acceptodds
Under review as a conference paper at ICLR 2027

The Sharp Plane Is Tilted: Anisotropy Priors Must Follow the Stored Affine, Not the Grid

Abstract

Restoration models for anisotropic 3-D data need one piece of physics: the direction of worst resolution. Recent 3-D microscopy restoration methods hard-code it as the grid axis . We show that this is a statement about the resampling that produced the stored volume: under a linear resample of the acquisition, the plane of best resolution maps to and its normal to , so a grid-axis prior is exact only when the transported axis still lands on a grid axis. For stage-scanned light-sheet data is the deskew affine the file records; when it leaves the camera plane tilted, the worst axis is or , both read from the file and ordered by the ratio of 's column norms. On 107 blocks from 18 open-top light-sheet biopsies (7 patients) the spectrum's sharp lobe lies along the camera axis predicts in 103/107 blocks, and a sweep that uses no network finds the cheapest direction to under-sample on the median over blocks. On held-out blocks decimated along , the restorer that assumes leaves them worse than unrestored in explained variance (EV) at decimation factors and (0/4 patient-disjoint folds improve them). The same network pointed at improves them in 4/4 folds and beats the convention by , and in EV at , and , and by F1 on a nucleus-detection endpoint that does not use EV (). The gap survives the published CARE recipe, and eight restorers trained at fixed axes from to trace a curve that peaks at the file's angle (). On a second instrument (oblique-plane microscopy) a degradation whose stored-frame direction the file sets (every second camera frame dropped before the deskew) is improved only by the restorer pointed at the file's scan direction once its kept-slice mask input is removed; on brain cone-beam CT with an injected oblique axis and a true reference the rotated prior beats the prior in 9/9 foldseed pairs. The fix is one signed angle per stored volume, read from the file.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anisotropic 3D restoration, light-sheet microscopy, deskew affine, inductive bias, self-supervised super-resolution, image restoration, microscopy image analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dThjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the misalignment between assumed and actual anisotropy directions in 3D microscopy restoration. The authors argue that conventional methods often assume the lowest-resolution direction to be aligned with the grid's z-axis, whereas geometric transformations during image acquisition and deskewing may produce oblique anisotropy directions. To address this issue, the paper proposes using the stored affine transformation to determine the appropriate anisotropy direction and align the restoration prior accordingly. Experiments on light-sheet microscopy and synthetically degraded CT data demonstrate improvements over conventional z-axis assumptions.

Critical strengths
  1. The paper identifies a practically relevant issue in anisotropic 3D image restoration, emphasizing the importance of incorporating acquisition geometry into restoration models.
  2. The experimental evaluation includes multiple control settings, different restoration architectures, and analyses of directional misalignment.
  3. The proposed correction is straightforward to implement and can potentially improve existing restoration methods without modifying their network architectures.
Critical weaknesses
  1. The central contribution relies on standard affine geometry, particularly the transformation of plane normals through the inverse transpose of the affine matrix. Similar geometric considerations are already established in optical imaging and PSF-aware deconvolution. The proposed solution primarily consists of identifying the anisotropy direction from metadata and rotating the input or restoration prior accordingly. While practically useful, this represents a correction of an existing modeling assumption rather than a substantial methodological advancement in machine learning. The paper does not introduce a fundamentally new restoration formulation, learning objective, or network architecture.
  2. A major concern is that the primary microscopy experiments evaluate restoration against the already acquired and deskewed volumes, rather than an independent high-resolution or isotropic ground truth. The evaluation therefore primarily measures how well different methods recover synthetically removed information, rather than whether they recover genuinely missing structural details caused by the original imaging process.
Questions for authors
  1. Can the authors provide stronger evidence that the proposed correction improves the recovery of genuinely missing structures in real microscopy data, rather than primarily improving reconstruction under synthetic decimation?
  2. How reliably does the proposed approach transfer to other acquisition systems with different scanning geometries, PSFs, and spatially varying anisotropy?
Reviewer_bwngOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines how affine resampling can misalign the resolution direction of 3D microscopy data with the grid axis assumed by restoration models. It uses metadata and directional measurements to identify a suitable axis and correct the restoration pipeline. Controlled experiments show that aligning the prior with the degradation direction improves reconstruction.

Critical strengths
  1. The paper identifies a practical coordinate mismatch and proposes a simple correction that can be incorporated into existing restoration pipelines.

  2. Shared-input comparisons, patient-disjoint folds, rotation controls, and mask ablations provide convincing evidence that axis alignment matters in the tested settings.

Critical weaknesses
  1. Most restoration gains are evaluated after additional downsampling, using the already anisotropic stored volume as the reference. Official-model comparisons also involve re-pooling confounds, leaving the impact on existing methods in their intended workflows insufficiently established.

  2. The geometric transformation is established knowledge, so the contribution relies mainly on its empirical significance. Moreover, the axis-selection rule depends on PSF anisotropy and interpolation, limiting the claim that the appropriate axis can be determined from the affine alone.

Questions for authors
  1. Can you demonstrate the correction within an existing method’s intended restoration workflow, using identical inputs and eliminating the additional pooling confound?

  2. Under what conditions can the affine and sampling ratio reliably determine the appropriate axis without additional PSF information?

  3. Could the authors evaluate the proposed correction on real restoration tasks, such as isotropic reconstruction of 3D fluorescence microscopy volumes? It would help determine whether the correction improves structural recovery and strengthen the practical relevance and reliability of the conclusions.

Overall, I find this paper quite interesting. If the authors can provide evidence demonstrating its effectiveness when applied to downstream tasks, I would be inclined to raise my score.

Reviewer_tJwYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper raises a practical issue that degradation-direction mismatch can hurt reconstruction performance.

However, the proposed correction is based on well-established geometry, and the main experiments largely demonstrate the benefit of matching synthetic training and test degradation.

The lack of convincing qualitative results for the main comparison and limited evidence on actual image restoration make it difficult to assess the practical significance of the reported gains.

This paper is really hard to read for human readers. It's probably a document generated directly by AI. There are many issues about grammar, format, figures, etc.

Critical strengths

The paper identifies a practical issue in existing 3D restoration methods. The degradation direction may change after geometric transformations, while some learning-based methods still assume a fixed axis. This is a useful observation.

Critical weaknesses
  1. Limited novelty. The transformation of directions and plane normals under an affine matrix is a standard geometric result. Accounting for acquisition geometry is also common in traditional image restoration, particularly PSF-based deconvolution. The proposed correction mainly applies this existing knowledge to the training and inference procedures of learning-based restoration methods. It is unclear whether this constitutes a sufficiently new methodological contribution for ICLR.

  2. No experimental results on real data, and the actual restoration quality is not demonstrated convincingly. In the main experiments, the authors take already reconstructed microscopy volumes, artificially downsample them along a chosen direction, and evaluate how well the models recover the original stored volumes. This is useful for studying direction mismatch, but it does not show whether the method can recover details lost during actual image acquisition. More importantly, the main paper provides very limited qualitative restoration results. Figure 4 in the appendix contains one visual comparison, but it uses the proxy-task setting rather than the shared-degradation setting used for the main quantitative results. For an image restoration paper, I would expect representative visual comparisons on the main experiments, including difficult cases and actual microscopy images, to understand what the reported improvements mean in practice.

  3. A mismatch between training and test degradation. The proposed model is trained using degradation along the direction used to generate the test inputs, while the baseline is trained with degradation along the z-axis. It is therefore not surprising that the direction-matched model performs better. The reverse experiment in Appendix J, where the z-axis model performs better when the test degradation is along z, further supports this explanation. These results show that matching the degradation direction matters, but they do not fully establish a new restoration capability.

  4. The paper is unnecessarily complicated for the main finding it presents. A large number of experimental variants, controls, and numerical comparisons are introduced, but many support essentially the same conclusion: a model performs better when its assumed degradation direction matches the test data. The repeated emphasis on fold counts, seed counts, and control experiments makes it difficult to identify what is genuinely new. A more focused presentation, together with clear visual results, would make the contribution easier to assess.

Questions for authors
  1. Can the authors provide qualitative restoration results for the shared-degradation experiments in Table 1, rather than only the proxy-task examples in Appendix Figure 4? What visible structures or details are actually improved?

  2. Can the authors demonstrate improvements on real microscopy restoration tasks without adding artificial downsampling to the stored volumes? If such evaluation is not possible, how should the practical restoration claims be interpreted?

  3. Since affine transformation of PSFs and resolution directions is already known in computational microscopy, what is the main methodological contribution beyond applying this correction to existing learning-based pipelines?

  4. Could a model trained with sufficient directional augmentation achieve similar results without explicitly reading the affine metadata? How does the proposed method compare under matched training budgets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.