluation.","tokens_in":15053,"tokens_out":1483,"duration_ms":194163,"significance":"The problem formulation—NA transfer as a diffusion-prior inverse problem under an explicit band-limiting operator—is well-motivated and distinct from prior artifact-suppression work in HT. The deterministic ODE sampler yielding reproducible inference in five denoiser evaluations is a practical contribution for screening workflows. The Fourier-domain cone-preservation check provides a falsifiable, modality-agnostic acceptance criterion. The bead-phantom validation of the emulation operator (Fig. 2a,b) and the multi-cell-type training corpus are commendable design choices. However, the quantitative headline results are computed on emulated pairs generated by the same analytic operator used in training, which the paper itself acknowledges measures operator-inversion fidelity rather than real-world NA-transfer performance. This limits the current evidence to a proof-of-principle demonstration under controlled physics.","major_comments":[{"comment":"§3.2–3.3: The headline metric (>99% of voxels within RI error 0.002–0.003) is, as the paper transparently states, 'dominated by the preserved low-NA passband, where recovery reduces to passing through measured frequencies.' The decisive quantity—recovery accuracy in the lateral annulus outside K_L but within K_H—is examined only qualitatively in Fig. 3d. No quantitative metric for spectral power recovery in this annulus is provided. Without such a metric, the central claim of NA transfer (as opposed to passband preservation) is supported by visual inspection alone. A quantitative out-of-band recovery metric (e.g., spectral power ratio in the annulus K_H ∖ K_L, or annulus-restricted PSNR/SSIM) would substantially strengthen the evidence and is necessary for the claim to be load-bearing.","section":null},{"comment":"§3.1: The evaluation uses three held-out test volumes (one per cell type), and all three cell types are represented in training. The paper acknowledges that 'genuine transfer to held-out cell types remains untested.' While this is honestly stated, it means the cross-cell-type generalization claim is not statistically supported. A leave-one-class-out evaluation, even on a single held-out class, would provide meaningful evidence that the prior is anchored to the shared forward operator rather than to cell-type-specific image statistics. If this is infeasible with the current dataset size, the claim of cross-cell-type transfer should be further qualified in the abstract and conclusion.","section":null},{"comment":"Table 1, HepG2 row: ResShift-ODE achieves PSNR 40.068 dB versus 3D U-Net's 39.677 dB and 3D ResShift's 39.923 dB. The paper correctly notes that differences below ~1 dB should not be interpreted as statistical ranking. However, the four-model comparison is performed on a single volume, making it impossible to assess whether ResShift-ODE offers a meaningful accuracy advantage over the simpler 3D U-Net. Given that the U-Net is ~5.8× faster, the practical case for ResShift-ODE rests almost entirely on the Fourier-domain cone-preservation behavior and the measurement-anchoring argument. The paper should more explicitly frame the contribution as a generative-prior framework with interpretable spectral behavior rather than a metric improvement over deterministic regression.","section":null}],"minor_comments":[{"comment":"§2.2: The bead-phantom validation (Fig. 2a,b) compares emulated and measured low-NA profiles qualitatively. A quantitative comparison metric (e.g., RI-profile RMSE or FWHM ratio) would strengthen the operator-fidelity claim.","section":null},{"comment":"§2.3, Eq. (2): The scalar noise scale κ is stated to be in RI units, with its numerical value deferred to Supplementary S2. Including the value in the main text would help readers assess the noise regime.","section":null},{"comment":"§3.4, Table 1: The relative wall-clock cost for the Yeast and K562 rows is listed as ~5.8×, identical to the HepG2 row, but the wall-clock time is also listed as 217.98 s for all three. If these are the same measurement, this should be clarified; if they are independent measurements that happen to coincide, a note would help.","section":null},{"comment":"Fig. 3: The Fourier-domain amplitude projections in panel (d) are shown for one cell type (HepG2). Including the corresponding projections for yeast and K562, or at minimum confirming in text that the same behavior holds, would aid interpretation.","section":null},{"comment":"§4, Discussion: The sentence beginning 'A residual-shifted diffusion prior is a natural fit...' could benefit from a more specific justification of why the residual-shifting formulation is preferable to standard DDPM conditioning for this particular inverse problem, beyond the computational advantage of fewer steps.","section":null},{"comment":"Reference [28] is dated 2026; if this is a preprint, the arXiv identifier or DOI should be included. Similarly, reference [51] is dated 2026 with no volume/page information.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent about its limitations, which is commendable but also creates a tension: the abstract and conclusion headline the >99% voxel accuracy and cross-cell-type transfer, while the body text acknowledges these are single-volume, emulated-pair, passband-dominated results. The authors should align the framing of the headline claims with the caveats they themselves identify. The core method is sound and the problem formulation is novel; the revision should focus on adding at least one quantitative out-of-band recovery metric and, if feasible, a leave-one-class-out test. I do not view the emulation-based evaluation as circular in a misleading sense—the paper is explicit that it tests operator inversion—but the current evidence does not yet support the stronger claim of real NA-transfer performance."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. All three major comments are well-taken. We address each below.","responses":[{"response":"The referee is correct. The headline voxel-error metric is dominated by the preserved low-NA passband, and the decisive out-of-band recovery is currently supported only qualitatively. We will add a quantitative annulus-restricted metric in the revised manuscript. Specifically, we will compute the spectral power ratio |X̂(k)|² / |X_H(k)|² averaged over the lateral annulus K_H ∖ K_L (and separately over K_L for reference), as well as an annulus-restricted PSNR, for each held-out test volume. These quantities are computable from the existing test data and Fourier transforms already used for Fig. 3d, so no new acquisitions are required. We agree that without such a metric the NA-transfer claim is not load-bearing, and we will present the annulus metric alongside the full-volume metrics in a revised Table 1 and/or an expanded §3.3.","revision_made":"yes","referee_comment":"§3.2–3.3: No quantitative metric for spectral power recovery in the lateral annulus K_H ∖ K_L; the central NA-transfer claim rests on visual inspection of Fig. 3d alone."},{"response":"We agree that the current evaluation does not test genuine cross-cell-type transfer. We will attempt a leave-one-class-out experiment: for each of the three cell types, we will retrain on the remaining two classes (approximately 34–40 paired volumes depending on the split) and evaluate on the held-out class. We acknowledge that with only two classes in training and ~34–40 volumes, the statistical power will be limited and the results may show degraded performance relative to the multi-class model. However, even a single leave-one-class-out result per class would provide a first indication of whether the prior is anchored to the shared forward operator or to cell-type-specific statistics. If the leave-one-class-out results are inconclusive due to the small training set, we will state this explicitly and further qualify the cross-cell-type transfer claim in the abstract and conclusion, as the referee suggests. In either case, the abstract and conclusion will be revised to avoid implying statistically supported cross-cell-type generalization.","revision_made":"partial","referee_comment":"§3.1: Cross-cell-type generalization is not statistically supported because all three cell types appear in training; a leave-one-class-out evaluation is needed, or the claim should be further qualified."},{"response":"We agree with this framing. The manuscript already notes that sub-1-dB differences on a single volume should not be interpreted as a statistical ranking, and the Discussion states that ResShift-ODE should not be viewed as a scalar-metric replacement for a deterministic regressor. However, the referee is right that this framing should be more prominent. In the revision we will: (i) state explicitly in the abstract that the contribution is a generative-prior framework with interpretable spectral behavior, not a PSNR improvement over deterministic regression; (ii) move the 'not a metric replacement' qualification from the Discussion into the Results section adjacent to Table 1; and (iii) emphasize that the practical case for ResShift-ODE rests on the Fourier-domain cone-preservation behavior and measurement-anchoring argument, not on scalar metric gains over the faster U-Net.","revision_made":"yes","referee_comment":"Table 1, HepG2 row: ResShift-ODE's PSNR advantage over 3D U-Net is below 1 dB on a single volume; given the U-Net is ~5.8× faster, the paper should more explicitly frame the contribution as a generative-prior framework with interpretable spectral behavior rather than a metric improvement over deterministic regression."}],"tokens_in":14168,"tokens_out":1305,"duration_ms":49771,"standing_objections":["The leave-one-class-out evaluation (Comment 2) will be attempted, but with only ~34–40 training volumes per fold, the results may not be statistically conclusive. If performance degrades substantially, we will not be able to distinguish insufficient training data from failure of operator-anchored generalization. We will report the results transparently regardless of outcome."]},"desk_editor":{"model":"glm-5.2","letter":"Short version: the paper formulates low-NA-to-high-NA RI transfer in holotomography as a diffusion-prior inverse problem under an explicit Fourier-support operator, solves it with a deterministic 5-step ODE sampler, and is unusually transparent that its quantitative results measure operator inversion, not real-world NA transfer. The formulation is new and the physics is sound. The validation doesn't yet support the broader claims, but the authors say so themselves, repeatedly and clearly. It deserves a serious referee. What's genuinely new is the problem formulation, not the sampler — the authors explicitly state the ODE is equivalent to conditional Flow-Matching / rectified flow, and the residual-shifting scheme extends prior 2D work (ResShift). The contribution is scoping NA transfer as band completion under a known coherent forward operator, with measurement-anchored behavior that preserves the axial missing cone. The Fourier-domain analysis in §3.3 is a real physical check: the model populates the lateral annulus (K_H minus K_L) while leaving the axial cone near the low-NA input level. This asymmetry is explained mechanistically — the training construction never presents cone energy to synthesize — and it serves as a falsifiable hallucination safeguard. The bead phantom comparison (Fig. 2a,b) validating the emulation operator against a physically acquired low-NA measurement is a nice piece of grounding. The five-step deterministic inference at ~218 s per volume is practical for screening workflows. Now the soft spots, which are real but proportionate. The headline metric — >99% of voxels within RI error 0.002–0.003 — is, as the paper itself states, dominated by the preserved low-NA passband. It measures passband fidelity, not recovery of the annihilated lateral annulus, which is the actual task. There is no quantitative metric for spectral power recovery in K_H minus K_L; the only evidence there is qualitative (Fig. 3d). The cone-preservation behavior, while physically correct, is guaranteed by construction rather than independently validated. The test set is one volume per cell type, so no population-level statistics. And the load-bearing premise — that the analytic Fourier-support operator captures the dominant information loss of physical low-NA optics — omits pupil apodization, partial coherence, and aberration. The paper calls this an upper bound on fidelity, which is fair. The stress-test note is correct on all these points, but I'd emphasize that the paper acknowledges every one of them in its own text. This isn't a case of hidden circularity; it's a case of a method paper that ships with honest scoping and needs one more validation step — registered physical low-NA/high-NA pairs with quantitative out-of-band metrics — to close the loop. Recommendation: send to review. The formulation is a legitimate methodological advance, the physics is grounded, and the limitations are tractable. A good referee should push for (1) at least one physical low-NA/high-NA validation pair, (2) a quantitative spectral recovery metric in the lateral annulus, and (3) a leave-one-cell-type-out transfer test. If the authors can address even (1) and (2), this clears the bar.","headline":"Sound formulation, honest about its own limits, but validation is operator-inversion not NA transfer — needs physical data to earn its claims.","tokens_in":15066,"tokens_out":1116,"would_cite":false,"duration_ms":62460,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.30.Wb","87.57.-s","42.40.-i"],"model":"glm-5.2","headline":"Diffusion prior recovers high-NA 3D refractive index volumes from low-NA holotomography in five steps","keywords":[],"falsifier":"Apply the trained model to a physically acquired low-NA holotomography volume of a specimen for which a registered high-NA acquisition of the same specimen exists. If the recovered volume fails to match the high-NA reference—particularly in the lateral high-frequency annulus outside the low-NA support—then the method inverts the analytic operator but does not achieve physical NA transfer.","tokens_in":14355,"feed_emoji":"🔬","tokens_out":1022,"duration_ms":70293,"temperature":0.7,"pith_summary":"The paper presents ResShift-ODE, a method that uses a deterministic diffusion prior to transfer low-numerical-aperture (NA) refractive-index (RI) tomograms to high-NA-equivalent 3D volumes without changing microscope hardware. The authors formulate NA transfer as an inverse problem: a low-NA measurement is modeled as a Fourier-support restriction of the true high-NA volume, meaning certain spatial frequencies are not merely blurred but entirely absent. To recover them, the method anchors a residual-shifting diffusion process to the measured low-NA volume and reformulates the reverse sampling as a probability-flow ordinary differential equation (ODE), yielding deterministic, reproducible output in five network evaluations. On held-out emulated test volumes spanning yeast, HepG2 cells, and K562 lymphoblasts, the recovered volumes match high-NA references with refractive-index errors of 0.002–0.003 for over 99% of voxels. Crucially, Fourier analysis shows the method populates the lateral high-frequency band that the low-NA aperture truncates while leaving the axial missing cone unfilled, providing a physics-grounded check against hallucinated structure. The central object carrying the argument is the combination of the explicit NA-limited Fourier-domain forward operator and the measurement-anchored deterministic ODE sampler: together they constrain the prior to complete only the frequencies the optics annihilated, rather than synthesizing arbitrary detail.","feed_headline":"Five-step diffusion prior recovers high-NA 3D refractive index from low-NA holotomography","feed_subtitle":"Deterministic ODE sampler inverts Fourier band-limiting operator, completing missing lateral frequencies while preserving the axial missing—","key_machinery":"ResShift-ODE: a residual-shifting diffusion prior whose reverse process is reformulated as a probability-flow ODE, anchored to the low-NA measurement as its endpoint rather than to a Gaussian noise prior. The forward operator is an explicit NA-limited Fourier-support mask that models what a low-NA holotomography system can and cannot measure. Together, the anchor and the operator ensure the prior completes only the annihilated lateral frequency band while preserving the axial missing cone.","core_discovery":"The key result is that a residual-shifted diffusion prior, made deterministic via a probability-flow ODE, can invert a known Fourier-band-limiting operator to recover high-NA-equivalent 3D refractive-index structure from low-NA holotomography inputs, while preserving the axial missing cone that should remain empty. The method achieves this in five denoiser evaluations per volume, making it roughly 166 times faster than a standard 1000-step diffusion model while maintaining comparable reconstruction fidelity to a deterministic U-Net regressor. The measurement-anchored design ensures that recovered high-frequency content is constrained by the measured low-NA support, and the asymmetric Fourier","pith_inferences":[],"forward_implications":["If the method generalizes to physical low-NA acquisitions, existing multiwell-plate holotomography systems could gain high-NA-equivalent resolution through a software update, enabling label-free subcellular imaging in high-throughput drug screening without new optics.","The measurement-anchored, cone-preserving behavior offers a falsifiable criterion for generative priors in other coherent imaging inverse problems: the model should not fill spectral regions that the physics says are unmeasurable.","The deterministic five-step ODE sampler makes diffusion-prior inference practical for post-acquisition screening pipelines, where stochastic sampling chains taking hours per volume would be prohibitive.","The same construction—explicit band-limiting operator plus measurement-anchored diffusion prior—could extend to other coherent modalities with defined information gaps, such as limited-angle tomography, synthetic-aperture extension, or sub-pupil imaging."],"fun_headline_variants":["Deterministic diffusion prior recovers high-NA refractive index from low-NA holotomography","Five-step ODE diffusion prior inverts Fourier band-limiting operator in holotomography","Deterministic diffusion prior recovers high-NA volumes in five denoiser evaluations","Low-NA to high-NA refractive index transfer via deterministic residual-shifted diffusion","Deterministic ODE diffusion prior recovers lateral high frequencies in holotomography"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that a simple analytic Fourier-support mask faithfully captures the dominant information loss of a real physical low-NA microscope. The mask omits pupil apodization, partial coherence, and optical aberration. If this idealized operator is not a faithful surrogate for actual low-NA optics, the quantitative results measure inversion of a mathematical operator rather than real-world NA transfer performance.","fun_headline_variants_meta":{"raw":{"variants":["Deterministic diffusion prior recovers high-NA refractive index from low-NA holotomography","Five-step ODE diffusion prior inverts Fourier band-limiting operator in holotomography","Deterministic diffusion prior recovers high-NA volumes in five denoiser evaluations","Low-NA to high-NA refractive index transfer via deterministic residual-shifted diffusion","Deterministic ODE diffusion prior recovers lateral high frequencies in holotomography"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":716,"prompt_tokens":604,"completion_tokens":112,"prompt_tokens_details":null},"tokens_in":604,"tokens_out":112,"duration_ms":19372,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T23:08:24.064110+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Apply the trained model to a physically acquired low-NA holotomography volume of a specimen for which a registered high-NA acquisition of the same specimen exists. If the recovered volume fails to match the high-NA reference—particularly in the lateral high-frequency annulus outside the low-NA support—then the method inverts the analytic operator but does not achieve physical NA transfer.","supporting_citations":[],"review_version":1}