Pith. sign in

REVIEW 4 major objections 5 minor 31 references

SaSi: A Self-augmented and Self-interpreted Deep Learning Approach for Few-shot Cryo-ET Particle Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes SaSi, a few-shot cryo-ET particle detection method that densifies training sub-volumes with transformed particles and enforces transform-consistent segmentations, reporting F1 improvements over DeepFinder and…

desk verdict Plausible first few-shot cryo-ET particle picker with a fair baseline upgrade, but single-run F1s without error bars leave the headline margin unproven. read the letter →

arxiv 2505.19948 v1 pith:G4W3E67B submitted 2025-05-26 cs.CV cs.AIcs.LG

Gokul Adethya , Bhanu Pratyush Mantha , Tianyang Wang , Xingjian Li , Min Xu This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords cryo-electrontomographyparticledetectionfew-shotlearningself-supervisedvolumeinfillconsistencyguidance3DsegmentationAugMix
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

The paper tries to establish that a deep segmentation model can localize and classify particles in 3D cryo-electron tomograms from as few as three to ten labeled examples per class, a regime where existing state-of-the-art pickers degrade sharply. It proposes SaSi, a U-Net-based method with two novel components: a self-augmented Volume Infill that densifies sub-volumes with transformed copies of the few labeled particles, and a self-interpreted Consistency Guidance loss that forces the predicted segmentation map to transform in step with the input. On the simulated SHREC2021 benchmark, SaSi reports F1 of 0.409 with ten labels per class, versus 0.276 for DeepETPicker and 0.260 for DeepFinder under the same connected-component post-processing, with similar gains on a real three-tomogram dataset. The claim matters because cryo-ET ground truth is expensive and sparse, so a picker that works with tens of annotations rather than thousands would make tomogram analysis practical in new biological contexts.

What carries the argument

The load-bearing machinery is a pair of self-generated training signals layered on a 3D U-Net with spherical pseudo-masks built from point annotations. Volume Infill adapts AugMix to volumetric data: several random chains of spatial-only operations (shift, rotation, flip) are mixed via a Dirichlet draw and combined with the original sub-volume, and the same mixing is applied to the label mask, increasing particle density and orientation diversity without introducing intensity corruption. Consistency Guidance computes the segmentation on an input and on a spatially transformed copy, then penalizes disagreement between the transformed prediction and the prediction on the transformed input, thereby enforcing transformation-aware equivariance on both encoder and decoder. The pipeline closes with 3D connected-component labeling (cc3d), which the paper finds is more stable and hyperparameter-free than the mean-shift or MP-NMS post-processing used by the baselines.

What would settle it

Repeat the SHREC2021 N=10 comparison with, say, ten random draws of the labeled particles per class, training SaSi and DeepETPicker with cc3d from the same seeds, and compare the F1 distributions; the claim of significant improvement fails if the interquartile ranges overlap substantially.

Watch

Extended reading notes

Core claim

The central discovery is that the difficulty of few-shot particle picking in cryo-ET is less about the segmentation backbone and more about particle sparsity and label scarcity, and that both can be attacked with self-generated supervisory signals. SaSi increases the effective particle density inside each training sub-volume by mixing a chain of spatial transformations of the input (shifts, rotations, flips) with the original volume, carrying the same mixing through to the point-derived spherical masks, so the network sees more particles per sample. At the same time, the self-interpreted Consistency Guidance term applies the same spatial transformation to the predicted segmentation map and to the network's own output on the transformed input, training the decoder to respect the geometry of the structures without needing extra labels. On SHREC2021, this combination is reported to outperform both DeepFinder and DeepETPicker at N=3, 5, and 10, and to transfer to real tomograms, where SaSi improves F1 from 0.132 to 0.136 on the test tomogram and from 0.123 to 0.163 on validation-plus-test at N=10.

Load-bearing premise

The entire evaluation rests on single F1 runs per setting, so the claimed gains assume that the random choice of the few labeled particles (and training randomness) does not produce differences as large as the reported margins.

Editorial extensions

If this is right

  • If the reported gains hold, cryo-ET practitioners can obtain usable particle localizations from roughly 10 annotated particles per class instead of thousands.
  • The components transfer across architectures: applying Volume Infill to DeepETPicker raises its N=10 F1 from 0.276 to 0.351, so the augmentation strategy is not tied to the U-Net used for SaSi.
  • Self-interpreted Consistency Guidance provides a label-free regularization that should help whenever segmentation outputs are expected to follow known spatial transforms, not just in cryo-ET.
  • Connected-component post-processing eliminates manually tuned parameters from the picking pipeline, which should improve reproducibility of few-shot benchmarks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next experiment is a multi-seed benchmark: with only one run per setting, the paper's margins over the baselines (e.g., 0.409 vs 0.276 at N=10) have not yet been separated from run-to-run variation caused by which few particles are chosen.
  • Because the method assumes the particle radius is known to build spherical masks, an extension to estimate per-particle radii or to use anisotropic shapes could broaden it to heterogeneous macromolecular aggregates.
  • Consistency Guidance is essentially an equivariance prior; the same self-interpreted loss could be dropped into other sparse-label 3D segmentation tasks, such as vesicle or neuron tracing in volume electron microscopy, where label scarcity is the same bottleneck.
  • The Shapley analysis attributes more value to Volume Infill than to Consistency Guidance; a testable prediction is that even simpler density-boosting tricks, such as tiling particles into larger crops, would capture part of the same gain.
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

4 major / 5 minor

Summary. The manuscript proposes SaSi, a few-shot cryo-ET particle detection method built on a 3D U-Net/DeepFinder architecture. It adds three ingredients: self-supervised contrastive pre-training, a self-augmented Volume Infill strategy adapted from AugMix to increase particle density in subvolumes, and a self-interpreted Consistency Guidance loss that regularizes predicted segmentation maps under spatial transformations. The method uses spherical masks generated from weak point labels and replaces the original post-processing of DeepFinder and DeepETPicker with cc3d connected components. Experiments on the SHREC2021 benchmark and a real cryo-ET dataset report F1 scores for N=3, N=5, and N=10 labeled particles per class, with SaSi outperforming DeepFinder and DeepETPicker baselines in most settings. An ablation using Shapley values attributes positive contributions to both Volume Infill and Consistency Guidance.

Significance. If the reported gains are replicable, the paper would make a useful contribution to cryo-ET particle picking in label-scarce settings, where existing pipelines are designed for larger annotation budgets. The authors are to be credited for evaluating improved cc3d-based baseline variants rather than only the original post-processing methods, for testing on both simulated and real tomograms, and for explicitly targeting the few-shot regime. The proposed components are simple and architecture-agnostic in principle, and the use of spherical weak labels is well aligned with real annotation practice. However, the headline claim of "significant" improvement is currently supported only by single F1 measurements per configuration, and the evaluation design also averages over overlapping test sets, so the quantitative basis for the central claim is not yet established. The ablation section additionally contains an internal inconsistency in the interpretation of the Volume Infill Shapley values.

major comments (4)
  1. [Section 5.1, Tables 1 and 2] The central claim that SaSi significantly outperforms existing methods is supported only by a single F1 value per condition, obtained from one randomly selected set of labeled particles and one training run. The manuscript itself states in Section 5.1 that baselines are unstable, "sometimes getting lower F1 score on the N=5 setting than that of N=3 due to high randomness of the few-shot learning problem." Without standard deviations, confidence intervals, or repeated label-set and seed trials, the observed margins cannot be distinguished from run-to-run noise. This is especially true on the real dataset, where the N=10 Test margin is 0.136 for SaSi versus 0.132 for DeepFinder (cc3d). Because Tables 1 and 2 are the sole evidence for the paper's main claim, the authors should report mean and standard deviation over multiple random label subsets and training seeds, and ideally provide per-condition significance tests.
  2. [Section 4.1.1 and Table 1] The SHREC evaluation averages F1 over T9, T8-9, and T6-9, but these test sets are nested and heavily overlapping: T9 is contained in T8-9, and T8-9 is contained in T6-9. Averaging the three numbers does not provide three independent evaluations, and it implicitly overweights the T9 tomogram. This weakens the claim that the evaluation is "more reliable" than using the official T9 alone. Please report per-tomogram F1 scores and either use disjoint test sets or make the official T9 split the primary basis for aggregate conclusions.
  3. [Section 5.2, Tables 3 and 4] The interpretation of the Volume Infill Shapley analysis is internally inconsistent. Table 3 shows phi_VI increasing with N (0.052 at N=3, 0.097 at N=5, 0.130 at N=10), meaning Volume Infill contributes more as the number of labeled particles grows, yet the text states that "VI shows more improvements under fewer-shot settings." The subsequent explanation then describes exactly the opposite behavior, namely that more initial examples provide more diversity for Volume Infill to exploit. The sentence appears to reverse the direction of the finding and should be corrected. In addition, these Shapley values are computed from single runs per configuration, so the same lack of repeated-trial variance applies to the ablation conclusion as to the main results.
  4. [Section 4.3 and Section 5.2] Several hyperparameters are selected using the test benchmark itself: lambda_dice = 20 and lambda_focal = 1 are said to be set "according to empirical observations," the consistency guidance loss is applied before epoch 4000 "as it achieved better empirical performance," and the self-supervised pre-training duration is fixed at 10 epochs. Since the central comparison is against baselines on the same test sets, tuning on the test benchmark creates a selection-bias risk that should at least be disclosed explicitly as a limitation. Ideally, hyperparameters should be fixed on a validation split or the sensitivity of the main conclusions to these choices should be reported.
minor comments (5)
  1. [Section 3.4 and Section 3.5] The symbol T is used both for the set of augmentation operations in Volume Infill and for a single spatial transformation in Consistency Guidance; please use distinct notation to avoid confusion.
  2. [Section 2] There is a typo in "DoG applies a band passband-pass filter"; the phrase should be cleaned up.
  3. [Table 1] The baseline label "DeepETpicker (nmsv2)" is not introduced in the text, which only mentions MP-NMS; please align the notation between Section 4.2 and Table 1.
  4. [Section 4.3] The statement that all SaSi experiments use the DeepFinder architecture is clear, but the claim in the introduction that the components are compatible with ViT and other architectures is not experimentally demonstrated in this paper; please mark it as a claim about compatibility rather than an empirical result.
  5. [Section 4.1.2 and Table 2] For the real dataset, the "Val+Test" evaluation pools the validation and test tomograms, but the validation tomogram is used as the split described in the text; a brief justification of why pooling is appropriate would help the reader interpret the numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is tested on external benchmarks with held-out tomograms, and the proposed components are training-time regularizers rather than constructions that encode the answer.

full rationale

The paper's claimed derivation chain is self-contained. The supervised loss (Eq. 1) is a standard focal and Dice segmentation objective on pseudo-labels generated from point annotations, and the Volume Infill is an AugMix-style augmentation whose labels are obtained by transforming the same point-derived masks, not by reading off the test answer. The Consistency Guidance loss L(m', T(m)) is a transformation-consistency regularizer that uses the model's own predictions as targets, which is a training objective rather than a circular definition of the final detection. Evaluation is against external benchmarks (SHREC2021 and real tomograms) with held-out test tomograms and no test-set fitting. The few self-citations (e.g., refs. 4, 5, 18, 23, 25) appear in related-work context and are not load-bearing for the central improvement claim. The Shapley-value ablations are computed by retraining the system with component subsets, so they measure empirical marginal contributions rather than being defined by the target result. The acknowledged randomness in few-shot selection (Section 5.1) and the absence of repeated-trial error bars are statistical-reproducibility concerns, not circularity, and therefore do not raise the circularity score.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper does not postulate new physical entities or forces. It relies on standard network architectures, well-known losses, and a set of hand-selected hyperparameters. The main load-bearing assumptions are domain-specific: spatial transformations do not distort particle patterns, spherical pseudo-masks are adequate supervision, and the training and test tomograms are distributionally similar. These assumptions are stated in Sections 3.2, 3.4, 3.6, and 4.1.1.

free parameters (6)
  • lambda_dice = 20
    Weight of dice loss in combined loss (Eq. 1), set by empirical observations in Section 3.2.
  • lambda_focal = 1
    Weight of focal loss in combined loss (Eq. 1), set empirically.
  • self-supervised pretraining epochs = 10
    Number of epochs for contrastive pretraining before switching to supervised learning (Section 4.3).
  • consistency guidance stop epoch = 4000
    Self-interpreted loss is applied only before the 4000th epoch for better empirical performance (Section 4.3).
  • NT-Xent temperature = 0.1
    Temperature parameter for the contrastive loss (Section 4.3).
  • AugMix parameters = alpha=1; k, m, beta unspecified
    Dirichlet alpha set to 1; chain length, number of chains, and Beta mixing coefficient are not reported (Sections 3.4 and 4.3).
assumptions (5)
  • standard math U-Net with focal and dice losses is an appropriate objective for particle segmentation
    The method builds on the standard practice in DeepFinder and DeepETPicker (Section 3.2).
  • domain assumption Spatial transformations (shift, rotation, flip) preserve particle structure
    Section 3.2 states that any augmentation that affects voxel values could distort the underlying pattern; the method limits augments to spatial operations, and Volume Infill and Consistency Guidance rely on this.
  • domain assumption Spheres generated around point labels are valid pseudo-ground-truth masks
    Section 3.2 converts weak point labels into spherical masks of minimum radius; this assumes the radius is known and that spheres are a faithful segmentation target.
  • domain assumption cc3d connected components with 26-connectivity produce particle centers matching point annotations
    Section 3.6 uses cc3d as post-processing for all methods; the centroid of each cluster is taken as a detected particle position.
  • domain assumption The training tomogram T0 and test tomograms T6-T9 are from the same distribution
    Section 4.1.1 uses only T0 for training and evaluates on T9, T8-9, T6-9; this assumes the few-shot-trained model generalizes across tomograms without domain shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SaSi: A Self-augmented and Self-interpreted Deep Learning Approach for Few-shot Cryo-ET Particle Detection." pith.science (2026). https://pith.science/paper/G4W3E67B

@misc{pith2026250519948,
  author       = {Pith},
  title        = {Pith review of: SaSi: A Self-augmented and Self-interpreted Deep Learning Approach for Few-shot Cryo-ET Particle Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4W3E67B}},
  note         = {Machine review of arXiv:2505.19948}
}
read the original abstract

Cryo-electron tomography (cryo-ET) has emerged as a powerful technique for imaging macromolecular complexes in their near-native states. However, the localization of 3D particles in cellular environments still presents a significant challenge due to low signal-to-noise ratios and missing wedge artifacts. Deep learning approaches have shown great potential, but they need huge amounts of data, which can be a challenge in cryo-ET scenarios where labeled data is often scarce. In this paper, we propose a novel Self-augmented and Self-interpreted (SaSi) deep learning approach towards few-shot particle detection in 3D cryo-ET images. Our method builds upon self-augmentation techniques to further boost data utilization and introduces a self-interpreted segmentation strategy for alleviating dependency on labeled data, hence improving generalization and robustness. As demonstrated by experiments conducted on both simulated and real-world cryo-ET datasets, the SaSi approach significantly outperforms existing state-of-the-art methods for particle localization. This research increases understanding of how to detect particles with very few labels in cryo-ET and thus sets a new benchmark for few-shot learning in structural biology.

Figures

Figures reproduced from arXiv: 2505.19948 by the authors.

Figure 1
Figure 1. This figure illustrates the training phase incorporating self-supervised learning using augmented pair [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. During testing and self-supervised learning, a sliding window approach with window size W and stride W/2 is applied, while [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the composition process of our self-augmented volume infill strategy. The input volume is filled with more [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Cellular struc- tural biology as revealed by cryo-electron tomogra- phy

    Irobalieva RN, Martins B, Medalia O. Cellular struc- tural biology as revealed by cryo-electron tomogra- phy. Journal of cell science. 2016;129(3):469–476. 1

    work page 2016

  2. [2]

    Cryo-electron tomography: gaining insight into cellular processes by structural approaches

    Yahav T, Maimon T, Grossman E, Dahan I, Medalia O. Cryo-electron tomography: gaining insight into cellular processes by structural approaches. Current opinion in structural biology. 2011;21(5):670–677. 1

    work page 2011

  3. [3]

    The architecture of inactivated SARS-CoV-2 with postfusion spikes revealed by cryo-EM and cryo- ET

    Liu C, Mendonc ¸a L, Yang Y , Gao Y , Shen C, Liu J, et al. The architecture of inactivated SARS-CoV-2 with postfusion spikes revealed by cryo-EM and cryo- ET. Structure. 2020;28(11):1218–1224. 1

    work page 2020

  4. [4]

    Gum-net: Unsupervised geometric matching for fast and accurate 3d subtomogram im- age alignment and averaging

    Zeng X, Xu M. Gum-net: Unsupervised geometric matching for fast and accurate 3d subtomogram im- age alignment and averaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition; 2020. p. 4073–4084. 1

    work page 2020

  5. [5]

    De novo structural pattern min- ing in cellular electron cryotomograms

    Xu M, Singla J, Tocheva EI, Chang YW, Stevens RC, Jensen GJ, et al. De novo structural pattern min- ing in cellular electron cryotomograms. Structure. 2019;27(4):679–691. 1

    work page 2019

  6. [6]

    Cryo-electron tomography of cellular mi- crotubules

    Koning RI. Cryo-electron tomography of cellular mi- crotubules. Methods in cell biology. 2010;97:455–

    work page 2010

  7. [7]

    Structural biology in situ—the potential of subtomogram averaging

    Briggs JA. Structural biology in situ—the potential of subtomogram averaging. Current opinion in structural biology. 2013;23(2):261–267. 1

    work page 2013

  8. [8]

    Deep learn- ing improves macromolecule identification in 3D cel- lular cryo-electron tomograms

    Moebel E, Martinez-Sanchez A, Lamm L, Righetto RD, Wietrzynski W, Albert S, et al. Deep learn- ing improves macromolecule identification in 3D cel- lular cryo-electron tomograms. Nature methods. 2021;18(11):1386–1394. 1, 2, 3, 5

    work page 2021

Show all 31 references
  1. [9]

    SHREC 2020: Classi- fication in cryo-electron tomograms

    Gubins I, Chaillet ML, van Der Schot G, Veltkamp RC, F ¨orster F, Hao Y , et al. SHREC 2020: Classi- fication in cryo-electron tomograms. Computers & Graphics. 2020;91:279–289. 1

    2020

  2. [10]

    DeepET- Picker: Fast and accurate 3D particle picking for cryo- electron tomography using weakly supervised deep learning

    Liu G, Niu T, Qiu M, Zhu Y , Sun F, Yang G. DeepET- Picker: Fast and accurate 3D particle picking for cryo- electron tomography using weakly supervised deep learning. Nature Communications. 2024;15(1):2090. 1, 3, 5

    2024

  3. [11]

    Augmix: A simple data pro- cessing method to improve robustness and uncertainty

    Hendrycks D, Mu N, Cubuk ED, Zoph B, Gilmer J, Lakshminarayanan B. Augmix: A simple data pro- cessing method to improve robustness and uncertainty. arXiv preprint arXiv:191202781. 2019;. 2, 4

    2019

  4. [12]

    The surprising effectiveness of representation learning for visual imitation

    Pari J, Shafiullah NM, Arunachalam SP, Pinto L. The surprising effectiveness of representation learning for visual imitation. arXiv preprint arXiv:211201511. 2021;. 2

    2021

  5. [13]

    Unsupervised visual representation learning by context prediction

    Doersch C, Gupta A, Efros AA. Unsupervised visual representation learning by context prediction. In: Pro- ceedings of the IEEE international conference on com- puter vision; 2015. p. 1422–1430. 2

    2015

  6. [14]

    Boosting self-supervised learning via knowledge transfer

    Noroozi M, Vinjimoor A, Favaro P, Pirsiavash H. Boosting self-supervised learning via knowledge transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2018. p. 9359–9367. 2

    2018

  7. [15]

    Identification of macromolecu- lar complexes in cryoelectron tomograms of phantom cells

    Frangakis AS, B ¨ohm J, F ¨orster F, Nickell S, Nicas- tro D, Typke D, et al. Identification of macromolecu- lar complexes in cryoelectron tomograms of phantom cells. Proceedings of the National Academy of Sci- ences. 2002;99(22):14153–14158. 2 8

    2002

  8. [16]

    DoG Picker and TiltPicker: software tools to facilitate particle selection in single particle electron microscopy

    V oss N, Yoshioka C, Radermacher M, Potter C, Car- ragher B. DoG Picker and TiltPicker: software tools to facilitate particle selection in single particle electron microscopy. Journal of structural biology. 2009;166(2):205–213. 2

    2009

  9. [17]

    Detection and identification of macromolecu- lar complexes in cryo-electron tomograms using sup- port vector machines

    Chen Y , Hrabe T, Pfeffer S, Pauly O, Mateus D, Navab N, et al. Detection and identification of macromolecu- lar complexes in cryo-electron tomograms using sup- port vector machines. In: 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI). IEEE

    2012

  10. [18]

    Improved deep learning-based macromolecules structure classification from electron cryo-tomograms

    Che C, Lin R, Zeng X, Elmaaroufi K, Galeotti J, Xu M. Improved deep learning-based macromolecules structure classification from electron cryo-tomograms. Machine vision and applications. 2018;29:1227–1236. 2

    2018

  11. [19]

    Very deep convolutional networks for large-scale image recognition

    Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556. 2014;. 2

    2014

  12. [20]

    Deep residual learning for image recognition

    He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recogni- tion; 2016. p. 770–778. 2

    2016

  13. [21]

    SuRV oS: super-region volume seg- mentation workbench

    Luengo I, Darrow MC, Spink MC, Sun Y , Dai W, He CY , et al. SuRV oS: super-region volume seg- mentation workbench. Journal of Structural Biology. 2017;198(1):43–53. 2

    2017

  14. [22]

    Convolutional neural networks for auto- mated annotation of cellular cryo-electron tomograms

    Chen M, Dai W, Sun SY , Jonasch D, He CY , Schmid MF, et al. Convolutional neural networks for auto- mated annotation of cellular cryo-electron tomograms. Nature methods. 2017;14(10):983–985. 2

    2017

  15. [23]

    Automatic localization and identification of mi- tochondria in cellular electron cryo-tomography using faster-RCNN

    Li R, Zeng X, Sigmund SE, Lin R, Zhou B, Liu C, et al. Automatic localization and identification of mi- tochondria in cellular electron cryo-tomography using faster-RCNN. BMC bioinformatics. 2019;20:75–85. 2

    2019

  16. [24]

    Faster R-CNN: To- wards real-time object detection with region proposal networks

    Ren S, He K, Girshick R, Sun J. Faster R-CNN: To- wards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence. 2016;39(6):1137–1149. 2

    2016

  17. [25]

    Few-shot learning for classification of novel macro- molecular structures in cryo-electron tomograms

    Li R, Yu L, Zhou B, Zeng X, Wang Z, Yang X, et al. Few-shot learning for classification of novel macro- molecular structures in cryo-electron tomograms. PLoS computational biology. 2020;16(11):e1008227. 3

    2020

  18. [26]

    Un- derstanding and improving the role of projection head in self-supervised learning

    Gupta K, Ajanthan T, Hengel Avd, Gould S. Un- derstanding and improving the role of projection head in self-supervised learning. arXiv preprint arXiv:221211491. 2022;. 4

    2022

  19. [27]

    cc3d: Connected components on mul- tilabel 3D & 2D images.; 2021

    Silversmith W. cc3d: Connected components on mul- tilabel 3D & 2D images.; 2021. 5

    2021

  20. [28]

    In situ structure of neuronal C9orf72 poly-GA aggregates reveals proteasome re- cruitment

    Guo Q, Lehmer C, Mart ´ınez-S´anchez A, Rudack T, Beck F, Hartmann H, et al. In situ structure of neuronal C9orf72 poly-GA aggregates reveals proteasome re- cruitment. Cell. 2018;172(4):696–705. 6

    2018

  21. [29]

    SHREC 2021: Classification in Cryo-electron Tomograms

    Gubins I, Chaillet ML, White T, Bunyak F, Papoulias G, Gerolymatos S, et al. SHREC 2021: Classification in Cryo-electron Tomograms. In: Biasotti S, Dyke RM, Lai Y , Rosin PL, Veltkamp RC, editors. Euro- graphics Workshop on 3D Object Retrieval. The Eu- rographics Association; ...

    2021

  22. [30]

    A value for n-person games

    Shapley LS. A value for n-person games. Contribution to the Theory of Games. 1953;2. 7

    1953

  23. [31]

    A unified approach to inter- preting model predictions

    Scott M, Su-In L, et al. A unified approach to inter- preting model predictions. Advances in neural infor- mation processing systems. 2017;30:4765–4774. 7 9

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.