Pith. sign in

REVIEW 5 major objections 7 minor 51 references

Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that self-supervised pretraining for 3D medical images improves when anatomical structures are treated as semantic units, enforced by two losses that push different-structure patches apart and pull same-structure patches…

desk verdict Useful new SSL idea for 3D medical images, but the SOTA claim is not yet verifiable because baselines are borrowed and the headline metric is undefined. read the letter →

arxiv 2507.02581 v1 pith:7CS4OOQ6 submitted 2025-07-03 cs.CV

Tan Pan , Zhaorui Tan , Kaiyu Guo , Dongli Xu , Weidi Xu , Chen Jiang , Xin Guo , Yuan Qi
show 1 more author
Yuan Cheng
This is my paper
classification cs.CV
keywords self-supervisedlearning3Dmedicalimagingstructure-awarerepresentationsemanticconsistencydiscrepancyoptimaltransportcontrastiveSwinTransformer
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 argues that self-supervised pretraining for 3D medical images should respect anatomical structure: patches inside the same organ should be pulled together into one semantic cluster, while patches from different organs should be pushed apart. To do this it proposes S2DC, which first builds reliable patch-to-patch correspondences between two augmented views using the known affine transform and pushes all patches apart, then reweights that push by a Sharpe-ratio measure of how concentrated each patch's similarity distribution is, so patches that have several similar neighbors (likely the same structure) are less aggressively separated. The paper reports that this two-step objective, added to a standard contrastive loss, outperforms prior medical SSL methods on segmentation, classification, image-to-image translation, and reconstruction across CT, MRI, and PET, with an average score of 81.43% versus 80.65% for the previous best method and 77.93% for training from scratch.

What carries the argument

The central objects are (i) the ground-truth patch correspondence matrix $M_{gt}$ computed from the affine transform $H$ between two views, which records which patch in the augmented view corresponds to which patch in the source view; (ii) the dual-softmax operator applied to the feature similarity map, which refines soft correspondences and brings mutual nearest neighbors closer; (iii) the neighborhood similarity distribution $D_n$ of each anchor patch, whose dispersion is summarized by a simplified Sharpe ratio $sr_{V_i}^n = (\max(D_n) - \bar{D}_n)/\sigma_{D_n}$; and (iv) the soft weighting of the patch-to-patch loss by these Sharpe ratios, so that patches with several similar neighbors (likely same structure) are treated less harshly. The total loss is $L = L_g + L_{p2p} + L_{p2s}$, where $L_g$ is the InfoNCE contrastive loss and the two added terms implement discrepancy and consistency respectively.

What would settle it

Take a trained S2DC model, apply a range of affine warps with increasing rotation and scale to volumes, and compute the fraction of patches whose most similar feature in the augmented view matches the ground-truth correspondence $M_{gt}$; if this matching accuracy drops sharply toward chance for moderate deformations, the equivariance premise underlying the patch-to-patch supervision fails. Alternatively, an ablation that replaces the Sharpe-ratio weighting with a constant and shows no downstream drop would falsify the claim that the consistency step drives the gains.

Watch

Extended reading notes

Core claim

The central claim is that anatomical structure, not individual voxels or fixed patches, is the right semantic unit for 3D medical image self-supervision. Concretely, S2DC assumes that patches sharing a structure share semantics (intra-structure consistency) and patches from different structures have distinct semantics (inter-structure discrepancy). The framework enforces both: patch-to-patch discrepancy via a dual-softmax correspondence loss supervised by ground-truth correspondences computed from the affine warp between two views, and patch-to-structure consistency via a neighborhood similarity distribution whose dispersion is measured by a simplified Sharpe ratio and used as a soft weight to modulate the patch-to-patch loss. The paper's experiments claim consistent improvements over the state of the art, including an average score of 81.43% across 10 datasets versus 80.65% for the previous best method VoCo and 77.93% for training from scratch.

Load-bearing premise

The paper assumes that after a geometric augmentation, each patch token in the Swin Transformer moves exactly as the corresponding image patch does, so that the pixel-space affine correspondence computed from $H$ is also a semantically correct correspondence in feature space; if learned positional encodings and strided downsampling break that equivariance, the patch-to-patch loss supervises incorrect pairs.

Editorial extensions

If this is right

  • Pretrained S2DC weights transfer to organ segmentation (BTCV, MSD-Liver/Lung/Spleen, BraTs21), classification (CC-CCII, ADNI), image-to-image translation (BraTs23), and low-dose PET reconstruction (UDPET), improving over both general and medical SSL baselines.
  • The structure-aware principle transfers across CT, MRI, and PET modalities, suggesting it captures a general property of anatomical and functional imaging rather than a single-modality artefact.
  • The two losses are complementary: $L_{p2p}$ alone sparsifies inter-structure features, $L_{p2s}$ alone compacts intra-structure clusters, and the combination yields the best silhouette scores and Dice scores.
  • On CC-CCII, S2DC pretraining is most beneficial with little labeled data, reaching 88.29% accuracy with 10% of the training set versus 87.87% for the prior best method.
  • Larger structures are learned earlier in training than smaller ones, so the framework's difficulty with small structures (e.g., bronchioles) remains a known limitation.

Reading between the lines

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

  • The Sharpe-ratio weighting is a form of confidence or uncertainty gating: it down-weights the discrepancy loss exactly for patches whose feature distribution is ambiguous, so the method could be read as a principled alternative to hard clustering or group-contrast methods.
  • The equivariance assumption could be tested directly by measuring how the correspondence accuracy between $\hat{M}_t$ and $M_{gt}$ degrades as the affine deformation $H$ becomes more extreme; a sharp degradation would show where the method's supervision breaks.
  • The same structure-aware principle could be lifted to object-centric SSL in natural images where objects play the role of anatomical structures, or to 2D medical slices, which is not tested in this paper.
  • Since the paper only reports pre-training on 1k-10k volumes, scaling to larger corpora (e.g., whole-body CT) is an untested but plausible way to amplify the reported gains.
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

5 major / 7 minor

Summary. The paper proposes S2DC, a self-supervised learning framework for 3D medical images that combines patch-level semantic discrepancy and structure-level semantic consistency. Patch-level discrepancy is enforced through a patch-to-patch correspondence loss with ground-truth correspondences derived from affine transformations and a dual-softmax assignment, while structure-level consistency is imposed by weighting the correspondence loss with a Sharpe-ratio-based soft label computed from the neighborhood similarity distribution. The authors evaluate S2DC on 10 datasets, 4 tasks, and 3 modalities using a Swin-B backbone, reporting consistent improvements over prior medical image SSL methods, with a headline average score of 81.43% versus 80.65% for VoCo and 77.93% for training from scratch.

Significance. If the empirical claims are substantiated, S2DC offers a conceptually clear and well-motivated approach to incorporating anatomical structure awareness into 3D medical image self-supervised learning. The idea of using affine-geometry correspondences for patch-level discrepancy and distribution-derived soft labels for structure-level consistency is original and does not require extra annotations. The evaluation breadth across 10 datasets, 4 tasks, and 3 modalities is a strength, as is the promise of code release. However, the current evidence does not establish the SOTA claim: several baseline numbers are borrowed from other papers rather than produced under matched protocols, the aggregate metric used for the headline comparison is undefined, and a missing appendix with key implementation details prevents reproducibility. These issues are fixable and do not invalidate the underlying approach, but they must be addressed before the central claim can be accepted.

major comments (5)
  1. [Section 4.2 (Overall comparisons on 10 downstream datasets)] The aggregate 'average score' is not defined. Values from Dice, accuracy, PSNR, SSIM, and NMSE are combined into a single percentage without stating any normalization or aggregation rule. Since these metrics have different scales and directions (NMSE is lower-better), the reported 81.43% vs 80.65% and the 3.5% improvement are not interpretable. Please specify the exact aggregation formula, or report per-task averages separately and avoid the combined average.
  2. [Section 4.1 and Tables 1, 3] Baseline numbers are not produced under matched conditions. Table 1 states that 'Most of the results are drawn from [40, 44]', and Table 3 notes that the VoCo CT result uses the official released weight rather than being retrained under S2DC's pretraining and fine-tuning protocol. Because downstream fine-tuning details strongly affect Dice and accuracy, the claimed margin over VoCo (0.78%) may be an artifact of protocol differences. Please retrain all baselines under identical pretraining data, augmentations, schedule, and fine-tuning settings, or clearly indicate which numbers are borrowed and provide a justification for their comparability.
  3. [Section 3.3, Eq. (6)] The simplified Sharpe ratio is undefined when the variance of the neighborhood similarity distribution D_n is zero (e.g., a uniform similarity vector or a constant feature map), causing division by zero and producing NaN in the softmax weighting of Eq. (7). Add a small epsilon to the denominator or define a fallback value for degenerate distributions, and discuss whether such cases occur during training.
  4. [Throughout (Appendix references)] Key implementation details are consistently deferred to an Appendix that is not present in the submitted manuscript. This includes subvolume sampling, dataset splits, pre-training and fine-tuning hyperparameters, and Sinkhorn-algorithm results. Without these details the experiments are not reproducible. Please include the appendix or move the essential hyperparameters into the main text.
  5. [Section 3.1 and Section 3.2, Eq. (2)] The ground-truth patch correspondence M_gt is computed purely from the affine transformation H, relying on the assumption that token features are geometrically equivariant. If learned positional encodings or downsampling break this equivariance, the patch-to-patch loss supervises incorrect pairs, and this error would propagate into the Sharpe-ratio weighting. This is a correctness-risk concern; please provide a quantitative evaluation of feature equivariance (for example, correspondence accuracy against M_gt on a held-out set) or demonstrate empirically that the training loss corrects the equivariance error over time.
minor comments (7)
  1. [Eq. (1)] There is a typo in the numerator: 'q∗′ i . τ' should read 'q∗′ i / τ'.
  2. [Eq. (2)] The notation is incomplete: the second condition appears as '⟨H−1(cj), ci⟩ ∧' with missing angle brackets around the second metric, and the meaning of the ∧ symbol should be clarified.
  3. [Eq. (4)] There is an extra parenthesis in 'softmax((Mt(·,j))'; it should be 'softmax(Mt(·,j))'.
  4. [Section 3.3] In Eq. (6), σ_Dn is described as the 'variance' of the vector D_n, but the symbol σ conventionally denotes standard deviation; please clarify which is used.
  5. [Table 1] The caption states 'Best and second best are highlighted', but no highlighting is visible in the text, and the organ abbreviations (Spl, RKid, LKid, Gall, Eso, Liv, Sto, Aor, IVC, Veins, Pan, RAG, LAG) are not defined in the caption.
  6. [Table 4] The column header 'A VG Acc(%)' appears to be a typo for 'AVG Acc(%)'.
  7. [Section 4.1] The description of pretraining data is ambiguous: 'BTCV and TCIA COVID-19 are used as pre-training datasets (1k data) for the BTCV segmentation task' should specify the exact number of volumes from each dataset and how the 10k collection is composed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SSL objective uses an external geometric target M_gt and is evaluated on external downstream benchmarks; the soft-label weighting is self-distillation, not a fitted parameter renamed as a prediction.

full rationale

The paper's central derivation is a self-supervised pretraining objective, not a predictive claim derived from fitted inputs. Eq. (2)-(3) build the patch correspondence ground truth M_gt from the known affine augmentation H, an external transformation, and Eq. (5) trains the model to match this target with a dual-softmax similarity map. The patch-to-structure loss (Eq. (6)-(8)) reweights Lp2p using a Sharpe ratio computed from the model's own neighborhood similarity distribution D_n; although this is self-referential in a broad sense, it is a soft self-distillation/regularization mechanism, not the paper 'predicting' a value from the very parameter fitted to it. The reported superiority claim is supported by external downstream metrics (Dice, accuracy, PSNR/SSIM/NMSE) on 10 datasets and is not derived from the SSL loss by construction. No load-bearing self-citations, no imported uniqueness theorem, and no fitted-input-called-prediction steps were found. The concerns that 'most results are drawn from [40,44]' and that the aggregate 'average score' is not normalized are verification/correctness issues about baseline matching and metric definition, not circularity under the stated rubric.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework is built on existing SSL and image-matching machinery. The load-bearing assumptions are the semantic-structure premise and the geometric equivariance of token features, both stated but not formally validated. The Sharpe-ratio weighting introduces a self-referential soft label.

free parameters (2)
  • InfoNCE temperature tau = not reported (MoCo default)
    Appears in Eq. 1 and is a standard hyperparameter for contrastive loss; not fitted to downstream data.
  • Equal loss weights = 1.0, 1.0, 1.0
    The final loss L = L_g + L_p2p + L_p2s uses unit weights with no reported sensitivity analysis.
assumptions (5)
  • standard math Dual-softmax operator approximates optimal transport and yields differentiable soft correspondences.
    Used in Eq. 4 to convert the raw similarity map into a doubly-stochastic-like matrix for the patch-to-patch loss.
  • standard math InfoNCE contrastive loss (MoCo) is a valid SSL objective.
    Used as the baseline loss L_g in Eq. 1 and provides the global contrastive signal.
  • domain assumption Patches within the same anatomical structure share the same semantics, and patches from different structures have distinct semantics.
    Stated in Section 1 and Section 3.3; it is the premise for the entire structure-aware formulation.
  • domain assumption The token features of the Transformer encoder are geometrically equivariant to affine transformations of the input.
    Introduced in Section 3.1 and used to build the ground-truth correspondence matrix M_gt in Eq. 2; if violated, the patch-to-patch supervision is misaligned.
  • ad hoc to paper The neighborhood similarity distribution D_n reflects semantic structure membership, so its Sharpe ratio is a valid soft label for structure consistency.
    Introduced in Section 3.3; this self-referential assumption is used to weight the patch-to-patch loss but is not validated independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning." pith.science (2026). https://pith.science/paper/7CS4OOQ6

@misc{pith2026250702581,
  author       = {Pith},
  title        = {Pith review of: Structure-aware Semantic Discrepancy and Consistency for 3D Medical Image Self-supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7CS4OOQ6}},
  note         = {Machine review of arXiv:2507.02581}
}
abstract

3D medical image self-supervised learning (mSSL) holds great promise for medical analysis. Effectively supporting broader applications requires considering anatomical structure variations in location, scale, and morphology, which are crucial for capturing meaningful distinctions. However, previous mSSL methods partition images with fixed-size patches, often ignoring the structure variations. In this work, we introduce a novel perspective on 3D medical images with the goal of learning structure-aware representations. We assume that patches within the same structure share the same semantics (semantic consistency) while those from different structures exhibit distinct semantics (semantic discrepancy). Based on this assumption, we propose an mSSL framework named $S^2DC$, achieving Structure-aware Semantic Discrepancy and Consistency in two steps. First, $S^2DC$ enforces distinct representations for different patches to increase semantic discrepancy by leveraging an optimal transport strategy. Second, $S^2DC$ advances semantic consistency at the structural level based on neighborhood similarity distribution. By bridging patch-level and structure-level representations, $S^2DC$ achieves structure-aware representations. Thoroughly evaluated across 10 datasets, 4 tasks, and 3 modalities, our proposed method consistently outperforms the state-of-the-art methods in mSSL.

Figures

Figures reproduced from arXiv: 2507.02581 by the authors.

Figure 1
Figure 1. A. Two views of mSSL. B. Similarity heatmaps on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of our SSL framework. S 2DC is established on patch features (i.e., token feature in vision transformer) and incorporates two main steps: (1) Patch-to-patch discrepancy. The model aligns two patch sets, which treats every patch as distinct semantics to enlarge the distribution gap between all semantics. (2) Patch-to-structure consistency. The model applies a soft regularization to align patches within t… view at source ↗
Figure 3
Figure 3. Illustration of the similarity distribution [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: We visualize the first principal components after apply [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: The t-SNE visualization of features from 13 organs on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 42 canonical work pages

  1. [1]

    Vicregl: Self-supervised learning of local visual features

    Adrien Bardes, Jean Ponce, and Yann LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems , 35:8799–8810,

    work page

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 8

    2021

  3. [3]

    Truly shift-invariant convolutional neural networks

    Anadi Chaman and Ivan Dokmanic. Truly shift-invariant convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773–3783, 2021. 3

    work page 2021

  4. [4]

    A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling

    Gaoyu Chen, Sheng Liu, Wenxiang Ding, Li Lv, Chen Zhao, Fenghua Weng, Yong Long, Yunlong Zan, and Qiu Huang. A total-body ultra-low dose pet reconstruction method via image space shuffle u-net and body sampling. IEEE Trans- actions on Radiation and Plasma Medical Sciences, 2023. 1, 5, 6

    work page 2023

  5. [5]

    Jigsaw clustering for unsupervised visual representation learning

    Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526–11535, 2021. 6

    work page 2021

  6. [6]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on ma- chine learning, pages 1597–1607. PMLR, 2020. 6

    work page 2020

  7. [7]

    Masked image modeling advances 3d medical image analysis

    Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 1970–1980, 2023. 6

    work page 1970

  8. [8]

    Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms

    Jiaqi Cui, Pinxian Zeng, Xinyi Zeng, Yuanyuan Xu, Peng Wang, Jiliu Zhou, Yan Wang, and Dinggang Shen. Prior knowledge-guided triple-domain transformer-gan for direct pet reconstruction from low-count sinograms. IEEE Trans- actions on Medical Imaging, 2024. 7

    work page 2024

Show all 51 references
  1. [9]

    Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images

    Tugba Akinci D’Antonoli, Lucas K Berger, Ashraya K In- drakanti, Nathan Vishwanathan, Jakob Weiß, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas K¨ustner, et al. Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr ima...

    2024 arXiv

  2. [10]

    The autopet challenge: towards fully automated lesion segmentation in oncologic pet/ct imaging

    Sergios Gatidis, Marcel Fr ¨uh, Matthias Fabritius, Sijing Gu, Konstantin Nikolaou, Christian La Foug `ere, Jin Ye, Jun- jun He, Yige Peng, Lei Bi, et al. The autopet challenge: towards fully automated lesion segmentation in oncologic pet/ct imaging. 2023. 1, 6

    2023

  3. [11]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272–284. Springer, 2021. 5

    2021

  4. [12]

    Unetr: Transformers for 3d med- ical image segmentation

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d med- ical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574–58...

    2022

  5. [13]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 9729–9738, 2020. 3, 6

    2020

  6. [14]

    Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d med- ical image segmentation

    Yufan He, Vishwesh Nath, Dong Yang, Yucheng Tang, An- driy Myronenko, and Daguang Xu. Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d med- ical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Interven...

    2023

  7. [15]

    Geometric vi- sual similarity learning in 3d medical image self-supervised pre-training

    Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean- Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric vi- sual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9...

    2023

  8. [16]

    Attentive symmetric autoencoder for brain mri segmenta- tion

    Junjia Huang, Haofeng Li, Guanbin Li, and Xiang Wan. Attentive symmetric autoencoder for brain mri segmenta- tion. In International Conference on Medical Image Com- puting and Computer-Assisted Intervention, pages 203–213. Springer, 2022. 2

    2022

  9. [17]

    Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis

    Yankai Jiang, Mingze Sun, Heng Guo, Xiaoyu Bai, Ke Yan, Le Lu, and Minfeng Xu. Anatomical invariance modeling and semantic alignment for self-supervised learning in 3d medical image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 158...

    2023

  10. [18]

    Oasis-3: longitudinal neuroimaging, clin- ical, and cognitive dataset for normal aging and alzheimer disease

    Pamela J LaMontagne, Tammie LS Benzinger, John C Mor- ris, Sarah Keefe, Russ Hornbeck, Chengjie Xiong, Eliza- beth Grant, Jason Hassenstab, Krista Moulder, Andrei G Vlassenko, et al. Oasis-3: longitudinal neuroimaging, clin- ical, and cognitive dataset for normal aging and alz...

    2019

  11. [19]

    Miccai multi-atlas la- beling beyond the cranial vault–workshop and challenge

    Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas la- beling beyond the cranial vault–workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 6

    2015

  12. [20]

    Self- supervised alignment learning for medical image segmenta- tion

    Haofeng Li, Yiming Ouyang, and Xiang Wan. Self- supervised alignment learning for medical image segmenta- tion. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2024. 2

    2024

  13. [21]

    Robust perception through equivariance

    Chengzhi Mao, Lingyu Zhang, Abhishek Vaibhav Joshi, Jun- feng Yang, Hao Wang, and Carl V ondrick. Robust perception through equivariance. In International Conference on Ma- chine Learning, pages 23852–23870. PMLR, 2023. 3

    2023

  14. [22]

    The alzheimer’s disease neuroimaging initiative

    Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer’s disease neuroimaging initiative. Neuroimaging Clinics of North America, 15(4):869, 2005. 5, 6

    2005

  15. [23]

    Repre- sentation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3

    2018 arXiv

  16. [24]

    Neighbourhood con- sensus networks

    Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovi ´c, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood con- sensus networks. Advances in neural information processing systems, 31, 2018. 4

    2018

  17. [25]

    Silhouettes: a graphical aid to the inter- pretation and validation of cluster analysis

    Peter J Rousseeuw. Silhouettes: a graphical aid to the inter- pretation and validation of cluster analysis. Journal of com- putational and applied mathematics, 20:53–65, 1987. 8

    1987

  18. [26]

    A geometric approach to robust medical image segmentation

    Ainkaran Santhirasekaram, Mathias Winkler, Andrea Rock- all, and Ben Glocker. A geometric approach to robust medical image segmentation. Medical Image Analysis , 97: 103260, 2024. 2

    2024

  19. [27]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4

    2020

  20. [28]

    Mutual fund performance

    William F Sharpe. Mutual fund performance. The Journal of business, 39(1):119–138, 1966. 4

    1966

  21. [29]

    A large annotated medical image dataset for the development and evaluation of segmentation algo- rithms

    Amber L Simpson, Michela Antonelli, Spyridon Bakas, Michel Bilello, Keyvan Farahani, Bram Van Ginneken, An- nette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, et al. A large annotated medical image dataset for the development and evaluation of segmentation a...

    1902 arXiv

  22. [30]

    A relationship between arbitrary positive matrices and doubly stochastic matrices.The annals of math- ematical statistics, 35(2):876–879, 1964

    Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices.The annals of math- ematical statistics, 35(2):876–879, 1964. 4

    1964

  23. [31]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8922–8931, 2021. 4

    2021

  24. [32]

    3d self-supervised methods for medical imaging

    Aiham Taleb, Winfried Loetzsch, Noel Danz, Julius Severin, Thomas Gaertner, Benjamin Bergner, and Christoph Lippert. 3d self-supervised methods for medical imaging. Advances in neural information processing systems, 33:18158–18172,

  25. [33]

    Self-supervised pre-training of swin trans- formers for 3d medical image analysis

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin trans- formers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

    2022

  26. [34]

    Revisiting rubik’s cube: Self-supervised learning with volume-wise transformation for 3d medical image seg- mentation

    Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik’s cube: Self-supervised learning with volume-wise transformation for 3d medical image seg- mentation. In Medical Image Computing and Computer As- sisted Intervention–MICCAI 2020: 23rd International ...

    2020

  27. [35]

    Elements of informa- tion theory

    MTCAJ Thomas and A Thomas Joy. Elements of informa- tion theory. Wiley-Interscience, 2006. 4

    2006

  28. [36]

    Self-supervised learning from a multi-view perspective

    Yao-Hung Hubert Tsai, Yue Wu, Ruslan Salakhutdinov, and Louis-Philippe Morency. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576 ,

    2006 arXiv

  29. [37]

    Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding

    Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bo- hao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

    2024

  30. [38]

    Dense contrastive learning for self-supervised visual pre-training

    Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3024–3033, 2021. 3

    2021

  31. [39]

    Swinmm: masked multi-view with swin trans- formers for 3d medical image segmentation

    Yiqing Wang, Zihan Li, Jieru Mei, Zihao Wei, Li Liu, Chen Wang, Shengtian Sang, Alan L Yuille, Cihang Xie, and Yuyin Zhou. Swinmm: masked multi-view with swin trans- formers for 3d medical image segmentation. InInternational Conference on Medical Image Computing and Computer- ...

    2023

  32. [40]

    V oco: A simple- yet-effective volume contrastive learning framework for 3d medical image analysis

    Linshan Wu, Jiaxin Zhuang, and Hao Chen. V oco: A simple- yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 22873–22882, 2024. 1, 2, 3, 5, 6, 7, 8

    2024

  33. [41]

    Simmim: A simple framework for masked image modeling

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653–9663, 2022. 6

    2022

  34. [42]

    Keypoint-augmented self-supervised learn- ing for medical image segmentation with limited annota- tion

    Zhangsihao Yang, Mengwei Ren, Kaize Ding, Guido Gerig, and Yalin Wang. Keypoint-augmented self-supervised learn- ing for medical image segmentation with limited annota- tion. Advances in Neural Information Processing Systems , 36, 2024. 2

    2024

  35. [43]

    Generative adversar- ial network in medical imaging: A review

    Xin Yi, Ekta Walia, and Paul Babyn. Generative adversar- ial network in medical imaging: A review. Medical image analysis, 58:101552, 2019. 7

    2019

  36. [44]

    Dive into the de- tails of self-supervised learning for medical image analysis

    Chuyan Zhang, Hao Zheng, and Yun Gu. Dive into the de- tails of self-supervised learning for medical image analysis. Medical Image Analysis, 89:102879, 2023. 6 10

    2023

  37. [45]

    Clinically applicable ai system for accu- rate diagnosis, quantitative measurements, and prognosis of covid-19 pneumonia using computed tomography

    Kang Zhang, Xiaohong Liu, Jun Shen, Zhihuan Li, Ye Sang, Xingwang Wu, Yunfei Zha, Wenhua Liang, Chengdi Wang, Ke Wang, et al. Clinically applicable ai system for accu- rate diagnosis, quantitative measurements, and prognosis of covid-19 pneumonia using computed tomography. Cel...

    2020

  38. [46]

    Positional label for self- supervised vision transformer

    Zhemin Zhang and Xun Gong. Positional label for self- supervised vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 3516–3524,

  39. [47]

    Preservational learning improves self- supervised medical image models by reconstructing diverse contexts

    Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. Preservational learning improves self- supervised medical image models by reconstructing diverse contexts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3499–3509, 2021. 6

    2021

  40. [48]

    Pcrlv2: A unified visual information preserva- tion framework for self-supervised pre-training in medical image analysis

    Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. Pcrlv2: A unified visual information preserva- tion framework for self-supervised pre-training in medical image analysis. arXiv preprint arXiv:2301.00772, 2023. 2, 3, 6, 7

    2023 arXiv

  41. [49]

    Models genesis

    Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Got- way, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 6

    2021

  42. [50]

    Mim: Mask in mask self- supervised pre-training for 3d medical image analysis

    Jiaxin Zhuang, Linshan Wu, Qiong Wang, Varut Vardhan- abhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self- supervised pre-training for 3d medical image analysis. arXiv preprint arXiv:2404.15580, 2024. 2

    2024 arXiv

  43. [51]

    Advanc- ing volumetric medical image segmentation via global-local masked autoencoder

    Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advanc- ing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913 ,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.