acceptodds
Under review as a conference paper at ICLR 2027

CryoWaveGS: Wave-Optical Gaussian Splatting for Cryo-Electron Tomography

Abstract

Gaussian Splatting has recently been adapted to tomographic reconstruction, but existing formulations assume a ray model that discards the depth at which scattering occurs. This approximation fails past a computable switch point that the resolution of current cryo-ET reconstructions does not reach, and that atomic-resolution series do. We introduce CryoWaveGS, a Gaussian Splatting reconstruction with a wave-optical forward model. Because Fresnel propagation maps a Gaussian to a Gaussian with complex covariance, depth-resolved diffraction can be rendered in a single rasterization pass at a cost similar to that of the ray-based renderer, and within end to end on our densest run. We validate the renderer against independent multislice simulation and map where the ray approximation is adequate and where it is not. We evaluate on measured cryo-ET and atomic-resolution tilt series against conventional, implicit-field and matched ray-based baselines under a single scoring pipeline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gaussian splatting, Cryo ET, Tomography

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ByjDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposes a Gaussian-based reconstruction method with a wave-optical forward model for electron tomography. Starting from the thick-phase-grating approximation (TPGA), the authors approximate each Gaussian primitive as sufficiently thin and incorporate its depth-dependent Fresnel propagation into a complex projected covariance. The resulting renderer is integrated with microscope image formation and joint refinement of the reconstruction and nuisance parameters. Evaluation includes forward-model comparisons against independent multislice simulations, experimental cryo-ET reconstructions, and an atomic-resolution Co3O4 tilt series.

Critical strengths
  1. Clear exposition and physically motivated method design. The paper explains the scientific background accessibly and connects the limitations of the projection model to the proposed operator. The main derivation is coherent, and the resulting depth-dependent complex covariance provides a compact formulation that fits naturally into Gaussian rasterization.

  2. A promising operator supported by controlled forward-model validation. Figure 3 compares different renderers on the same fixed potential against an independent multislice implementation, isolating forward-model accuracy from reconstruction and optimization. This supports the accuracy of the proposed approximation under the tested high-frequency conditions. The matched wave/ray comparisons are also useful for distinguishing benefits from the reconstruction framework and those from the physical correction.

Critical weaknesses
  1. The experimental settings provide limited evidence for the practical reconstruction benefit of the central physical contribution. The motivation is to improve image formation where specimen depth makes the projection approximation inaccurate. However, the reconstructions of EMPIAR-10164 and EMPIAR-10643 mostly remain at resolutions of tens to over one hundred angstroms, where the relevant propagation effect is small under the reported conditions.

    This mismatch can be quantified using the Fresnel phase expression and parameters in §5.2. At 300 kV, with an electron wavelength of approximately 0.0197 Å and a propagation distance of 81 nm, the depth-dependent Fresnel phase is πλzq2\pi\lambda zq^2. At spatial periods of 50 Å and 100 Å, this gives approximately 0.020 rad (1.15°) and 0.005 rad (0.29°), respectively. The corresponding complex propagation factors differ from unity by only about 2% and 0.5%. For comparison, the phase reaches approximately 2 rad at a spatial period of 5 Å under the same conditions. These are frequency-specific propagation-factor differences, rather than relative image errors, but they illustrate the mismatch between the physical motivation and the tested reconstruction regime.

    Consistently, the matched wave and ray reconstructions are essentially identical on these datasets. The authors explicitly state in §5.2 that the gains arise from the representation and joint refinement rather than wave optics. These experiments therefore do not establish an additional reconstruction benefit from the proposed physical correction.

    Figure 3 provides useful forward-model validation on a fixed simulated potential, and the Co3O4 experiment offers additional evidence at high spatial frequencies. However, the practical benefit for recovering macromolecular structures from experimental cryo-EM images remains insufficiently demonstrated. A high-resolution single particle cryo-EM experiment would provide a more direct test of whether the improved forward model enables recovery of additional reliable structure.

  2. The structural claims should be narrowed to match the available evidence. The higher half-map FSC demonstrates stronger agreement between the reconstructed halves under the reported evaluation protocol. However, this does not by itself establish a corresponding improvement in structural resolution or the recovery of additional fine detail. The cryo-ET visualizations primarily show reduced background noise and improved contrast, without clearly identifiable additional structural features commensurate with the reported FSC gains. In Figure 13, some shell regions appear broader or smoother than in WBP, rather than more distinctly resolved.

    My concern is therefore with the interpretation of the results, rather than the correctness of the reported FSC measurements. The evidence supports improved half-map consistency and visual contrast, but claims of “finer detail” or higher structural resolution require additional corroboration. Unless such evidence is provided, these claims should be narrowed accordingly.

Questions for authors
  1. Evaluation in a more suitable resolution regime. Could you evaluate your operator either on a high-resolution single particle cryo-EM dataset, such as EMPIAR-10202 in the CryoSPARC tutorial, or through high-resolution subtomogram averaging within cryo-ET? Since you already use EMPIAR-10164, the NovaCTF study provides a relevant precedent for demonstrating the benefit of depth-dependent correction using this data source.

  2. Structural evidence and claim scope. Can the authors identify specific structural features that become newly distinguishable in their cryo-ET reconstructions, supported by matched local views or density profiles, particularly in the shell regions of Figure 13? Comparison with appropriately filtered or denoised WBP would also help distinguish improved visibility from improved structural resolvability. If the principal observable improvements are background suppression and contrast enhancement, would the authors revise the claims of finer structural detail and higher resolution to reflect the evidence more precisely?

  3. Initialization and computational implementation. Please specify how the Gaussians are initialized for the cryo-ET experiments, including positions, masses, initial widths, and width limits at each binning stage. How are primitives smaller than the coarse sampling interval handled? Does every projection process all approximately 12 million Gaussians, and what fraction survives any culling? Are the multiple views per update rendered concurrently or processed sequentially with gradient accumulation? Please also clarify the modifications made to the original 3DGS rasterizer.

  4. Minor presentation issue. Could you adjust the layout of the pipeline diagram? Several text labels overlap with graphical elements.

The proposed operator is promising, but the current experimental settings do not sufficiently establish its practical benefit for macromolecular reconstruction, and the claims of finer structural detail exceed the evidence presented. A convincing high-resolution evaluation through either single particle cryo-EM or subtomogram averaging, demonstrating an advantage over the matched ray operator with corresponding structural evidence, would address my main concern and lead me to raise my score.

3 comments
AuthorsOpenReview ↗

Common Response

1. Testing the relevant regime [CW1, CW2, Q1, Q2]

Following the reviewer’s suggestion, we tested EMPIAR-10202, the AAV2 capsid dataset, at 0.788 Å/pixel and 300 kV (Section 5.3, Appendix I). This sampling allows testing at 2.1 Å, where the 250 Å capsid gives a half-depth Fresnel phase of about 1.75 rad, unlike the small phases in the cryo-ET tests. Wave/ray pairs share initialization and fitting settings with fixed CryoSPARC J2 poses and halves, giving four paired tests from two starts within one acquisition.

ReadoutWaveRay
Model FSC, 2.0 to 2.2 Å.7691.7361
Model FSC, 6 to 10 Å.9393.9394
Opposite-half image correlation, 2.0 to 2.2 Å.005535.003908
Mean residue Q-score, start 42.8704.8619
Mean residue Q-score, start 43.8698.8615

Model FSC improves by 0.0330 (SD 0.0020), but 6E9D is a same-data reference, not independent truth. All four high-frequency prediction intervals exclude zero, conditional on calibration using 48 test particles, and the 42% gain is relative correlation, not error reduction. Rendering as ray without refitting removes 90.6% of that gain (Table 25), while Q-scores use fixed coordinates without per-map fitting or sharpening, with residues not independent replicates (Figure 23). These results support a matched wave/ray benefit.


2. Cryo-ET structural claims [CW2, Q2]

Following the reviewer’s request, we compared matched regions with filtered/Topaz-denoised WBP and SIRT/SART, finding the same shells and broader submitted profiles, so we now claim improved contrast and background suppression rather than new structure (Figure 2). Views share a 1300 Å window, 80 Å slab and common 40 Å low-pass, with contrast scaled per panel.

Table 1 shows agreement gains on 10643 but not 10164, not structural resolution or missing-wedge recovery. The datasets use alternating-tilt and frame halves, respectively, and the 10164 visualization uses a separate 12M zero-mass fit, not the submitted WBP-initialized fit or the 3M baseline.


3. Initialization and subpixel Gaussians [Q3]

The submitted fits started from 12M WBP-based centers, sampled 90% by WBP weights and 10% uniformly, with equal masses totaling the WBP integral and σ=12 Å, then inherited bin 4, split to 10M at bin 2 and depth-cropped at bin 1.

The new baseline instead starts from 2.4M zero-mass lattice sites with small random displacements, aligned axes and σ=0.8 spacing, growing mass-conservingly to 3M at passes 20/40/60 with alignment/defocus fixed (Table 4 gives stage counts and bounds).

Gaussians remain continuous and are sampled at pixel centers, with physical width floors of 5.4/13.5 Å per axis/geometric mean, whereas the submitted p²/12 covariance addition approximated pixel-area integration.


4. Projection and rasterizer [Q3]

As requested, we checked culling: every projection processes all Gaussians, with none culled at bins 8/4/2 and at most 0.39% at bin 1 across five sampled views, versus 0 to 0.52% in submitted fits (Appendix A). Pixels evaluate only their 16 × 16 tile's list, with footprints extending about 5.7σ.

To clarify the view batching, each update processes eight views sequentially, averages their gradients and takes one optimizer step. Our renderer keeps the tiled design of 3DGS but replaces RGB compositing with additive complex splatting before transmission and microscope modeling, with batching and implementation details given in Appendix A.


5. Diagram [Q4]

Following the reviewer's suggestion, we corrected Figure 1 to separate labels from graphical elements and arrows.

Thank you for pointing us to EMPIAR-10202 (AAV2), which gave us the experimental macromolecular test missing from the submission and let us assess reconstruction benefits beyond forward-model accuracy. References are to the revised manuscript unless mentioned as submitted.

I appreciate the authors' clarifications and additional experiments. In particular, the experiments on AAV2 largely address my main concerns. The visual results support the improvements achieved by TPGA, which are also consistent with the results from cryoSPARC. I have therefore raised my rating.

Additionally, as SPA cryo-EM experiments have now been included in the paper and are crucial to supporting its central claim, I believe it is necessary to include relevant cryo-EM literature in the related work, such as conventional reconstruction methods like cryoSPARC[1] and RELION[2], as well as Gaussian-based methods like cryoSplat[3].

The following are merely recommendations, not critical weaknesses:

  1. As I understand it, the central claim of this paper is to demonstrate the necessity and practical benefits of TPGA for Gaussian-based high-resolution reconstruction. To support this claim, I believe the AAV2 experiment should serve as the primary evidence and be given greater prominence, since it directly compares TPGA with WPOA and validates the proposed formulation.

  2. Figure 1 could be further improved. I recommend rendering the equations using LaTeX for better visual quality.

References: [1] Punjani, A., Rubinstein, J. L., Fleet, D. J., & Brubaker, M. A. (2017). cryoSPARC: Algorithms for rapid unsupervised cryo-EM structure determination. Nature Methods, 14, 290–296. https://doi.org/10.1038/nmeth.4169

[2] Scheres, S. H. W. (2012). RELION: Implementation of a Bayesian approach to cryo-EM structure determination. Journal of Structural Biology, 180(3), 519–530. https://doi.org/10.1016/j.jsb.2012.09.006

[3] Chen, S., & Ling, H. (2026). CryoSplat: Gaussian Splatting for Cryo-EM Homogeneous Reconstruction. International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=dLaUZKBzta

We thank the reviewer for raising the score and for the helpful suggestions.

Cryo-EM related work. Related Work (Section 3) now includes a paragraph on single-particle reconstruction: cryoSPARC [1] and RELION [2] as conventional voxel-based methods, and CryoSplat [3] as a Gaussian-based method. It also states that our AAV2 experiment uses cryoSPARC poses and CTFs, and that the TPGA and WPOA fits differ only in the projection operator.

Figure 1. All labels and equations in Figure 1 are now typeset in LaTeX.

Once again, thanks for the follow up and great insights!

Reviewer_83WROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed CryoWaveGS for 3D reconstruction of cryo-electron tomography (cryo-ET). Their Gaussian representation is built on previous works such as R^2-Gaussian and CryoETGS, their main contribution here is incorporating depth-dependent wave propagation into Gaussian splatting. They started with the thick-phase-grating approximation (TPGA), proposed by Cowley & Moodie back in 1962 for high-resolution electron microscopy (HREM) of perfect crystals to account for multiple scattering events and the strong phase shifts that electrons experience when passing through a specimen (applies to both thick, ordered crystal matrix and thick cryo-ET samples of disordered and weakly-scattering macromolecules), then made a further approximation so each Gaussian's complex covariance describes its projected shape and propagation from its depth to a plane through the specimen's rotation centre. This enabled them to account for depth-dependent defocus in one rendering pass, rather than propagating through successive slices as in the case of the multislice approach. They also proposed a ray version of the same method, when the internal propagation term is removed.

They first compared their renderer with independent multislice simulations using the multislice software abTEM, then tested both ray and wave versions on two well-known biological cryo-ET datasets and a cobalt oxide nanoparticle dataset also acquired as a cryo-ET dataset (by authors of PhaseT3M). On the biological datasets, the wave and ray results are practically the same. So the gains over weighted back-projection (WBP) cannot be attributed to the wave term there. On the nanoparticle dataset, the wave version gave results closer to reconstruction made with PhaseT3M than the ray version did. After 100 passes through the data, the Fourier correlation at 2 A is 0.85 for the wave version and 0.42 for the ray version, over the particle and surrounding carbon together. The wave version takes about 1.3 times longer to run. But PhaseT3M is itself a reconstruction, not the ground truth as acknowledged by the authors. Agreement with PhaseT3M at 2 A does not by itself show that the method has recovered the true structure at 2 A resolution. The authors also mentioned that neither version could reproduce details finer than ~7 A in a tilt image left out during fitting. These comparisons demonstrate different conclusions.

Critical strengths

The authors’ main idea is interesting. But their Gaussian representation itself follows previously published work, R^2-Gaussian and CryoETGS, as they mentioned in Sections 3 and 4. Accounting for specimen thickness and depth-dependent defocus are not new either. They also mentioned NovaCTF for considering the variation of defocus through the tomogram thickness, TPGA and multislice for accounting for electron propagation within the specimen. So here their main contribution lies in how they put depth-dependent wave propagation into Gaussian splatting.

In Equation 7, each Gaussian's complex covariance encodes its projected shape and wave propagation from its depth to a plane perpendicular to the beam through the specimen's rotation centre. This accounts for depth-dependent defocus in one rendering pass. Multislice instead divides the specimen into thin slices, with the wave being transmitted through a slice and then propagated to the subsequent one. Avoiding computationally-heavy successive propagation is a merit, especially if it can be done at a cost similar enough to their ray renderer as they reported. This is the main technical strength of the paper. It should still be clear that this remains an approximation of the wave model, because they started from TPGA and dropped the cross and axial terms in Equation 6 to get the Gaussian form. As the authors acknowledged, their renderer can break down under strong multiple scattering, its agreement with multislice is demonstrated for the tested conditions.

The comparison between their wave and ray versions is useful, there they kept the representation and fitting settings the same. They first checked the renderers using the same fixed simulated specimen and compared the predicted images with images calculated using the multislice software abTEM. So this checks the rendering itself, where they reported that the wave renderer kept the image correlation of about 0.9 down to 1.6 A, while the ray renderer lost agreement below ~5 A in this test.

On the biological EMPIAR datasets, the two versions gave pretty much the same results. On the nanoparticle dataset, their performances separated at higher spatial frequencies: at pass 100, the wave version had a Fourier correlation of 0.85 with PhaseT3M at 2 A over the particle and surrounding carbon, compared with 0.42 for their ray version, with about just 1.3 times longer running time. Their reporting of the simulated renderer test and this matched reconstruction comparison was clear in showcasing the difference made by including wave propagation against their own ray version. They also clearly stated PhaseT3M is not ground truth and that their biological gains are not due to wave optics.

Critical weaknesses
  1. A main concern lies in the authors’ treatment of tested datasets which was not clearly reported in the manuscript if any additional procedure was performed after data retrieval. Specifically, whether the contrast transfer function (CTF) has been estimated and corrected to account for defocus variation across the tilt series if not also corrected to adapt to sample thickness in 3D as could be done with NovaCTF (the only reference the authors have included regarding CTF). Since CryoWaveGS included CTF in Equation 2, it accounts for how defocus can affect image contrast. If other methods that were compared against CryoWaveGS were applied to uncorrected images, then some of the improvement could come from this difference. Particularly because CTF-correction is routinely applied to tilt images before 3D reconstruction for all mentioned existing methods except PhaseT3M, as these methods do not integrate contrast transfer considerations on their own and assume to have corrected tilt series as their input.

Metal oxides are routinely imaged with room-temperature electron microscopy rather than cryo-ET as they are robust against irradiation damage unlike macromolecules or biological cellular matter that require cryogenic imaging conditions. Here the authors chose to use the benchmarking dataset generated by authors of PhaseT3M to demonstrate benefit of including wave optics in Gaussian splatting with their method. And as they have stated in their abstract, such a dataset is not representative of the mass majority of biological cryo-ET datasets commonly seen in the community in terms of the reachable resolution. While CTF estimation was performed for this cobalt oxide dataset where the initial defocus values were obtained with CTFFIND4, no CTF correction was applied as described in the PhaseT3M publication when they compared their method against existing methods which would require CTF-corrected inputs.

As the EMPIAR datasets have only been motion-corrected, it was not reported if additional procedures such as CTF fitting and correction were performed by authors, nor the specific method or associated settings used. Moreover, the ‘halos’ around gold fiducial beads shown in WBP-reconstructed volume in Figure 14 are typically seen in tomograms where CTF-correction was skipped. It is uncertain how much the FSC/FRC curves would change after CTF correction, but this is something the authors should examine.

CTF-correction also matters for comparison with CryoETGS, a method much closer to the spirit of CryoWaveGS than WBP and its derivatives, as its authors clearly stated in their publication their method was intended to work with CTF-corrected images as it does not account for CTF in its Gaussian renderer. They also stated ‘applying CTF pre-correction yields consistent improvements in both PSNR and SSIM of the reprojected tilt images’. So here it is important information whether the CryoETGS comparison was made with CTF-corrected input it needs. Though, the PSNR and SSIM result in that paper cannot suggest how much the FSC/FRC would change on these data. Again, this should be examined by the authors.

In Table 1b and the last paragraph of Section 5.2, the wave and ray versions gave practically identical results for the biological Cryo-ET datasets. Here the authors clearly explained these data do not resolve fine enough detail for propagation through the specimen to make a measurable difference. So their improvement over WBP here is not due to wave optics. For the biological datasets, the authors attributed the improvement to the Gaussian representation and refinement of alignment, defocus and ice model. But the authors did not report separate contributions of these measures nor adapting Gaussians. And the improvement did not span all binning as for the EMPIAR-10164 dataset at bin 1, Table 1a gave 31.7 A for CryoWaveGS (wave version) and 31.4 A for WBP, which are practically identical resolutions with the latter even a little bit better while the input tilt series was not CTF-corrected. Whereas, in the publication of PhaseT3M, defocus before and after refinement was clearly showcased for comparison. In addition, the authors did not report on where their initial defocus values to be refined came from, as in which software and fitting settings were used. They only briefly mentioned Thon rings fitting, but it is not clear what the actual procedure was.

  1. The authors used significantly different numbers of Gaussians for different datasets, but their criteria for these choices is not clearly explained. Despite generally describing splitting, cloning and pruning of the Gaussians, it is not reported how the authors would set the initial numbers of Gaussians or deciding when to stop adding them. Even though the authors reported Gaussian counts, run times, and memory in Appendix F, it is not shown whether and how sensitive varying numbers of Gaussians can result in quality changes of reconstruction.

In Equation 8, the authors used the square root to downweight errors in bright pixels and make the counting noise more similar across bright and dark pixels. But they did not give the exact intensity normalisation used before this step. They also used a squared L2 loss, but the noise assumptions behind it were not clearly explained for the errors after the square root and frequency weighting. It would have been useful to see whether an L1 or another robust loss gave a different result. For Adam, the authors did not show whether the fitting had converged or whether starting from different initial values gave similar reconstructions. So it was not clear how much the result depended on these choices.

For the cobalt oxide dataset, using PhaseT3M as a reference was reasonable given the lack of ground truth, and the authors acknowledged this limitation. But their main reported result came from fitting only the Gaussian masses for 100 passes with fixed centres and widths. They then pruned the Gaussians and allowed the centres to move for 50 passes. They reported a lower image R-factor but less agreement with PhaseT3M-generated reconstruction. However, as the comparison included pruning followed by more passes with movable centres, the effects of pruning, centre refinement and additional fitting cannot be separated.

Questions for authors
  1. Were the images used for comparison with existing methods CTF-corrected ? If not, does the improvement persist when appropriately corrected inputs are used as inputs ?

  2. For the test datasets, which software and settings were used for the initial defocus estimates, and how much did these change after refinement ? Were alignment refinement, defocus refinement and ice model refinement contributing differently ?

  3. How were the initial number of Gaussians and rules for cloning, splitting and pruning chosen ? Was the reconstruction quality or computational cost sensitive to changes in these choices ?

  4. For the test datasets, how did refining the different Gaussian parameters affect reconstruction quality ? Were the effects of changing their centres, widths and orientations tested separately from fitting their masses ?

  5. How were the image intensities normalised before calculating Equation 8, and what noise assumptions were made for the errors after the square root and frequency weighting ? How was convergence checked here, and did different initialisations give similar or different reconstruction qualities ?

Suggestions :

  • Directional FSC was mentioned by authors in the introduction as an informative way of exhibiting differences in directional resolution of reconstructions especially concerning the missing wedge effect in cryo-ET, but was not presented in results.
  • The visual comparisons would be easier to follow with the same region of reconstructed tomograms obtained with different methods shown next to each other. E.g., as shown in the PhaseT3M publication on the HIV-1 dataset (Figure S16).
  • The R-factor should be defined with its equation and normalisation, or a clear reference to the appendix where these were given as not all readers are familiar with such a concept from the field of crystallography. The authors should also make their introduction of TPGA and related concepts that are community-specific more accessible for non-experts.
  • The thin/thick description at the start of Section 5 needs correction. It compared t_crit with an unspecified threshold, but it should compare specimen thickness with t_crit.
  • The claim of improvement at every binning should be revised. For EMPIAR-10164 at bin 1, Table 1a gave 31.7 A for CryoWaveGS (wave version) and 31.4 A for WBP, these are identical results with WBP even a little bit better.
  • Section 5.2 mentioned an IsoNet2 improvement, but the results were not shown. The relevant results and settings should be stated or this statement should be removed.
  • A comparison on the same dataset using N/8, N/4, N/2, N, 2N, 4N and 8N Gaussians would help show how the number of Gaussians affected reconstruction quality, running time and memory.
  • Comparison with an L1 or another robust loss, or a different optimiser, could help show how these choices may affect the optimisation.
1 comment
AuthorsOpenReview ↗

Common Response

1. CTF fairness [CW1, CW3, CW4, CW5, Q1, S5]

After the reviewer’s comment, we confirmed that WBP and CryoETGS used uncorrected inputs while CryoWaveGS modeled CTF, so we repeated the comparisons with corrected inputs and fixed halves, alignment and masks for a fairer comparison. The 3M fit improves over WBP on 10643 but not 10164, so we withdrew the all-binning claim (Table 1).

We reproduced CryoETGS on phase-flipped inputs with the paper's 10643 settings: 10,000 iterations, at most 5M Gaussians. Our separate 12M fit is scored against opposite-half phase-flipped WBP using one-direction full-shell FSC, not ΔFSC (Appendix J). Triples are 100 to 60 / 60 to 40 / 40 to 30 Å.

Method1016410643
CryoWaveGS.380 / .378 / .181.263 / .147 / .089
CryoETGS.250 / .260 / .084.149 / .088 / .041
WBP.406 / .406 / .182.168 / .069 / .038

WBP correction changes ΔFSC by at most about 0.002 here. Gains over SIRT/SART on 10643 are smaller and mixed at bin 2. Gold halos on 10453 persist after phase flipping, so appearance cannot establish CTF treatment.


2. Co₃O₄ scope [CW2, Q1]

We agree Co₃O₄ is not routine biological cryo-ET: acquired cryogenically without ice encapsulation, we use it to test high-frequency propagation alongside AAV2 (Section 5.3). Our fits model CTF on raw images, WBP uses phase-flipped foci, and CryoETGS/IsoNet2 average three separately flipped foci with IsoNet2 correction disabled.


3. Defocus [CW3, CW5, Q2, Q4]

To clarify defocus, Appendix A gives CTFFIND 4.1.14 settings for 10164 and deposited IMOD ctfplotter estimates for 10643, whose settings are unreported. Mean shifts near +1900/+1830 Å did not consistently improve agreement, so we fix alignment and defocus, with larger effects from growth and widths than wave propagation (Table 5). Submitted Co₃O₄ optics stayed fixed, while new fits refine them (Appendix B.1).


4. Gaussian counts [CW6, Q3, S7]

As suggested, we tested seven budgets on 10164 (Table 16), finding that quality improves above our practical 3M budget, but doubling 12M to 24M adds at most 0.003 while time and memory roughly double. Both datasets follow the same growth rule, starting from 0.8N zero-mass, randomly displaced lattice sites and adding N/15 mass-conserving splits/clones at passes 20/40/60 until reaching N without pruning, with finer-stage continuation given in Table 4.


5. Loss [CW7, Q5, S8]

To clarify the normalization, we clip images at zero, divide by their valid-pixel mean and take the square root, which approximately stabilizes Poisson variance without removing processing correlations, with planar gain fitted to cryo-ET predictions. Co₃O₄ uses W=1, while current cryo-ET only excludes periods above 1000 Å perpendicular to the tilt axis, without SSNR weighting.

The submitted cryo-ET fits instead used fixed gains and Huber on whitened intensity residuals, with W estimated from residual power rather than frame SSNR (Appendix A), while the new controls favor L2 over 200 to 30 Å, with L1/Huber leading in outer bands and momentum SGD improving finer but worsening coarser bands (Table 6).


6. Convergence [CW7, Q5]

We agree that training loss alone is not enough, so we tested eight held-out 10164 tilts, where MSE ×10⁴ falls from 1.82 to 1.53 at pass 100 but rises to 1.59 at pass 300 even as training loss keeps falling (Table 7). Three starts agree within 0.001 in band scores, and unseen-tilt prediction improves over WBP, although fitted ice and absorption also contribute to that gain.


7. Pruning [CW8, Q4]

As requested, we rescored immediately after pruning and found little change, while later fitting lowers wave correlation near 2 Å from .931/.824 to .879/.629 (all/wedge). Without a matched frozen-center continuation, we cannot isolate center motion. New fits fix geometry without pruning (Appendices B/C).


8. Structural claims [CW5, S1, S2, S5, S6]

Following the reviewer's suggestion, we matched regions and filtering (Figures 2, 7 to 8), finding better contrast and less background, not newly resolved structure. 10164 visuals use 12M, not the 3M baseline. Directional scores and curves (Table 8, Figure 6) do not prove missing-wedge recovery. We removed the unmatched biological IsoNet2 claim.


9. Physics and R-factors [S3, S4]

As requested, Appendix F defines amplitude R as each image's summed absolute amplitude error / summed observed amplitude, averaged without affine fitting. Submitted .161/.168 are affine-calibrated intensity scores. Sections 2/4 and Appendix G explain TPGA's single phase step, depth dependence and dropped cross/axial terms, with multislice the reference for strong scattering. Section 5 compares thickness with d²/λ at the measured period.

We appreciate the reviewer's questions and feedback. References above are to the revised manuscript unless mentioned as submitted.

Reviewer_YEXrOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CryoWaveGS, a Gaussian representation for electron tomography with a differentiable wave-optical renderer based on the thick-phase-grating approximation (TPGA). Under an additional approximation that treats each primitive as axially thin, the depth-dependent Fresnel phase is absorbed into a 2D complex covariance. The propagated effective potential can then be accumulated in one rasterization pass, followed by exit-wave formation and microscope and detector modeling. Removing the internal depth-dependent phase recovers the matched ray renderer.

The experiments cover two biological cryo-ET datasets and an atomic-resolution Co3O4 series. At the resolutions supported by the biological data, wave and ray reconstructions are nearly identical, while the full Gaussian pipeline improves over WBP at several binnings. For Co3O4, TPGA-GS agrees better than Ray-GS with independent multislice forward simulations. In measured-data reconstruction, its pass-100 Fourier correlation with the deposited PhaseT3M volume at 2 Å is 0.85 versus 0.42 over particle plus carbon, and 0.93 versus 0.74 for the particle alone. The paper also reports a held-out image test and computational benchmarks.

Critical strengths

A useful adaptation of wave-optical Gaussian rendering to electron tomography. Encoding the depth-dependent Fresnel phase in a complex covariance combines an explicit Gaussian representation with established TPGA physics while avoiding sequential multislice propagation. The derivation explains the relation to ray projection and explicitly identifies the dropped terms. The comparison against independent abTEM multislice simulations provides convincing evidence that the renderer captures high-frequency propagation effects missed by the ray model in the tested conditions. This is the paper’s strongest technical contribution.

Experiments organized around a clear physical prediction. The depth-of-field argument motivates testing both a regime where wave and ray models should agree and one where they should differ. The near-identical biological cryo-ET results are a useful control, while the Co3O4 forward experiment demonstrates the expected separation at finer periods. The authors appropriately distinguish the biological pipeline gains from the contribution of wave propagation, disclose the unsuccessful compact representation on measured data, and provide detailed reproduction settings and compute measurements.

Critical weaknesses
  1. Independent evidence of improved high-resolution 3D reconstruction is limited. First, stronger agreement with the PhaseT3M reconstruction does not independently establish specimen accuracy (Section 5.3, Table 2). The held-out 0° defocus triplet provides no corroborating predictive advantage: Ray-GS has lower amplitude MSE (0.0195 versus 0.0218), and neither prediction correlates reliably with the measured images at spatial periods finer than about 7 Å (Figure 22). This noisy, single-angle test cannot determine which 3D volume is more accurate or establish its resolution. Second, each reported reconstruction in Table 2 is based on a single optimization run, so the stability of the differences across repeated optimizations is unquantified (Section 5.3). Third, the definition and phase-randomization normalization of the {400} score need clarification before it can be interpreted as structural evidence (Section 5.3, Appendix C). Phase-only randomization would preserve both described power-based statistics on a fixed Fourier grid and shell, with fixed lattice selection and no further processing. These issues limit the demonstrated inverse-reconstruction contribution beyond the validated forward operator. A targeted inverse experiment with known 3D ground truth, independently simulated noisy multislice data, and repeated reconstructions would strengthen the accuracy and robustness evidence.

  2. The incremental benefit of the Gaussian renderer needs clearer attribution and an accuracy-cost comparison. First, TPGA-GS and Ray-GS change propagation for both the Gaussian particle and the voxel carbon (Appendix C, Table 3). Their full-field difference therefore does not isolate the proposed Gaussian renderer’s contribution. Particle-only scoring does not fully resolve this attribution issue because the particle and carbon are jointly fitted, so changes in the carbon model can also affect the recovered particle. Second, the best reference agreement occurs at pass 100 with fixed centers on a dense 1.6M-Gaussian grid, alongside a 216M-voxel carbon field (Section 5.3, Table 2; Appendix C). The operator speedups over multislice, 3.7x at 1.6M Gaussians and 139x at 20k (Table 7), do not establish that a compact Gaussian representation retains comparable reconstruction quality or an end-to-end cost advantage. At the pass-100 checkpoints with the highest reference agreement, TPGA-GS takes 125 minutes versus 97 for Ray-GS (Table 2). The complete 150-pass runs take 188 and 146 minutes, respectively, but the additional center refinement reduces reference agreement. These timings quantify the overhead of wave physics; they do not establish Gaussian efficiency relative to other representations. The roughly 183-minute PhaseT3M figure is extrapolated under different optimization settings, while the substantially faster voxel-TPGA run lacks a corresponding reconstruction-quality comparison (Appendix F). A matched quality, runtime, and memory comparison with voxel TPGA would help determine what the Gaussian representation adds beyond the propagation approximation.

Questions for authors
  1. Independent 3D accuracy.

What limits the held-out predictions in Figure 22, and how should this test be interpreted alongside the reference-volume advantage? A small inverse experiment using a known 3D specimen and noisy tilt images generated by an independent multislice simulator would materially affect my assessment: does TPGA-GS recover more accurate 3D structure than Ray-GS near the frequencies where the forward models separate? Please use matched support and optics and report a direct 3D error or Fourier correlation, including the missing-wedge region.

For the measured data, repeats with different optimization seeds would assess stability. For the simulated experiment, independent noise realizations would assess robustness to observation noise.

  1. Attribution to Gaussian propagation.

Keeping the carbon propagation model and all other settings fixed, how much improvement remains when only the particle renderer changes from ray projection to complex-covariance propagation, with both components jointly refitted? A full 2x2 ablation independently switching particle and carbon propagation, if feasible, would additionally distinguish their contributions and interactions. Please report both particle-only and full-field metrics, since joint fitting can couple the two components.

Practical benefit of the representation. Can the authors first report reconstruction quality for the existing voxel-TPGA timing benchmark and compare it with TPGA-GS under matched data, optics, support, objective, and hardware? Please report optimization and stopping settings, and pair quality with runtime and peak memory at the same checkpoints. A comparison over additional Gaussian counts would be a useful extension. Please distinguish measured comparisons from the extrapolated PhaseT3M runtime. Demonstrating a useful accuracy-cost advantage would strengthen my assessment; reproducing the specific 20k-primitive benchmark speedup in reconstruction is not a prerequisite.

Definition of the {400} score. Section 5.3 describes shell power at lattice positions normalized by a phase-randomized copy with identical Fourier magnitudes. Appendix C instead describes the shell’s 99.95th-percentile-to-mean power ratio, again normalized after phase replacement. Phase-only randomization preserves every Fourier power value. With the Fourier grid, shell, and lattice-selection mask held fixed and no further processing, either statistic would therefore be unchanged and the normalized score would be 1. Are additional real-space masking, cropping, mask re-estimation, or other processing steps applied after randomization? Please reconcile the two descriptions and provide the operation order or scoring code needed to reproduce the reported values in Table 2, including 1.04 and 1.78. This may be an omitted processing detail, but it affects whether the score supplies interpretable structural evidence of lattice recovery.

1 comment
AuthorsOpenReview ↗

Common Response

1. Independent 3D accuracy and repeats [CW1, Q1]

Following the reviewer’s request, we tested 3D accuracy using independent abTEM multislice images and two Poisson noise realizations. Wave and ray share support, initialization, starting optics, optics refinement and a 100-pass budget. Fourier correlations to the known potential over 2.4 to 1.6 Å favor wave, including in the missing wedge (Figure 5, Table 10).

ConditionWave: covered / wedgeRay: covered / wedge
Simulation, noise 10.695 / 0.3300.542 / 0.212
Simulation, noise 20.705 / 0.3550.251 / −0.158
Measured, seed 20290.8276 / 0.74080.6566 / 0.6097
Measured, seed 20300.8312 / 0.76470.6504 / 0.6058

As requested, we checked stability by changing the tilt order in new all-Gaussian fits with refined optics, which differ from the submitted hybrid and measure agreement with PhaseT3M rather than ground truth (Appendix B.1). The simulated noise draws use the same specimen, and “covered” means directions within the tilt range, not exact CTF coverage.


2. Held-out prediction [CW1, Q1]

We agree that reference agreement does not override the unfavorable holdout, so we expanded it to five angles. Ray still has lower amplitude error in the six finer bands, and both methods have NMSE above 1 throughout (Table 9). The cause remains unresolved, and we do not claim measured 2 Å prediction or structural resolution.

Simulation favors wave in every evaluated band finer than 10 Å against clean and noisy targets (Figure 10, Table 12). However, its full-band optics calibration and amplitude correlations differ from the measured test, so it does not resolve that discrepancy.


3. Propagation attribution [CW2, Q2]

To address the reviewer’s concern about carbon coupling, we tested all four wave/ray combinations in the particle-containing inner region and the outer carbon. These regional controls show that switching the inner region to wave improves agreement when outer-carbon propagation is wave-based, but reduces covered-direction particle agreement when it is ray-based (Table 11, particle and full-field scores).

The benefit therefore depends on the carbon model, and nearby carbon within the inner region prevents particle-only attribution. The regional ray/ray control also differs from the main Ray-GS baseline because it sums propagated potentials before applying one exponential.


4. Reconstruction quality and cost [CW2, Q3]

We agree that operator timings alone do not establish reconstruction efficiency, so we paired quality, runtime and memory at pass 100 on an RTX 4090. The 143.9M-site Gaussian fits achieve comparable reference agreement in about 42% less recorded time, with higher memory (Tables 3 and 18).

MethodCovered / wedgeMinutesLive CuPy GiB
TPGA-GS, seed 20300.8312 / 0.764760.5513.77
TPGA-GS, seed 20290.8276 / 0.740861.8513.77
Voxel TPGA0.8310 / 0.7526104.159.92

All use zero starts, optics refinement and fixed 100-pass budgets without early stopping. Support, preconditioning, regularization and optics updates differ, so this compares complete solvers, not representation alone. Gaussian time includes preparation and fitting, voxel time reconstruction (Appendix E).

We then evaluated reconstruction quality for the voxel-TPGA timing benchmark, as the reviewer requested. With deposited optics held fixed, voxel TPGA reaches a correlation of 0.949 near 2 Å in 29.3 minutes, compared with 0.931 in 125 minutes for the submitted hybrid (Table 18). Different optics and scoring bands prevent direct comparison with the table above, and the memory measurements use different counters. The submitted operator speedups measure individual updates, not end-to-end reconstruction speed.


5. Gaussian-count sensitivity [CW2, Q3]

As the reviewer suggested, we tested seven Gaussian counts (Table 19), and halving 143.9M to 72.1M cuts runtime from 65.8 to 37.5 minutes while lowering covered-direction correlation from 0.828 to 0.797. Wave stays above ray throughout, but further reductions lose quality, so we do not claim an atom-like compact solution. Timings come from the separate count study.


6. The {400} descriptor [CW1, Q4]

The reviewer is correct that phase replacement alone preserves Fourier power, but we then inverse-transform, subtract the mean and reapply the Hann window, which changes the control’s power. The score divides the original shell’s 99.95th-percentile-to-mean power ratio by the control’s, with crop, mask and shell fixed. Appendix C corrects the reference to lattice positions and reproduces 1.04 and 1.78 (Table 14). This window-dependent descriptor is not independent evidence of accuracy or resolution.


We appreciate the reviewer's questions and feedback. References above are to the revised manuscript unless mentioned as submitted.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response

We appreciate the reviewers’ time, questions and feedback. Common concerns were the strength of the structural evidence, fairness of the baseline comparisons, attribution of improvements to wave propagation, and practical benefit of the Gaussian representation. The additions and revisions are marked in blue in the manuscript.

The main additions address those concerns:

  • 3D accuracy and robustness: TPGA-GS recovers the known simulated potential more accurately than Ray-GS in both independent noise realizations, including within the missing wedge. Repeated measured-data reconstructions assess sensitivity to optimization seeds.
  • A high-resolution propagation test: matched AAV2 comparisons favor wave propagation in high-frequency reference agreement, opposite-half image prediction and atom-local Q-scores.
  • Baseline fairness and attribution: CTF-corrected biological comparisons and component controls assess the contributions of image preparation, Gaussian adaptation, alignment and defocus refinement, and propagation.
  • Practical benefit and reproducibility: we report reconstruction quality alongside runtime and memory, separately from operator timings, and clarify initialization, Gaussian counts, optimization settings and the {400} scoring procedure.

We have revised the claims to match the evidence and address the remaining limitations explicitly in the individual responses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.