acceptodds
Under review as a conference paper at ICLR 2027

TomoDETR: Point-Set Prediction with Cross-Window Refinement for Cryo-ET Particle Localization

Abstract

Particle localization is a critical step in cryo-electron tomography (cryo-ET), providing the coordinates needed for downstream structural analysis. State-of-the-art supervised methods have achieved strong localization performance by first predicting dense particle masks or response maps and then extracting center coordinates. Yet the target itself is a labeled set of particle centers. We therefore develop TomoDETR, a DETR-based framework that formulates particle localization as point-set prediction from point annotations. Large tomograms require overlapping-window inference, causing the same particle to be predicted repeatedly from different window contexts. We find that these repeated observations differ in localization quality, while classification confidence does not reliably identify the most accurate one. Motivated by this finding, TomoDETR incorporates a Cross-Window Query Refiner (CWQR), which learns interactions among nearby candidates from different windows to jointly refine candidate centers and class scores before final selection. On the CZII particle localization benchmark, TomoDETR achieves the highest weighted F4 among the evaluated methods. Under a shared backbone and matched training setup, direct point-set prediction improves weighted F4 by 2.30 points over dense prediction with center extraction, while CWQR provides a further 2.85-point improvement with only 2.9% measured inference-time overhead. Additional evaluations on cellular and simulated tomograms further support the effectiveness of cross-window refinement. Our code will be released publicly upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-Electron Tomography, Particle Localization, Set Prediction, 3D Object Detection, DETR

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XtB1OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes TomoDETR, a DETR-style detector for cryo-ET particle localization. Instead of predicting dense masks or heatmaps and then extracting centers, it regresses particle centers and classes directly from 64 grid-initialized queries in each overlapping 96³ window (Deformable DETR decoder, Hungarian matching). They add a second stage, the Cross-Window Query Refiner (CWQR). It is a sparse transformer over candidates coming from different windows, which refines centers and class logits before NMS. Validation is peformed on CZII, SHREC2021 and DeePiCt (VPP and defocus) benchmark datasets.

Critical strengths

The method is applied to an interesting problem, on several datasets.

Critical weaknesses
  1. The method is very hard to understand.
  • The abstract, introduction and Fig. 1 never state what are the network outputs. Only Sec. 3.2 reveals that it regresses continuous centers from queries decoded over U-Net features.
  • The notation is heavily overloaded. x denotes annotated, candidate and final centers. r denotes reference points, predicted centers and the graph radius. R denotes class radius and recall. P denotes precision and window width. β denotes the F-score parameter, a protein class and a focal-loss exponent. g denotes window geometry and the prediction heads.
  • Several quantities are undefined or only defined in the appendix: g^win_i, E_q, ϕ_g, ϕ_w (Eq. 4), z and Ω (Eq. 6). The value r_g = 9.4614 is never justified.
  1. The baseline comparison may not be fair, and the gains outside CZII are not convincing.
  • The CZII calibration set (40 tomograms) is larger than the training set (7). SHREC2021 uses 1 selection and 1 test tomogram, and DeePiCt uses 2 calibration and 2 test tomograms per split. The clearest gains appear on CZII, the dataset with by far the largest selection and test budget.
  • On DeePiCt, the margins over DeepETPicker (+0.87 on VPP, +1.63 on defocus) fall within the standard deviations across splits, and no paired comparison is given.
  • On SHREC2021, TomoDETR without CWQR (86.99) is below DeepETPicker (92.06), and the evaluation relies on a single test tomogram.
  • External baselines are single runs. It is not stated whether they received the same tuning effort or the same select-then-retrain protocol (settings chosen on tomogram 8, retrained on 0–8).
  1. Limited methodological novelty. The first stage is a standard 3D Deformable DETR (Hungarian matching, L1+CE losses, deformable attention, auxiliary losses) with grid-initialized reference points. Point-based set prediction with Hungarian matching exists in other domains (e.g., crowd localization), and CWQR is conceptually close to learned NMS and relation-based refinement. The mismatch between confidence and localization quality is also known in object detection. The contribution is mainly an adaptation to cryo-ET plus a second-stage refiner, so its value depends on the empirical evidence, which is strongest on CZII only (with a lot of extra data for finetuning).
Questions for authors
  1. Did the baselines receive the same hyperparameters tuning and the same select-then-retrain protocol? Can you report paired per-split differences against DeepETPicker on DeePiCt?
  2. How was r_g = 9.4614 chosen? Please define g^win_i, z, Ω, E_q, ϕ_g and ϕ_w precisely, and state whether decoder updates are absolute or relative.
  3. Why is full CWQR 72.43 in Table 9 but 72.92 in Tables 1 and 2?
  4. Does the 1.16 GPU•h training cost include generating the cached candidates used to train CWQR?
5 comments
AuthorsOpenReview ↗

Response to Reviewer XtB1 (1/5)

W1 + Q2. Can the outputs, notation, and definitions be stated precisely, and how are r_g and the decoder updates defined?

Reviewer concern. The method is very hard to understand.

The abstract, introduction and Fig. 1 never state what are the network outputs. Only Sec. 3.2 reveals that it regresses continuous centers from queries decoded over U-Net features.

The notation is heavily overloaded. x denotes annotated, candidate and final centers. r denotes reference points, predicted centers and the graph radius. R denotes class radius and recall. P denotes precision and window width. β denotes the F-score parameter, a protein class and a focal-loss exponent. g denotes window geometry and the prediction heads.

Several quantities are undefined or only defined in the appendix: g^win_i, E_q, ϕ_g, ϕ_w (Eq. 4), z and Ω (Eq. 6). The value r_g = 9.4614 is never justified.

How was r_g = 9.4614 chosen? Please define g^win_i, z, Ω, E_q, ϕ_g and ϕ_w precisely, and state whether decoder updates are absolute or relative.

Response.

We are grateful for this careful reading and agree that the presentation needed to be clearer. We revised Sec. 3 as follows.

Network outputs. The abstract now states that the network outputs a set of continuous 3D particle centers with class labels and confidence scores.

Notation. The overloaded symbols are renamed:

PreviouslyNow
x for annotated, candidate, and final centers𝒴 = {(y_k, c_k)} for annotations; x_i, x′_i, x̂_n for candidates, refined candidates, and final predictions
r^ℓ_mi (decoder centers); r_g (graph radius)û^ℓ_mi; ρ
P, R in the F4 definition; P (window width)Prec, Rec; L
β, α (focal-loss exponents)described in words (Appendix A.3); β is now only the F-score parameter (the β in β-galactosidase is part of the protein name)
g^win_i; g_z, g_xexplicit window inputs (below); Head_z, Head_x
D × H × WS_z × S_y × S_x

Definitions. Sec. 3.3 now defines every term of Eqs. 4 and 6. The exact formulas and dimensions are in Appendix A.1.

  • E_q is a linear layer applied to the decoder query feature q_i.
  • φ_g is an MLP on the tomogram coordinate normalized to [−1, 1]³.
  • φ_w replaces g^win_i with two MLPs. One encodes the within-window position 2û_i − 1; the other encodes the normalized distance b_i to the nearest window face.
  • z_i ∈ ℝ^(C+1) are the local class logits, including no-object.
  • Ω = [0, S_z − 1] × [0, S_y − 1] × [0, S_x − 1] is the tomogram domain, so clip_Ω clamps refined centers to the tomogram.

Graph radius ρ (previously r_g). ρ is estimated on the CZII training tomograms only. It is 1.1 times the 99th percentile of the distances between candidates from different windows that are assigned to the same particle, which gives 9.4614 voxels. SHREC2021 and DeePiCt use 10 voxels, which equals the SHREC2021 particle radius and the DeePiCt matching tolerance (Appendix A.1).

Decoder updates. The updates are relative. As in Deformable DETR, each layer adds a predicted offset to the previous layer's center in inverse-sigmoid space (Sec. 3.2).

AuthorsOpenReview ↗

Response to Reviewer XtB1 (2/5)

W2 + Q1. Is the baseline comparison fair, and do the gains hold beyond CZII?

Reviewer concern. The baseline comparison may not be fair, and the gains outside CZII are not convincing.

The CZII calibration set (40 tomograms) is larger than the training set (7). SHREC2021 uses 1 selection and 1 test tomogram, and DeePiCt uses 2 calibration and 2 test tomograms per split. The clearest gains appear on CZII, the dataset with by far the largest selection and test budget.

On DeePiCt, the margins over DeepETPicker (+0.87 on VPP, +1.63 on defocus) fall within the standard deviations across splits, and no paired comparison is given.

On SHREC2021, TomoDETR without CWQR (86.99) is below DeepETPicker (92.06), and the evaluation relies on a single test tomogram.

External baselines are single runs. It is not stated whether they received the same tuning effort or the same select-then-retrain protocol (settings chosen on tomogram 8, retrained on 0–8).

Did the baselines receive the same hyperparameters tuning and the same select-then-retrain protocol? Can you report paired per-split differences against DeepETPicker on DeePiCt?

Response.

We appreciate these detailed questions on fairness and generality.

Tuning and protocol. All methods follow the same protocol, and each baseline was tuned on at least as much validation data as TomoDETR.

  • Each external baseline uses its official code and the recipe from its paper or official example.
  • We tuned each baseline on validation data with the selection mechanisms its code exposes (checkpoint or epoch, thresholds, cluster sizes, and post-processing settings; Appendix D.4, Table 21).
  • No method, including TomoDETR, is retrained on tomograms 0–8: all methods train on 0–7, select on 8, and test on 9. The submission's mention of retraining was a description error and is now corrected (Sec. 4.1).

Single runs. We retrained all external baselines with three seeds (DeepFinder: two, as one run produced no predictions). On CZII, TomoDETR (72.92 ± 0.45) leads the strongest baseline, OCTOPI (66.81 ± 0.55), by 6.1 points. Table 1 and Table 5 now report three-seed means.

Selection set on CZII. The 40 selection tomograms (Val40) never update network weights; they only select checkpoints and thresholds. TomoDETR needs only a few of them (weighted F4 on the 364 test tomograms; mean ± SD over the three training seeds, with the N tomograms drawn at random from Val40):

Selection tomograms N13510203040
TomoDETR68.58 ± 2.6070.94 ± 0.8572.03 ± 0.0971.87 ± 0.7072.52 ± 0.5372.34 ± 0.6072.92 ± 0.45

Five selection tomograms already give 72.03. Even with a single tomogram, the mean remains above the strongest baseline (OCTOPI, 66.81 ± 0.55).

DeePiCt, paired comparison. On DeePiCt, our claim is competitive rather than superior performance (Sec. 4.2). We added a per-test-tomogram comparison with DeepETPicker (Appendix D.3, Table 20). The differences (TomoDETR − DeepETPicker), two test tomograms per split, are:

  • VPP: +0.70 and +0.37 (split 0), +0.01 and +1.23 (split 1), +2.95 and +2.33 (split 2). All six are positive, with a mean of +1.27.
  • Defocus: +2.51 and −0.09 (split 0), +0.97 and −2.58 (split 1), +2.96 and +2.95 (split 2). Four of six are positive, with a mean of +1.12.

SHREC2021. To move beyond a single test tomogram, we added two further test tomograms (Appendix D.3, Table 19). TomoDETR has the highest F1 on all three: 93.08, 93.23, and 65.09, against the best baselines' 92.06, 92.69, and 64.28. We agree that TomoDETR without CWQR is below DeepETPicker here. However, CWQR is part of the method, and SHREC2021 shows its largest gain (+6.09).

AuthorsOpenReview ↗

Response to Reviewer XtB1 (3/5)

W3. Is the contribution more than an adaptation of Deformable DETR plus a second-stage refiner?

Reviewer concern. Limited methodological novelty.

The first stage is a standard 3D Deformable DETR (Hungarian matching, L1+CE losses, deformable attention, auxiliary losses) with grid-initialized reference points. Point-based set prediction with Hungarian matching exists in other domains (e.g., crowd localization), and CWQR is conceptually close to learned NMS and relation-based refinement. The mismatch between confidence and localization quality is also known in object detection. The contribution is mainly an adaptation to cryo-ET plus a second-stage refiner, so its value depends on the empirical evidence, which is strongest on CZII only (with a lot of extra data for finetuning).

Response.

We appreciate this candid assessment and agree that the first stage deliberately builds on Deformable DETR. We see the contribution in three points, each backed by controlled evidence.

  1. Point-set formulation for tomograms. Point-based set prediction exists for 2D crowd localization (P2PNet). To our knowledge, however, it has not been used for supervised particle localization in reconstructed tomograms, where dense prediction with center extraction is standard. Table 1(b) compares the two formulations under a shared backbone and training budget.
  2. Cross-window refinement, not single-image suppression. Learned NMS and relation networks operate on detections from one image under one context. CWQR instead jointly refines the coordinates and class scores of repeated observations of the same particle from different windows, under a single tomogram-level one-to-one assignment. Table 2 compares it with fixed fusion rules and with a parameter-matched self-only refiner.
  3. Confidence fails among repeated observations. The confidence–localization mismatch is known in detection (IoU-Net). Our finding concerns repeated observations of the same particle: the highest-confidence observation is the best-localized one only 16.72% of the time, the same as a random draw (16.53%) (Sec. 4.3). This motivates aggregating information across windows rather than selecting by confidence.

CWQR improves the local detector in all four settings, not only on CZII: +2.85 on CZII, +6.09 on SHREC2021, +2.18 on DeePiCt VPP, and +2.58 on DeePiCt defocus. On CZII, only the seven training tomograms update model weights; the 40 selection tomograms only select checkpoints and thresholds (see W2). We now cite P2PNet and IoU-Net in the related work (Sec. 2).

Reviewer_w5xSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses point-set prediction from center annotations to tackle cryo-ET particle localization. A window-level DETR-style detector predicts continuous centers and classes, and a Cross-Window Query Refiner (CWQR) exchanges information among nearby candidates from overlapping windows under tomogram-level Hungarian supervision. The paper reports an improvement in weighted F4 by 2.30 points over the same-backbone Gaussian-heatmap baseline on CZII from DETR, with CWQR further adds 2.85 points. The complete system reaches 72.92 weighted F4 and also improves the local detector’s aggregate results on SHREC2021 and both DeePiCt conditions.

Critical strengths

The paper identifies a practical limitation of independent overlapping-window predictions. CWQR provides a reasonable way to combine their information rather than only suppress duplicates. The experiments include meaningful controls. The dense comparison shares the backbone, data, and update budget. CWQR exceeds the best reported CZII fixed-fusion rule by 2.58 points and a parameter-matched self-only refiner by 1.83 points (Tables 1–2). These results support a contribution beyond additional refinement capacity alone. I also appreciate the level of detail in the results supplied in this paper.

Critical weaknesses
  • I find that the dense comparison does not really isolate the representation effect. While it is true that the shared-backbone comparison is a useful starting point, there are several changes when switching from dense prediction to point sets. From dense to point sets, I think there should be a certain treatment of the grid effect, such as a sub-voxel offset head, but I don't notice any in the current paper. It also seems that the point set prediction has around 50% more parameters than the dense case in the prediction head, if I read Appendix A.5 and understand it correctly.
  • I‘d appreciate more details about the external baselines. For example, the external CZII baselines are without uncertainties. Several details might also be worth documenting, such as the training and tuning, to ensure the reproducibility of the results and comparisons. Further, on SHREC2021, TomoDETR exceeds DeepETPicker in all three listed runs, but evaluation uses one held-out tomogram. DeePiCt reports variation across three fixed splits, rather than repeated training seeds on a fixed split. These evaluations are useful, but their error bars describe different sources of variation and should not be interpreted as equivalent evidence of robustness.
Questions for authors
  1. How much of the Direct-versus-Dense gain remains if the dense baseline gets more granular auxiliaries that compensate for the head parameter count?
  2. Can you provide fuller external-baseline training and selection configurations, and quote some uncertainties for the CZII baseline results, while distinguishing seed variation from split or tomogram variation?
  3. I noticed that in Table 9, the separate RTX 5090 output-replay experiment reaches 72.43 rather than the main 72.92. Which checkpoint, cache, or selection differences explain this gap, and can the mechanism comparison be reconciled with the main endpoint? Does this indicate some reproducibility issue, or am I overly concerned?
3 comments
AuthorsOpenReview ↗

Response to Reviewer w5xS (1/3)

W1 + Q1. Does the Dense-versus-Direct comparison isolate the effect of the representation?

Reviewer concern.

I find that the dense comparison does not really isolate the representation effect. While it is true that the shared-backbone comparison is a useful starting point, there are several changes when switching from dense prediction to point sets. From dense to point sets, I think there should be a certain treatment of the grid effect, such as a sub-voxel offset head, but I don't notice any in the current paper. It also seems that the point set prediction has around 50% more parameters than the dense case in the prediction head, if I read Appendix A.5 and understand it correctly.

How much of the Direct-versus-Dense gain remains if the dense baseline gets more granular auxiliaries that compensate for the head parameter count?

Response.

This is an insightful point, and we agree that the original comparison did not separate the grid effect from head capacity. Following the suggestion, we trained two additional Dense variants under the same protocol, with three seeds each:

  • Dense + offset head regresses each center's sub-voxel offset.
  • Dense, width 160 widens all modules after the backbone to match Direct's parameter count.

CZII, weighted F4 (%); differences are paired by seed.

ModelParams (M)Weighted F4Direct − this model
Dense (original)6.5367.77 ± 0.54+2.30 ± 0.15
Dense + offset head6.5368.26 ± 0.57+1.81 ± 0.21
Dense, width 1607.0468.00 ± 0.27+2.06 ± 0.14
Direct (TomoDETR w/o CWQR)6.9870.07 ± 0.40–

The offset head recovers about 0.5 points of the gap and the extra capacity about 0.2 points, so Direct still leads both new variants by 1.8–2.1 points. These rows are now in Table 1(b), and Appendix A.3 describes both variants. The abstract and introduction now report the gap to the strongest Dense variant (1.81 points) instead of 2.30.

AuthorsOpenReview ↗

Response to Reviewer w5xS (2/3)

W2 + Q2. Are the external baselines documented with uncertainties, and do the error bars reflect comparable sources of variation?

Reviewer concern.

I'd appreciate more details about the external baselines. For example, the external CZII baselines are without uncertainties. Several details might also be worth documenting, such as the training and tuning, to ensure the reproducibility of the results and comparisons. Further, on SHREC2021, TomoDETR exceeds DeepETPicker in all three listed runs, but evaluation uses one held-out tomogram. DeePiCt reports variation across three fixed splits, rather than repeated training seeds on a fixed split. These evaluations are useful, but their error bars describe different sources of variation and should not be interpreted as equivalent evidence of robustness.

Can you provide fuller external-baseline training and selection configurations, and quote some uncertainties for the CZII baseline results, while distinguishing seed variation from split or tomogram variation?

Response.

We are grateful for these detailed suggestions, which make the comparison more transparent and reproducible. We address each point below.

Uncertainties. We retrained every external baseline with three seeds. On CZII, weighted F4 is:

  • DeepETPicker: 61.01 ± 1.26;
  • DeepFinder: 52.80 ± 1.62 (two seeds; the third run produced no predictions);
  • OCTOPI: 66.81 ± 0.55;
  • TomoDETR: 72.92 ± 0.45.

The gap to the strongest baseline (6.1 points) far exceeds the seed variation. Table 1, Table 5, and the text now report these three-seed means instead of the single runs in the submission.

Training and selection. Appendix D.4 (Table 21) now documents the training and selection configuration of each baseline on each dataset.

  • Each baseline uses its official code and the recipe from its paper or official example.
  • We tuned each baseline on validation data with the selection mechanisms its code exposes (checkpoint or epoch, thresholds, cluster sizes, and post-processing settings). It uses the same or more selection data than TomoDETR.
  • The submission stated that all baselines select on 121 public CZII tomograms. In fact only DeepFinder does; DeepETPicker and OCTOPI use the same 40 tomograms as TomoDETR. This was an error in the description only and is now corrected (Sec. 4.1, Appendix A.2).

Sources of variation. We agree. On CZII and SHREC2021, ± reflects training seeds on a fixed split; on DeePiCt, it reflects three fixed splits without seed repetition. Sec. 4.1 now states this explicitly and notes that these reflect different sources of variation.

SHREC2021 test tomograms. All methods follow the DeepETPicker protocol: they train on tomograms 0–7, select on 8, and test on 9 without retraining. The submission's mention of retraining on 0–8 was a description error and is now corrected (Sec. 4.1).

We added two further splits with different test tomograms, one seed each (Appendix D.3, Table 19). TomoDETR has the highest F1 on all three test tomograms, although absolute scores vary across tomograms for all methods:

Test tomogramTomoDETRDeepETPickerDeepFinderOCTOPI
9 (original, three seeds)93.0892.0679.5985.40
8 (Split A)93.2372.2384.8392.69
7 (Split B)65.0941.7250.3064.28

DeePiCt. To remove split difficulty from the comparison, we added a per-test-tomogram comparison with DeepETPicker (Appendix D.3, Table 20). Per test tomogram, the two methods are comparable; on DeePiCt we claim competitive rather than superior performance.

AuthorsOpenReview ↗

Response to Reviewer w5xS (3/3)

Q3. What explains 72.43 in Table 9 versus 72.92 in the main tables, and is this a reproducibility issue?

Reviewer concern.

I noticed that in Table 9, the separate RTX 5090 output-replay experiment reaches 72.43 rather than the main 72.92. Which checkpoint, cache, or selection differences explain this gap, and can the mechanism comparison be reconciled with the main endpoint? Does this indicate some reproducibility issue, or am I overly concerned?

Response.

Thank you for this careful check. It is not a reproducibility issue.

The Table 9 experiment used the main-table detector checkpoints but an earlier CWQR checkpoint from a previous selection round. Only Table 9 was affected; all other results use the final Val40-selected checkpoints.

We reran Table 9 with the main-table checkpoints:

  • The full model now gives 72.92 ± 0.45, matching the main tables.
  • The mechanism comparison is unchanged: updating coordinates alone gives +0.63, scores alone +0.34, and both +2.85.

The revised Table 9 is in Appendix A.6.

Reviewer_eMsuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper the authors proposed to adapt deformable DETR (ICLR 2021) to the application of particle picking in cryo-et. The proposed model has been tested on a few synthetic and real cryo-et datasets. Overall the work appears to be technically sound. The proposed model appears to outperform several competing models such as DeepETPicker. I also have a few concerns. First, the performance metrics should be expanded to include specifically precision, recall, etc, as for DeepETPicker. Second, there is no reconstruction results in the paper. Third, related studies such as cryoDETR should be cited. https://link.springer.com/chapter/10.1007/978-981-92-3719-7_8

Critical strengths

The proposed CWQR module seems to perform well in enhancing the performance of particle picking. Testing the method on CZII datasets provides useful benchmark results.

Critical weaknesses
  1. Reported performance metrics are limited and should be expanded to include specifically precision, recall, etc, as for DeepETPicker.

  2. There are no reconstruction results in the paper, as acknowledged by the authors.

  3. Related studies such as cryoDETR should be cited. https://link.springer.com/chapter/10.1007/978-981-92-3719-7_8

Questions for authors
  1. The proposed method has been compared to three other methods, DeepETPicker, DeepFinder, OCTOPI, can the authors make a comparison with cryoDETR?
  2. Figure 5 shows the number of false positive (FP), false negatives (FN). But the results are very limited,can the authors provide a more comprehensive (i.e. dataset level) comparison of these numbers, so that comparison can be made against for example DeepETPicker?
3 comments
AuthorsOpenReview ↗

Response to Reviewer eMsu (1/3)

W1 & Q2. Can the evaluation report precision, recall, and dataset-level TP/FP/FN, as DeepETPicker does?

Reviewer concern.

Reported performance metrics are limited and should be expanded to include specifically precision, recall, etc, as for DeepETPicker.

Figure 5 shows the number of false positive (FP), false negatives (FN). But the results are very limited,can the authors provide a more comprehensive (i.e. dataset level) comparison of these numbers, so that comparison can be made against for example DeepETPicker?

Response.

We thank the reviewer for this helpful suggestion; dataset-level statistics make the comparison with DeepETPicker much more direct. Following DeepETPicker [1], we added precision, recall, F1, TP/FP/FN, and average center distance (AD) for all methods on all three benchmarks. The CZII results are summarized below (364 test tomograms, mean over three training seeds). The complete statistics are in Appendix D.1 of the revised paper (Table 15 for CZII, Table 16 for SHREC2021 and DeePiCt).

MethodWeighted F4PrecisionRecallTP / FP / FNAD (Å)
DeepETPicker61.0160.574.932,722 / 21,345 / 10,95416.2
OCTOPI66.8171.769.930,551 / 12,073 / 13,12516.7
TomoDETR w/o CWQR70.0734.482.736,114 / 68,929 / 7,56222.5
TomoDETR72.9245.486.337,703 / 45,320 / 5,97314.6

TomoDETR attains the highest recall and the lowest AD, with lower precision. This is the trade-off favored by the official CZII metric, the recall-weighted F4: one missed particle costs as much as 16 false positives, and all methods select their operating points by this metric. On SHREC2021, where thresholds are selected for F1, TomoDETR's precision and recall are balanced (94.6 / 91.7).

[1] G. Liu, T. Niu, M. Qiu, Y. Zhu, F. Sun, and G. Yang. DeepETPicker: Fast and accurate 3D particle picking for cryo-electron tomography using weakly supervised deep learning. Nature Communications, 15:2090, 2024. https://doi.org/10.1038/s41467-024-46041-0

AuthorsOpenReview ↗

Response to Reviewer eMsu (2/3)

W2. Can reconstruction (subtomogram averaging) results be provided?

Reviewer concern.

There are no reconstruction results in the paper, as acknowledged by the authors.

Response.

We appreciate this important point; reconstruction is the downstream purpose of particle picking. We added subtomogram averaging (RELION-5, CZII challenge pipeline) on 40 CZII test tomograms, using identical settings for TomoDETR and DeepETPicker. Map–model FSC and visualization use maps post-processed in RELION without sharpening (B = 0), ensuring consistent post-processing across methods. The full results and settings are in Appendix D.2 of the revised paper (Tables 17–18).

  • Equal numbers of picks. The two methods give comparable maps. TomoDETR is slightly better on ribosome (half-map FSC 0.143: 10.73 vs. 10.98 Å), DeepETPicker is better on apo-ferritin, and thyroglobulin is similar.
  • Picks that match no annotation. TomoDETR has about 1.8× as many unmatched ribosome picks as DeepETPicker. These picks still reconstruct the ribosome at equal or slightly better map–model resolution (15.69–15.79 vs. 15.71–15.91 Å), and the result holds on 37 additional tomograms (15.58 ± 0.04 vs. 15.72 ± 0.02 Å). Random positions do not reach FSC 0.5. Some of TomoDETR's "false positives" are therefore real particles absent from the annotations.

Figure 8 (Appendix D.2) shows these picks, their reconstructions, and the FSC curves, with random positions as a control. Given bin-3 data, we claim comparable rather than better reconstructions; the Limitations (Sec. 5) now state this scope.

AuthorsOpenReview ↗

Response to Reviewer eMsu (3/3)

W3 & Q1. Should CryoDETR be cited and compared?

Reviewer concern.

Related studies such as cryoDETR should be cited. https://link.springer.com/chapter/10.1007/978-981-92-3719-7_8

The proposed method has been compared to three other methods, DeepETPicker, DeepFinder, OCTOPI, can the authors make a comparison with cryoDETR?

Response.

Thank you for pointing us to CryoDETR, a closely related DETR-based particle picker. We now cite it in the related work (Sec. 2).

CryoDETR is designed for 2D single-particle cryo-EM micrographs, whereas TomoDETR localizes particles in reconstructed 3D cryo-ET tomograms. Applying CryoDETR to tomograms would require substantial changes to its encoder, queries, and window handling, so a faithful comparison is not straightforward. We therefore compare TomoDETR with established 3D cryo-ET pickers (DeepETPicker, DeepFinder, and OCTOPI).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.