{"id":"0d2390ff-cc7e-400d-a690-b7d322e20780","arxiv_id":"2509.04848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OmniFHT reconstructs 3D refractive index maps of flowing cells from sparse, unknown-angle measurements by alternating pose search and implicit neural representation-based volume reconstruction.","lead":"A new algorithm builds 3D refractive index images of cells flowing through a microfluidic channel without needing to know the cells' rotation angles. It jointly estimates each cell's tumbling motion and its 3D structure, which could let flow cytometers image every cell, including misshapen ones and clusters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on Rytov weak-scattering linearization (Eq. 3); the paper only validates at RI contrast 0.02 in BPM, so quantitative RI for dense aggregates and RBCs is unverified.","rationale":"The reader's weakest-assumption analysis identifies the Rytov weak-scattering approximation as the load-bearing physical premise. I agree: Eq. (3) is the forward model that connects measurements to the unknown scattering potential, and every subsequent pose-estimation and reconstruction step inherits its validity. The paper's simulation validation is performed at RI contrast 0.02 with BPM, which is a propagation model but still an approximation that does not fully exercise multiple-scattering rejection. Real samples contain RBCs and aggregates with higher contrast, so the quantitative accuracy of the reconstructed RI values and the assertion of unbiased population-scale imaging are not yet established. The experiments provide qualitative morphological plausibility, but not a quantitative check of RI fidelity. This is an addressable limitation rather than a fundamental flaw, so the conditional verdict is appropriate: the method is credible for low-contrast cells, and the central claim should be conditioned on validation at higher RI contrast and with multiple-scattering ground truth. The concrete test with a full-wave solver directly targets this gap. No other concern (e.g., pose-ambiguity or high-throughput overstatement) was judged more load-bearing, because those are either partially addressed by simulation or are engineering-scale issues that do not invalidate the core physics.","tokens_in":15602,"tokens_out":4502,"duration_ms":44511,"concrete_test":"Generate simulated holograms of a realistic biconcave RBC (RI 1.40, ~7 μm diameter) and a two-cell aggregate (each ~1.40) using a full-wave multiple-scattering solver (e.g., FDTD, rigorous coupled-wave, or Mie-series field propagation) at 532 nm. Run OmniFHT on the resulting complex fields. If the reconstructed RI values are biased by more than ~0.01 or the morphology shows artifacts (e.g., false vacuoles, shape distortion) compared with the low-contrast BPM case, the Rytov assumption in Eq. (3) is the limiting factor. Alternatively, compare the Rytov-based forward prediction against the full-wave field at each pose: a mean phase error above ~0.1 rad would quantify the model mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OmniFHT enables quantitative, in situ 3D RI reconstruction of entire flowing cell populations. This rests on Eq. (3), the Fourier diffraction theorem under the Rytov weak-scattering approximation, which linearly relates the 2D Rytov perturbation to the 3D scattering potential. The paper's only ground-truth validation of this model is a BPM simulation with a low-index vacuolated cell (RI 1.35, background 1.33, contrast 0.02) and two vacuoles. Real clinical samples include RBCs (typical RI ~1.40, contrast ~0.07) and multicellular aggregates, where phase retardation across the object can exceed several radians and multiple scattering among cells is significant. In that regime the Rytov approximation can introduce RI bias and artifacts not captured by the low-contrast BPM test. The experimental reconstructions in Figs. 3–5 are visually plausible but lack any independent RI ground truth or multiple-scattering simulation, so the quantitative accuracy of the reported RI distributions—and therefore the population-scale quantitative claim—is unverified precisely in the regime most relevant to the stated application.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OmniFHT, a self-supervised framework for three-dimensional refractive-index (RI) tomography of cells flowing in a microfluidic channel. The method jointly estimates each cell's unknown 3D rotation and in-plane translation and reconstructs the 3D scattering potential as an implicit neural representation in Fourier space, using the Rytov approximation and the Fourier diffraction theorem as the forward model. The authors validate the approach on beam-propagation-method simulations of a vacuolated cell, on experimental datasets including SW780 cells, RBCs, and multicellular aggregates, and on a clinical ascites specimen containing 21 cells. They report that OmniFHT outperforms a Rytov-based baseline, recovers useful structure from as few as 5–10 projections and from 120° angular coverage, and reconstructs cells with complex multi-axis rotations that violate the single-axis assumption of conventional flow holographic tomography.","tokens_in":15990,"tokens_out":6352,"duration_ms":60284,"significance":"If the central claims hold, the paper would make a useful contribution: it removes the known-pose and single-axis-rotation assumptions that currently restrict in-flow holographic tomography, and it demonstrates the possibility of population-scale, label-free 3D RI imaging in clinical biofluids. The forward model and loss are clearly stated, the BPM simulation provides a ground-truth check of both pose recovery and RI reconstruction, and the experimental demonstrations on heterogeneous clinical material are valuable. The main weaknesses are that quantitative accuracy is validated only at low RI contrast, the comparison baseline is weaker than existing angle-recovery methods, and the sparse-view evaluation uses the method's own full-view output as the reference. These issues affect the strength of the quantitative and comparative claims, but they are addressable with additional validation rather than being fundamental flaws.","major_comments":[{"comment":"The quantitative-accuracy claim rests on the Rytov weak-scattering approximation, but the only ground-truth simulation uses an RI contrast of 0.02 (cell 1.35 vs. medium 1.33) and a single vacuolated cell. The target samples in Figs. 3–5 include RBCs and multicellular aggregates with substantially higher RI contrast and stronger multiple-scattering effects; in this regime the linearization in Eq. (3) can introduce RI bias and artifacts that the current BPM test does not probe. The experimental reconstructions are visually plausible but are not validated against any independent RI ground truth, so the population-scale quantitative claim is not yet supported.","section":"§3.1, Eq. (3)"},{"comment":"The comparison baseline is described as a 'standard Rytov-based method' that assumes uniformly interpolated poses. Existing angle-recovery FHT methods (refs. 16–18) estimate the rotation from the projections rather than imposing uniform increments; because the baseline is handicapped by construction, the reported 1.75× FSC improvement and the comparisons in Figs. 2d–g and 3b–d do not establish superiority over the current state of the art. The authors should include a baseline that recovers angles from the data, such as phase-similarity matching or periodic-cycle detection, and report results for both single-axis and multi-axis rotation.","section":"§3.1, Fig. 2"},{"comment":"The sparse-view and limited-angle FSC curves are computed against the full-view OmniFHT reconstruction of the same dataset. Because the sparse and reference reconstructions share the same INR prior, pose-initialization strategy, and optimization procedure, these FSC numbers measure consistency with the method's own full-view output rather than absolute fidelity to the true structure. The claim that reconstructions from 10 views or 120° of coverage are 'high-fidelity' should be supported by ground-truth simulations at those sampling densities, or by a reference that does not come from the same algorithm.","section":"§3.2.4, Fig. 4b"},{"comment":"The description of the neural representation is incomplete: Eq. (6) requires the network to output a complex-valued scattering potential in Fourier space, but §2.4 states that the MLP outputs a single scalar, without explaining how the real and imaginary parts are represented or how the complex loss in Eq. (7) is computed. This is central to reproducibility and should be specified (e.g., two output heads, complex-valued output layer, or separate real/imaginary channels).","section":"§2.4, Eq. (6)"}],"minor_comments":[{"comment":"The caption contains a typo: 'Eular Angles' should be 'Euler Angles'; the text in §3.1 also uses 'slide' where 'slice' is intended.","section":"§3.1, Fig. 2 caption"},{"comment":"The sentence introducing Eq. (1) says 'The progress is', which appears to be a typo for 'The process is'.","section":"§2.2, Eq. (1)"},{"comment":"The abstract claims reconstruction of 'entire flowing cell populations', but the demonstration reconstructs 21 cells after explicitly excluding out-of-focus cells and cells near the downstream end; the abstract should be qualified to match the reported scope.","section":"§3.3"},{"comment":"The pose-search procedure is described only qualitatively (30° initial grid, five refinement stages, top-8 hypotheses); for reproducibility, the authors should give the exact grid sizes, translation bounds, and stopping criteria.","section":"§2.3"},{"comment":"The FSC threshold differs between the simulated (1/2, reference-based) and experimental (1/7, half-set) evaluations; the choice is plausible, but the text should state explicitly that the resolution values obtained under the two criteria are not directly comparable.","section":"§3.2.4"},{"comment":"The data availability statement is only 'Data will be made available on request'; providing code and trained models, or at least a detailed pseudocode protocol, would substantially strengthen the reproducibility of the claimed results.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the paper has the potential to be a strong contribution to quantitative phase imaging and flow cytometry. My main concern is the gap between the broad quantitative claims (RBCs, aggregates, entire populations) and the validation evidence, which is limited to low-contrast simulation and visually assessed clinical data. I would encourage the editor to ask for the additional validation and comparison experiments described in the major comments; I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a genuinely useful paper, and the central machinery — joint INR-based pose and volume recovery under the Fourier diffraction theorem — is a sensible extension of cryo-EM/NeRF ideas to flow holographic tomography. The forward model and loss are clearly stated, the simulation validation against BPM ground truth is real evidence, and the qualitative experimental results on RBCs, aggregates, and ascites cells are visually convincing. The paper deserves a serious referee.\n\nWhat is actually new: prior flow holographic tomography either assumes single-axis rotation and known or recovered angles, or filters cells to near-spherical ones. OmniFHT drops that assumption and jointly estimates SO(3) poses and the 3D RI volume from the Rytov linearized data. The coarse-to-fine pose search and Fourier-domain INR are not themselves novel, but the combination for this imaging modality is, and the paper makes the adaptation explicit.\n\nWhere I’d push back, in proportion to severity:\n\nFirst, the load-bearing physical premise, Rytov weak scattering, is validated only in a BPM simulation at Δn=0.02. The experimental samples include RBCs (Δn≈0.07) and aggregates, where phase retardation can be several radians. There is no experimental ground-truth RI, no multiple-scattering simulation, and no analysis of how RI bias grows with contrast. So the claim of quantitative RI accuracy at population scale is unverified in exactly the regime the paper cares about. That is a real gap, not a fatal one.\n\nSecond, the baseline is weak. Comparing against \"Rytov with uniformly interpolated poses\" bypasses the angle-recovery methods that practitioners actually use. Given the simulation shows pose error and RI error are coupled, a stronger baseline (e.g., periodic similarity-based angle recovery used in prior FHT work) would make the gain claim more credible.\n\nThird, the population-scale demonstration is 21 cells after exclusions. The qualitative diversity is nice, but \"entire flowing cell populations\" oversells n=21. Also, no code or data are released, which matters for a method whose value is in adoption.\n\nMinor: the sparse-view evaluation uses the method's own full-view reconstruction as reference; acceptable for relative comparisons, but it is not ground truth. The half-set FSC 1/7 criterion is standard but lenient.\n\nBottom line: the paper is a legitimate contribution with a clear formulation and credible simulation support. The quantitative and population-scale claims need tempering and stronger validation. Send it to review — it will improve with revision, and the approach is worth refereeing.","headline":"Solid pose-free INR framework for flow holographic tomography; quantitative claims outrun the validation, especially at RBC-level RI contrast.","tokens_in":16377,"tokens_out":1760,"would_cite":true,"duration_ms":16039,"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":"OmniFHT reconstructs 3D refractive-index maps of flowing cells without knowing their pose, using as few as ten views or 120 degrees of angular coverage.","keywords":["quantitative phase imaging","flow cytometry","holographic tomography","implicit neural representation","refractive index reconstruction","pose estimation","Fourier diffraction theorem","Rytov approximation"],"falsifier":"Flow a rigid phantom with known refractive-index distribution and a known rotation trajectory, such as a lithographically patterned microbead assembly on a piezo-driven rotator, through the same microfluidic channel; reconstruct with OmniFHT and compare the recovered poses and RI slices against the known ground truth. A systematic drift of RI values as scattering strength or rotation speed increases would refute the paper's implied claim that the Rytov linearization suffices for arbitrary flowing cells.","tokens_in":15406,"feed_emoji":"🔬","tokens_out":7736,"duration_ms":68523,"temperature":0.7,"pith_summary":"OmniFHT claims to remove the two assumptions that currently limit in-flow holographic tomography: that each cell rotates about a single fixed axis and that the rotation angle of every projection is known in advance. Instead of detecting rotational cycles and interpolating angles, the method jointly optimizes each cell's unknown three-dimensional rotation, in-plane translation, and volumetric refractive-index distribution through a maximum-likelihood formulation in Fourier space. The volumetric map is represented by an implicit neural network evaluated on Fourier coordinates, and its built-in smoothness acts as a regularizer that fills in missing spectral information. On simulated and experimental data the paper reports high-fidelity reconstruction from as few as 10 projections or 120 degrees of angular coverage, and it reconstructs every cell in a clinical ascites specimen without pre-selection. If the claims hold, label-free flow cytometry can become unbiased with respect to cell shape and rotational dynamics.","feed_headline":"Pose-free 3D cell imaging from 10 views","feed_subtitle":"Joint pose-structure optimization removes single-axis rotation limits in flow cytometry tomography.","key_machinery":"The load-bearing object is the implicit neural representation of the 3D scattering potential in Fourier space: a three-hidden-layer MLP with sinusoidal positional embedding that maps frequency coordinates $(k_x, k_y, k_z)$ to the complex scattering potential. Paired with the Fourier diffraction theorem, which linearly links a 2D Rytov-phase projection to a slice of this potential on an Ewald sphere, the network can synthesize predictions under any hypothesized pose; those predictions drive both the data-consistency loss and the pose search. The pose search is coarse-to-fine: it starts from a $30^\\circ$ rotation grid and a translation grid with spacing 0.1, keeps the top eight hypotheses by complex cross-correlation, and refines by bisection over five iterations, with pose updates interleaved every five training epochs.","core_discovery":"Under the Rytov weak-scattering approximation, each measured two-dimensional complex field is a slice of the cell's three-dimensional scattering potential on an Ewald sphere, so the pose-free problem becomes a joint maximum-likelihood estimation over the unknown poses $\\omega_i = (R_i, t_i)$ and the scattering potential (Eq. 5). OmniFHT solves it by alternating a coarse-to-fine search over $\\mathrm{SO}(3)$ and translation space for each projection with self-supervised training of a compact multilayer perceptron that maps Fourier coordinates to the complex scattering potential; the predicted projections are generated by the Fourier diffraction theorem (Eqs. 6-7). The neural representation's implicit regularity restores missing frequencies, allowing high-fidelity reconstructions from sparse views and restricted angular ranges. The central claim is that this combination removes the single-axis, known-pose restriction of flow-based holographic tomography and enables unbiased 3D refractive-index reconstruction of entire flowing cell populations.","pith_inferences":["Because the framework only requires a differentiable forward model linking structure to projections, the same joint pose-structure loop could be ported to other label-free tomographies with unknown orientation, such as X-ray or electron microscopy of isolated particles.","The reported resolution gains are measured against a Rytov baseline that assumes uniformly interpolated poses; comparing against a pose-free baseline given ground-truth poses would separate the benefit of pose refinement from the benefit of neural implicit regularization.","Swapping the Rytov model for a multiple-scattering forward model (for example a Born-series or beam-propagation operator) inside the same alternating loop is a direct testable extension that would show whether pose-free reconstruction survives outside the weak-scattering regime.","Since the neural representation acts as an implicit prior, it may smooth or plausibly hallucinate fine structure inside missing spectral regions; a phantom with known sub-resolution features would quantify that bias."],"forward_implications":["Cells with arbitrary geometry and multi-axis rotation, which current flow-cytometry tomography discards, can be reconstructed, removing a selection bias from population statistics.","Because reconstruction succeeds with as few as 10 projections or 120 degrees of angular coverage, faster flow rates and shorter in-channel dwell times become usable in high-throughput assays.","In simulations the method reaches 1.59 µm resolution versus 2.78 µm for the standard Rytov baseline, a 1.75-fold improvement by Fourier shell correlation.","On real red blood cells, multicellular aggregates, and collision-perturbed white blood cells, only OmniFHT resolves the expected morphologies (biconcave discs, distinct cell boundaries, vacuoles), where baseline reconstructions show severe artifacts.","In a clinical ascites specimen the method reconstructs all 21 retained cells in situ without trajectory pre-filtering, establishing a workflow for population-scale label-free RI cytometry."],"supporting_citations":[{"why":"Supplies the Rytov weak-scattering model and the Fourier diffraction theorem that linearly relate 2D measured fields to the 3D scattering potential.","marker":"[38]"},{"why":"Establishes the in-flow holographic tomography geometry that OmniFHT extends from known single-axis rotation to arbitrary pose.","marker":"[11]"},{"why":"Represents the prior single-axis rolling-angle recovery approach that OmniFHT removes the need for.","marker":"[17]"},{"why":"Introduces coordinate-based implicit neural representations with positional encoding that OmniFHT adapts to represent the 3D scattering potential.","marker":"[20]"},{"why":"Demonstrates joint pose-and-structure reconstruction with neural networks from unknown-orientation data, the strategy OmniFHT transfers to optical tomography.","marker":"[24]"},{"why":"Provides the beam-propagation method used to generate simulated diffraction datasets for numerical validation.","marker":"[42]"},{"why":"Supplies Fourier shell correlation as the resolution metric used to quantify reconstruction fidelity.","marker":"[31]"},{"why":"Supplies the FSC threshold criteria (1/2 vs 1/7) used in the resolution analysis.","marker":"[43]"}],"fun_headline_variants":["Pose-free 3D imaging of flowing cells","3D cell tomography without pose tracking","Implicit neural network unlocks 3D flow cytometry","From 10 views to full 3D: pose-free QPI","Arbitrary cell shapes imaged in 3D during flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing physical premise is that each cell is a weak scatterer, so the measured complex field depends linearly on the 3D scattering potential through the Rytov approximation; if real cells or clusters scatter strongly or exhibit significant multiple scattering, the recovered refractive indices and morphologies will be biased.","fun_headline_variants_meta":{"raw":{"variants":["Pose-free 3D imaging of flowing cells","3D cell tomography without pose tracking","Implicit neural network unlocks 3D flow cytometry","From 10 views to full 3D: pose-free QPI","Arbitrary cell shapes imaged in 3D during flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00036,"raw_usage":{"total_tokens":1961,"prompt_tokens":976,"completion_tokens":985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":905}},"tokens_in":592,"tokens_out":985,"duration_ms":8257,"temperature":1.0,"reasoning_tokens":905,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:25:07.580948+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Flow a rigid phantom with known refractive-index distribution and a known rotation trajectory, such as a lithographically patterned microbead assembly on a piezo-driven rotator, through the same microfluidic channel; reconstruct with OmniFHT and compare the recovered poses and RI slices against the known ground truth. A systematic drift of RI values as scattering strength or rotation speed increases would refute the paper's implied claim that the Rytov linearization suffices for arbitrary flowing cells.","supporting_citations":[{"cited_title":"Inverse-scattering theory within the rytov approximation,","cited_arxiv_id":null,"evidence_quote":"Supplies the Rytov weak-scattering model and the Fourier diffraction theorem that linearly relate 2D measured fields to the 3D scattering potential."},{"cited_title":"Three-dimensional holographic refractive-index measurement of continuously flowing cells in a microfluidic channel,","cited_arxiv_id":null,"evidence_quote":"Establishes the in-flow holographic tomography geometry that OmniFHT extends from known single-axis rotation to arbitrary pose."},{"cited_title":"Rolling angle recovery of flowing cells in holographic tomography exploiting the phase similarity,","cited_arxiv_id":null,"evidence_quote":"Represents the prior single-axis rolling-angle recovery approach that OmniFHT removes the need for."},{"cited_title":"Nerf: Representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Introduces coordinate-based implicit neural representations with positional encoding that OmniFHT adapts to represent the 3D scattering potential."},{"cited_title":"Cryodrgn: reconstruction of heterogeneous cryo-em structures using neural networks,","cited_arxiv_id":null,"evidence_quote":"Demonstrates joint pose-and-structure reconstruction with neural networks from unknown-orientation data, the strategy OmniFHT transfers to optical tomography."},{"cited_title":"Optical tomographic image reconstruction based on beam propagation and sparse regularization,","cited_arxiv_id":null,"evidence_quote":"Provides the beam-propagation method used to generate simulated diffraction datasets for numerical validation."},{"cited_title":"Eman: semiautomated software for high-resolution single-particle reconstructions,","cited_arxiv_id":null,"evidence_quote":"Supplies Fourier shell correlation as the resolution metric used to quantify reconstruction fidelity."},{"cited_title":"Fourier shell correlation threshold criteria,","cited_arxiv_id":null,"evidence_quote":"Supplies the FSC threshold criteria (1/2 vs 1/7) used in the resolution analysis."}],"review_version":2}