Pith. sign in

REVIEW 3 major objections 22 references

Tomo-center: an AI-based rotation-axis center finder for synchrotron micro- and nano-tomography

T0 review · 3 major / 0 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read A vision-transformer classifier finds the tomography rotation center to sub-pixel accuracy from try reconstructions, even under sparse or noisy projections.

desk verdict Solid engineering paper: modern DINOv2+MIL COR classifier with real beamline packaging and better noise robustness than Vo; the big OOD table is mostly agreement with Vo-as-GT, so treat the headline MAE carefully. read the letter →

arxiv 2607.09916 v1 pith:WMXRQRTV submitted 2026-07-10 eess.IV

classification eess.IV
keywords tomographycenterofrotationdeeplearningvisiontransformermulti-instancesynchrotronimagingDINOv2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

In parallel-beam synchrotron tomography the rotation-axis center must be known to a fraction of a pixel; a wrong value produces double-edge and streak artifacts that ruin the volume. Classical sinogram or keypoint methods work on high-contrast samples but break down on weakly absorbing or low-contrast specimens. This paper shows that the problem can be treated as binary classification of “try reconstruction” slices: a DINOv2-pretrained vision transformer plus attention-based multiple-instance learning is fine-tuned to score candidate centers, and the highest-scoring center is returned. On two independent test collections the method consistently keeps mean absolute error below one pixel and remains usable when the number of projections is cut by up to tenfold or Poisson noise is increased by a blank-scan factor of ten. The same model has already been wired into production reconstruction packages so that beamline users can obtain an automatic center without manual inspection.

What carries the argument

Attention-based multiple-instance learning over random windows of a DINOv2-pretrained ViT-B/14: each try-reconstruction slice is treated as a bag of patches whose class tokens are gated-attention-pooled into a binary correct/incorrect score; the candidate center with the highest score is selected.

What would settle it

Re-measure the same 302-scan thermal-ramp series with an independent mechanical or high-precision fiducial center and check whether the AI method’s mean absolute error still stays below one pixel when Vo is no longer the reference.

Watch

Extended reading notes

Core claim

A fine-tuned DINOv2 vision transformer with attention-based multiple-instance learning can discriminate correctly centered from mis-centered tomographic slices with enough accuracy that the highest-scoring candidate recovers the rotation-axis center to a mean absolute error below one pixel on two independent data sources, remaining robust under tenfold projection sparsity or blank-scan factor of ten.

Load-bearing premise

On the largest out-of-distribution benchmark the paper treats Vo’s operator-validated centers as ground truth, so reported accuracy is largely agreement with Vo rather than an independent mechanical standard.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The manuscript presents tomo-center, a learning-based rotation-axis (COR) estimator for parallel-beam synchrotron tomography. Candidate centers are swept to produce try-reconstruction slices; a DINOv2-pretrained ViT-B/14 backbone with gated attention MIL (Eq. 1) and a binary head classifies each slice as correctly or incorrectly centered, and the highest-scoring center is returned. Training uses ~110k try-reconstruction tomograms from 546 retrospective APS experiments with expert COR labels (±1 px tolerance). Evaluation covers a 302-scan out-of-distribution in-situ carbonation campaign (Tables 1–4), eight TomoBank cases with expert GT (Tables 5–6), sparse-angle and Poisson blank-scan stress tests, attention-map interpretability (Fig. 3), and production integration into tomocupy and TomoGUI. Reported headline results are mean absolute error below 1 px on both sources and superior robustness to blank-scan factor 100 relative to Vo.

Significance. COR calibration remains a practical bottleneck for automated high-throughput reconstruction at synchrotron beamlines, especially for low-contrast, evolving, or dose-limited samples where classical sinogram or 0°/180° methods can fail. A production-ready open-source tool (tomo-center) already integrated into tomocupy and TomoGUI, with explicit stress tests under sparse angles and Poisson noise and with attention-based interpretability, is a concrete contribution to beamline operations. The DINOv2 + attention-MIL design is a sensible transfer of modern vision backbones to a bag-level tomography task. If the accuracy claims hold under independent ground truth, the work would be a useful, deployable advance rather than a purely methodological curiosity.

major comments (3)
  1. Section 3.1 and Tables 1–4: the primary 302-scan OOD benchmark adopts Vo’s per-scan COR as ground truth (operator-validated production reference, already within ~5 px of nx/2). Absolute errors of the proposed method are therefore almost entirely measures of concordance with Vo (mean 0.32 px, 301/302 within 1 px), not independent recovery of a mechanical or expert axis. The abstract and §4.3.3 claim of “mean absolute error of below 1 pixel” on two independent sources is therefore overstated for the largest table; the TomoBank n=8 expert-GT set (Table 5, mean 1.00 px) is too small to carry the claim alone. Either re-label a substantial subset of the 302 scans with independent expert or mechanical GT, or reframe Tables 1–3 as agreement/stability relative to Vo and reserve absolute-accuracy language for expert-GT data.
  2. Section 3.3 and Tables 4, 6 (low-photon rows): stress-test absolute errors are still computed against the unperturbed Vo (or expert) reference. Superiority of the proposed method under blank-scan factor 100 (median 0.50 px vs Vo 8.25 px on the in-situ subset) is real relative to Vo, but does not by itself establish absolute accuracy if both methods share correlated failure modes on noisy or low-contrast data. Report at least a small expert re-annotation of the noisy reconstructions, or qualify the claim as relative robustness rather than absolute sub-pixel accuracy under noise.
  3. Section 2.2.3 and abstract: training labels use a ±1-pixel tolerance around expert COR, and inference selects the argmax over a 0.5-pixel sweep. The reported sub-pixel MAEs are therefore partly constrained by label granularity and by the discrete candidate grid. Clarify how continuous or sub-0.5-pixel accuracy is claimed, and whether the tolerance window systematically biases the validation MAE toward values ≤1 px.

Circularity Check

1 steps flagged · score 2.0 of 10

Empirical classifier; primary OOD table measures agreement with Vo-as-GT rather than independent mechanical truth, but no algebraic self-definition or forced prediction.

  1. fitted input called prediction [Section 3.1 and Table 1 (primary OOD benchmark)]
    "We adopt Vo's per-scan output as the ground truth for this benchmark: Vo's values were used to produce the production reconstructions delivered to the user and were operator-validated by inspecting the reconstructed volumes against the mechanically-aligned axis. Vo's per-scan values coincided with the geometric image centre nx/2 to within ∼ 5 pixels for every scan in this campaign. ... Table 1: Absolute deviation from the Vo reference on the 302-scan out-of-distribution benchmark."

    Absolute error of the proposed method is defined as deviation from Vo, which is itself the production reference already known to lie near the geometric centre. The reported mean 0.32 px / 301-of-302 within 1 px therefore measures agreement with the classical baseline rather than recovery of an independent mechanical or expert COR on that set. This is evaluation circularity (fitted/reference input treated as the accuracy claim), not algebraic self-definition of the classifier.

full rationale

This is a supervised deep-learning paper, not a first-principles derivation. Training labels are expert-annotated COR with a ±1-pixel tolerance (Section 2.2.3); inference selects the candidate with highest softmax score among try-reconstructions. No equation equates the network output to its training labels by construction, and no uniqueness theorem or ansatz is imported via self-citation to force the result. The only mild circularity is evaluation design: on the 302-scan in-situ OOD benchmark (Section 3.1, Tables 1–4) the authors adopt Vo’s per-scan COR as ground truth because it was operator-validated and used for production reconstructions, so the headline sub-pixel MAE largely reports concordance with the classical baseline under conditions where Vo already sits near nx/2. TomoBank (n=8) uses expert GT and is independent. Stress-test superiority under high Poisson noise is a genuine relative comparison against the same reference. Score 2 reflects that single non-load-bearing evaluation choice; the method itself is not circular.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The work is applied ML for instrument calibration. Load-bearing content is empirical: pretrained features transfer, bag-level labels from expert COR, and try-reconstruction quality as a proxy for axis correctness. Free parameters are training/inference hyperparameters and the label tolerance; axioms are standard tomography and MIL assumptions; no new physical entities are postulated.

free parameters (5)
  • COR label tolerance (±1 pixel around expert COR)
    Defines binary training labels; directly shapes what the classifier treats as “correct” (Section 2.2.3).
  • Patch size and count (518×518; K=24 train, K=3 inference)
    Chosen configuration; authors note smaller/fewer patches can miss artifacts and hurt accuracy (Discussion).
  • Learning rate 5e-6, AdamW, 10 epochs, cosine schedule, warmup 500 iters, batch size 2
    Optimization hyperparameters that determine the shipped checkpoint; epoch-1 weights selected by validation MAE.
  • Candidate sweep range/step (e.g. 200 or 400 steps at 0.5 px)
    Search grid at train/val/test; method returns argmax score on this discrete set.
  • Secondary crop scale set {66%, 75%, 87.5%, 100%} and 13 offsets; 90° rotations
    Augmentation design choices that affect invariance to ROI/orientation.
assumptions (5)
  • domain assumption Mis-set COR produces characteristic double-edge and tangential streak artifacts concentrated in the sample foreground of FBP-style reconstructions.
    Motivation and interpretability narrative (Introduction; Fig. 3 discussion); classifier is trained to detect these visual cues.
  • domain assumption DINOv2 ViT features pretrained on natural images transfer usefully to synchrotron tomograms after fine-tuning.
    Backbone initialization (Section 2.2.2); no in-domain pretraining from scratch.
  • domain assumption Bag-level binary supervision with attention MIL aggregation is sufficient to score whole-slice COR correctness from random windows.
    Architecture and loss (Eqs. 1–2; Ilse et al. MIL framing).
  • domain assumption Expert-annotated (or production Vo) COR is an adequate proxy for the true mechanical rotation axis within the stated pixel tolerance.
    Training labels and Section 3.1 ground-truth choice.
  • standard math Softmax / gated attention and cross-entropy training are standard and well-defined.
    Eqs. 1–2; no novel math claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tomo-center: an AI-based rotation-axis center finder for synchrotron micro- and nano-tomography." pith.science (2026). https://pith.science/paper/WMXRQRTV

@misc{pith2026260709916,
  author       = {Pith},
  title        = {Pith review of: Tomo-center: an AI-based rotation-axis center finder for synchrotron micro- and nano-tomography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WMXRQRTV}},
  note         = {Machine review of arXiv:2607.09916}
}
read the original abstract

Accurate determination of the rotation-axis position is a prerequisite for artifact-free reconstruction in parallel-beam synchrotron micro-tomography. Traditional approaches such as Vo's method rely on sinogram features that can fail for low-contrast or weakly absorbing specimens. We present a learning-based method that treats center selection as a binary classification problem, using a DINOv2-pretrained vision transformer aggregated with attention-based multiple-instance learning, fine-tuned end-to-end on tomographic images. At inference time, the proposed algorithm was applied to a stack of tomograms reconstructed at a sweep of candidate centers to select the optimal center for reconstruction. We tested the estimation accuracy of the proposed method on two independent data sources and consistently achieved a mean absolute error of below 1 pixel. We also tested the method robustness to sparse or noisy acquisitions with the same datasets and demonstrated consistent performance when the number of projections was reduced by a factor of up to 10 or the blank scan factor of the underlying Poisson's noise was increased to 10. We also illustrated the interpretability of the proposed method by mapping out the relative contributions of continuous spatial features to the overall classification task. This method, delivered as tomo-center, an open-source command-line tool, has been integrated into several tomography software packages to assist experiments during the routine beamline operations.

Figures

Figures reproduced from arXiv: 2607.09916 by the authors.

Figure 1
Figure 1. Schematic of the proposed framework to predict the correct COR based on the corre [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Training losses (A) and validation errors (B) of the proposed model to classify the COR [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Selected tomograms (row 1) and their corresponding distributions of spatial importance [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Selected tomographic slices reconstructed with the correct COR (columns A and B), [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: The functions and user interface (UI) of tomogui. (A) The main functions and workflow. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 3 canonical work pages

  1. [1]

    & Joulin, A

    Caron, M., Touvron, H., Misra, I., J´ egou, H., Mairal, J., Bojanowski, P. & Joulin, A. (2021). In Proceedings of the IEEE/CVF international conference on computer vision , pp. 9650–9660. De Andrade, V., Nikitin, V., Wojcik, M., Deriy, A., Bean, S., Shu, D., Mooney, T., Peterson, K.,

  2. [2]

    & Lethien, C

    Kc, P., Li, K., Ali, S., Fezzaa, K., G¨ ursoy, D., Arico, C., Ouendi, S., Troadec, D., Simon, P., De Carlo, F. & Lethien, C. (2021). Advanced Materials, 33(21), 2008653. https://advanced.onlinelibrary.wiley.com/doi/abs/10.1002/adma.202008653 De Carlo, F., G¨ ursoy, D., Ching, D. J., Batenburg, K. J., Ludwig, W., Mancini, L., Marone, F.,

  3. [3]

    M., Sijbers, J

    Mokso, R., Pelt, D. M., Sijbers, J. et al. (2018). Measurement Science and Technology, 29(3), 034004

  4. [4]

    G., Lathrop, R

    Dietterich, T. G., Lathrop, R. H. & Lozano-P´ erez, T. (1997).Artificial intelligence, 89(1-2), 31–71

  5. [5]

    & Schreyer, A

    Donath, T., Beckmann, F. & Schreyer, A. (2006). Journal of the Optical Society of America A , 23(5), 1048–1057

  6. [6]

    & Girshick, R

    He, K., Chen, X., Xie, S., Li, Y., Doll´ ar, P. & Girshick, R. (2022). InProceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 16000–16009

  7. [7]

    & Dean, J

    Hinton, G., Vinyals, O. & Dean, J. (2015). arXiv preprint arXiv:1503.02531

  8. [8]

    Ilse, M., Tomczak, J. M. & Welling, M. (2018). In Proceedings of the 35th International Conference on Machine Learning (ICML) , pp. 2127–2136

Show all 22 references
  1. [9]

    & Foster, I

    Liu, Z., Bicer, T., Kettimuthu, R. & Foster, I. (2019). In 2019 IEEE/ACM Third Workshop on Deep Learning on Supercomputers (DLS), pp. 20–28. IEEE

  2. [10]

    Lowe, D. G. (2004). International Journal of Computer Vision , 60(2), 91–110

  3. [11]

    Nikitin, V. (2023). Journal of Synchrotron Radiation , 30(1), 179–191. https://doi.org/10.1107/S1600577522010311

  4. [12]

    & De Carlo, F

    Nikitin, V., Tekawade, A., Duchkov, A., Shevchenko, P. & De Carlo, F. (2022). Journal of Syn- chrotron Radiation, 29(3). https://doi.org/10.1107/S1600577522003095 21

  5. [13]

    V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H. V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A. et al. (2024). In Transactions on Machine Learning Research. ArXiv:2304.07193

  6. [14]

    & Lamard, M

    Quellec, G., Cazuguel, G., Cochener, B. & Lamard, M. (2017). IEEE reviews in biomedical engi- neering, 10, 213–234

  7. [15]

    Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X. et al. (2021). Advances in neural infor- mation processing systems, 34, 2136–2147. Sim´ eoni, O., Vo, H. V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M. et a...

  8. [16]

    & Clark, S

    Tang, S., Bicer, T., Fezzaa, K. & Clark, S. (2025 a). Frontiers in High Performance Computing , 3, 1537080

  9. [17]

    & Clark, S

    Tang, S., Bicer, T., Sun, T., Fezzaa, K. & Clark, S. J. (2025 b). Synchrotron Radiation, 32(2)

  10. [18]

    & De Carlo, F

    Tang, S., Nikitin, V. & De Carlo, F. (2026). tomo-center: Ai rotation-axis estimation for parallel- beam tomography. Software release v1.0.0. https://github.com/xray-imaging/tomo-center/releases/tag/v1.0.0

  11. [19]

    T., Drakopoulos, M., Atwood, R

    Vo, N. T., Drakopoulos, M., Atwood, R. C. & Reinhard, C. (2014). Optics Express , 22(16), 19078–19086

  12. [20]

    Xie, Q., Luong, M.-T., Hovy, E. & Le, Q. V. (2020). In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 10687–10698

  13. [21]

    & G¨ ursoy, D

    Yang, X., De Carlo, F., Phatak, C. & G¨ ursoy, D. (2017).Synchrotron Radiation, 24(2), 469–475

  14. [22]

    & Kong, T

    Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A. & Kong, T. (2021). arXiv preprint arXiv:2111.07832. 22

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.