Pith. sign in

REVIEW 4 major objections 5 minor 49 references

ConceptVAE: Self-Supervised Fine-Grained Concept Disentanglement from 2D Echocardiographies

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a self-supervised VAE whose latent space is a grid of discrete concepts plus per-concept style vectors can find anatomical structures in 2D echocardiography and beat standard SSL baselines on retrieval, segmentation…

desk verdict A promising concept-style disentanglement architecture with nice qualitative results, but the quantitative claims of superiority over standard SSL are undercut by uncontrolled readout size and data leakage, so the evaluation needs major rework. read the letter →

arxiv 2502.01335 v1 pith:2F6MVJ24 submitted 2025-02-03 cs.CV

Costin F. Ciusdel , Alex Serban , Tiziano Passerini This is my paper · ORCID
classification cs.CV
keywords self-supervisedlearningconceptdisentanglementvariationalautoencoderechocardiographycontent-styleseparationout-of-distributiondetectionsemanticsegmentationinstanceretrieval
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 claims that the single-vector embeddings used by standard self-supervised learning are too coarse to capture fine-grained anatomical concepts, and that a model which explicitly factors an image into discrete concepts and per-concept style vectors can learn those concepts without any labels. It introduces ConceptVAE, a variational autoencoder whose latent space is a spatial grid of discrete concept probabilities plus continuous style vectors, trained with a set of losses that enforce concept-style separation. On 2D echocardiography, the model assigns stable concepts to anatomical structures such as blood pools, septum walls, and the left-ventricle free wall. The authors show this learned representation outperforms a VICReg baseline on region-based instance retrieval, semantic segmentation, near-OOD detection, and aortic-valve detection, and can generate in-distribution synthetic frames by altering only the style. If correct, this would make unsupervised pretraining more interpretable for medical imaging while also improving several concrete downstream tasks.

What carries the argument

The central mechanism is the concept discretizer and concept stylizer pair inside a VAE-like autoencoder. The discretizer maps each grid location to a probability distribution over $C$ discrete concepts, samples a one-hot concept via the Gumbel-Softmax trick with straight-through gradients, and indexes a shared learned embedding matrix; the stylizer then produces a continuous style vector at each location as a function of both the local features and the sampled concept, so style is concept-conditioned and different concept choices yield different style vectors. A set of losses drives the separation: reconstruction from concepts alone against a blurred image (forcing concepts to carry semantics), unit-covariance and zero-mean constraints on style (forcing style to be whitened), a consistency loss between EMA-computed concept maps of differently augmented views, KL priors that keep concept probabilities uniform inside the ultrasound cone and concentrated on a background concept outside it, and a cluster loss that penalises fine-grained concept transitions. The Feature Decoder reconstructs low-level features from $k \times k$ neighbourhoods of the latent grid ($k=5$), letting neighbouring concepts cooperate into larger anatomical structures.

What would settle it

Retrain the VICReg baseline with the same linear-probe kernel sizes ($5\times5$ to $9\times9$) and with PLAX views removed from its pretraining set, then re-run the Table 2 segmentation and Section 5.3 OOD experiments; if the mAP, Dice-loss and AuROC gaps do not exceed noise, the claim that concept–style disentanglement outperforms classical SSL is not supported. A second check is to use style-only descriptors for the retrieval task, which should perform near chance if styles truly carry no semantics.

Watch

Extended reading notes

Core claim

ConceptVAE demonstrates that a self-supervised variational autoencoder can discretise a 2D cardiac ultrasound into a small preset number of fine-grained concepts (16 in the experiments) arranged on a spatial grid, each with its own low-dimensional style vector, and that this concept–style disentanglement arises natively from the training objective without labels or anatomical annotations. The concept component carries semantic content—specific indices consistently correspond to blood pools, septum walls, the left ventricular free wall, and the right-heart side of the cone—while the style component carries local texture and speckle. The paper argues this is why the resulting embeddings support zero-shot region-based retrieval, better linear-probe segmentation, better near-OOD separation, and improved objectness localisation compared with a classical SSL baseline, and why perturbing only style produces realistic synthetic frames within the original data distribution.

Load-bearing premise

The reported advantages rest on comparing against a VICReg baseline that uses a 1×1 linear probe where ConceptVAE uses 5×5 to 9×9 kernels, and a baseline pretrained on data that already includes the supposedly out-of-distribution PLAX views; if these were matched, the gaps in segmentation and OOD detection could shrink or disappear.

Editorial extensions

If this is right

  • Region-based instance retrieval can be done zero-shot from concept probabilities alone, with more than double the mAP of the baseline (0.370 vs 0.163 averaged over five landmarks).
  • A single linear layer on the frozen ConceptVAE latent space gives segmentation Dice losses below the VICReg baseline when concept and style are combined, and concept-only input still outperforms the baseline.
  • Near-OOD detection of an unseen echocardiographic view reaches 0.753 AuROC with a normalising flow over concept descriptors, 10% above the baseline.
  • In aortic-valve detection on unseen PLAX views, ConceptVAE improves objectness AP (0.786 vs 0.665) while mean AP is comparable, suggesting concept grids localise anatomy even on views never seen in pretraining.
  • Injecting noise into style vectors while keeping concepts fixed generates synthetic frames that stay in-distribution and vary only texture, offering calibrated augmentation for dense prediction tasks.

Reading between the lines

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

  • The paper does not test whether the same concept indices survive across imaging modalities; if they do, the pretraining could transfer to other 2D ultrasound or radiograph tasks, a hypothesis that could be checked by fine-tuning the concept mapping on a second modality.
  • The $5 \times 5$ neighbourhood used for region descriptors means retrieval quality is tied to window size; varying the descriptor context should trade localisation against context, an implicit design knob the paper does not explore.
  • The OOD result suggests concept distributions are more view-invariant than raw features; a direct test would be to measure concept-map alignment between end-diastole and end-systole frames of the same acquisition.
  • The style-perturbation generator could be used as a plug-in augmentation for downstream supervised segmentation models; whether it improves robustness more than classical augmentations is a testable consequence the paper does not yet run.
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. ConceptVAE is a self-supervised pretraining framework for 2D echocardiography that learns a grid of discrete concept codes and per-location continuous style vectors, trained through a VAE-style reconstruction objective with a suite of auxiliary losses (concept consistency, style covariance, priors, and clustering). The paper claims that the resulting latent space is interpretable, with concepts aligning to anatomical structures such as blood pools and septum walls, and that ConceptVAE outperforms a VICReg baseline in region-based instance retrieval, semantic segmentation, near-OOD detection, and object detection, while also enabling style-based synthetic data generation.

Significance. If the quantitative claims were sound, ConceptVAE would be a meaningful contribution: it combines discrete concept discovery with content-style disentanglement in a self-supervised medical-imaging setting, with potentially useful downstream applications in retrieval, detection, and calibrated data augmentation. The qualitative results (Figures 3 and 4) do suggest that the model learns anatomy-aligned concepts without supervision, and the loss design is described in enough detail that the method is largely reproducible. However, the central empirical claims are not supported by the reported experiments: the segmentation comparison is confounded by an unmatched linear-probe kernel size, the object-detection result directly contradicts the abstract, and the OOD evaluation is invalid because the baseline saw the test distribution during pretraining. The manuscript also does not provide code or data, further limiting verification.

major comments (4)
  1. [Section 5.2, Table 2] The comparison between ConceptVAE and the VICReg baseline is confounded by the linear-probe kernel size. VICReg is evaluated only with a 1×1 convolutional kernel, whereas ConceptVAE is evaluated with kernels up to 9×9. The random-init control in the same table shows that this confound is real: a randomly initialized ConceptVAE with a 9×9 kernel on concept & style achieves a Dice loss of 0.1715, which is better than the trained VICReg baseline's 0.187 at 1×1. At the matched 1×1 setting, trained ConceptVAE's concept & style Dice loss is 0.4853, substantially worse than VICReg. Therefore the reported segmentation advantage cannot be attributed to the learned concept-style disentanglement; the readout receptive field is a plausible alternative explanation. The abstract's claim of outperforming traditional SSL in semantic segmentation is unsubstantiated as reported.
  2. [Section 5.4, Table 3; abstract] The abstract states that ConceptVAE "outperforms traditional self-supervised methods in tasks such as ... object detection," but Table 3 reports a mean AP of 0.362 for ConceptVAE versus 0.378 for the VICReg baseline. The text itself acknowledges the baseline is slightly better on mean AP. While the objectness AP is higher for ConceptVAE, the mean AP is the headline detection metric. This direct contradiction between the abstract and the reported results is a load-bearing inconsistency that must be corrected or supported by additional evidence.
  3. [Section 5.3] The near-OOD comparison is invalid as a test of OOD detection because the VICReg baseline was pretrained on a dataset that includes the PLAX views used as the OOD test set, as the text explicitly admits; ConceptVAE was not pretrained on PLAX. Consequently, the higher AuROC of ConceptVAE (0.753 vs 0.655) may simply reflect that the baseline has already seen the "OOD" distribution during pretraining. A valid comparison requires pretraining both models on identical in-distribution data only, or using a test distribution that neither model has seen.
  4. [Section 5.1, Table 1] The retrieval comparison uses different descriptor constructions: ConceptVAE uses a 5×5 neighborhood of concept-probability vectors, whereas the baseline uses a single 1×1 feature vector. Because the baseline descriptor has a much smaller context, the mAP advantage (0.370 vs 0.163) may be due to the descriptor size or the local context rather than to the disentangled representation. Please report a baseline with a comparable 5×5 (or otherwise equivalently sized) contextual descriptor.
minor comments (5)
  1. [Section 5.3] The statement that ConceptVAE has an AuROC "10% larger" than the baseline is imprecise; 0.753 vs 0.655 is a relative improvement of about 15% (or an absolute difference of 0.098).
  2. [Section 3.3 and Equation (3)] The loss weights β1...β9, the Gumbel temperature Tsamp, and the exact values of the number of concepts C and style components S are not all specified in the main text; a hyperparameter table would aid reproducibility.
  3. [Sections 4 and 5.2] The symbol k is used both for the Feature Decoder neighborhood size (Section 4) and for the linear-probe kernel size in Table 2; please use distinct notation to avoid confusion.
  4. [Throughout] The baseline method is spelled inconsistently as "Vicreg" in some places and "VICReg" in others; please standardize the notation.
  5. [Section 5 (all tables)] No confidence intervals or multiple-seed results are reported for any quantitative comparison. Given the small margins in Table 3, error bars or at least a statement of variance across seeds are needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method's inductive biases are not fitted predictions, and the quantitative claims rest on held-out evaluations; the kernel-size and OOD-set mismatches are correctness confounds, not circular steps.

full rationale

The paper does not reduce to its own inputs by construction. The cone prior and cluster loss are hand-set inductive biases that encourage spatially coherent concept maps with a dedicated background concept, but they do not specify which interior concept should correspond to blood pools, septum walls, or other anatomy; the anatomical correspondence is an empirical outcome, not a fitted label. The style/texture separation is encouraged by the blurred-target reconstruction, yet the swap and noise-injection experiments test whether the intended decomposition actually holds, rather than merely restating the loss. All quantitative tasks use independent held-out test sets and external annotations, and no fitted parameter is renamed as a prediction. No self-citations appear in the reference list, so the self-citation patterns do not apply. The Table 2 kernel-size mismatch (VICReg evaluated only at 1×1 while ConceptVAE uses up to 9×9) and the Section 5.3 pretraining-data mismatch (the baseline saw PLAX views) are legitimate evaluation-fairness concerns that undermine the strength of the comparative claims, but they are empirical confounds rather than instances of circular derivation under the stated criteria.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central claim rests on several hand-chosen hyperparameters (C, S, k, loss weights, Gumbel temperature), a domain-specific prior about the ultrasound cone, and a heuristic blur target for style removal. These are not derived from data or theory. The invented entities are latent variables internal to the model; they lack falsifiable handles outside the paper, so they are listed with independent_evidence false.

free parameters (6)
  • Number of concepts C = 16
    Chosen by hand as sufficient for echocardiographies; the paper states 'even a modest number of discrete concepts and styles (e.g., 16 concepts and 8 style components) are sufficient' (Sections 1 and 3.1).
  • Number of style components S = 8
    Same rationale as C; this hyperparameter defines the dimensionality of the style vectors.
  • Feature decoder neighborhood size k = 5
    Selected to balance smooth transitions and local descriptors (Section 4); downstream performance depends heavily on k (Table 2).
  • Loss weights beta1..beta9 = not reported
    The final loss in Eq. (3) is a weighted sum, but the beta values are never given, making the training objective under-specified.
  • Gumbel temperature Tsamp = not reported
    Used in Eq. (1) for concept sampling, but no value or schedule is provided.
  • Concept prior p0(c) for cone interior = uniform over last C-1 concepts, first concept zero mass
    Hand-designed prior based on domain knowledge of ultrasound cone geometry; encodes an expectation of uniform concept prevalence inside the cone.
assumptions (4)
  • domain assumption The ultrasound cone can be identified per image, and all non-cone regions should be modeled by a single background concept (concept 0).
    Used in the prior loss Lprior (Eq. 2) and in evaluation masking; assumes a stable cone geometry across acquisitions.
  • ad hoc to paper Blurring an image removes style while preserving concept content.
    Concept-only reconstruction loss Limg(Irec([xconcept, xstyle:=0]), Iblurred) uses a blurred target to force concepts to encode semantics; this is a heuristic proxy.
  • ad hoc to paper The VICReg baseline is a fair control for isolating the effect of concept-style disentanglement.
    The comparison protocols differ in kernel size (Table 2) and pretraining data (Section 5.3), so the attribution of gains to the proposed method is not established.
  • standard math EMA updates and stop-gradient prevent feature collapse.
    Standard SSL technique; assumed to work as in prior work (e.g., DINO).
invented entities (2)
  • Discrete concept latent variable (concept codebook)
    purpose: To represent semantic content such as anatomical structures at each grid location.
    No external falsifiable prediction; it is an internal representation. Its semantic alignment is only evidenced by qualitative overlays and downstream linear probes within the paper.
  • Continuous style latent variable
    purpose: To capture per-concept local texture and speckle.
    Style has no external handle; the claim that it encodes texture is supported by reconstruction experiments in Section 5.5 but not by an independent measurement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConceptVAE: Self-Supervised Fine-Grained Concept Disentanglement from 2D Echocardiographies." pith.science (2026). https://pith.science/paper/2F6MVJ24

@misc{pith2026250201335,
  author       = {Pith},
  title        = {Pith review of: ConceptVAE: Self-Supervised Fine-Grained Concept Disentanglement from 2D Echocardiographies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2F6MVJ24}},
  note         = {Machine review of arXiv:2502.01335}
}
read the original abstract

While traditional self-supervised learning methods improve performance and robustness across various medical tasks, they rely on single-vector embeddings that may not capture fine-grained concepts such as anatomical structures or organs. The ability to identify such concepts and their characteristics without supervision has the potential to improve pre-training methods, and enable novel applications such as fine-grained image retrieval and concept-based outlier detection. In this paper, we introduce ConceptVAE, a novel pre-training framework that detects and disentangles fine-grained concepts from their style characteristics in a self-supervised manner. We present a suite of loss terms and model architecture primitives designed to discretise input data into a preset number of concepts along with their local style. We validate ConceptVAE both qualitatively and quantitatively, demonstrating its ability to detect fine-grained anatomical structures such as blood pools and septum walls from 2D cardiac echocardiographies. Quantitatively, ConceptVAE outperforms traditional self-supervised methods in tasks such as region-based instance retrieval, semantic segmentation, out-of-distribution detection, and object detection. Additionally, we explore the generation of in-distribution synthetic data that maintains the same concepts as the training data but with distinct styles, highlighting its potential for more calibrated data generation. Overall, our study introduces and validates a promising new pre-training technique based on concept-style disentanglement, opening multiple avenues for developing models for medical image analysis that are more interpretable and explainable than black-box approaches.

Figures

Figures reproduced from arXiv: 2502.01335 by the authors.

Figure 1
Figure 1. ConceptVAE overview, where the blue blocks are trainable while the grey blocks are only updated using [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ConceptVAE model architecture and training setup, where the EMA blocks represent the exponential moving [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Concept maps for three randomly sampled inputs. The 16 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Effect of concept swapping. The left image is the reconstruction based only on the greedy concept map [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Region-based instance retrieval using conceptual search. The leftmost column displays query images, while [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The ROC curves comparison between ConceptVAE and the Vicreg baseline model, for distinguishing in [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Original images (left) displayed alongside reconstructions using [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Reconstructed images with unaltered xstyle (left) alongside three reconstructions with constant noise level β = 0.3. Each noisy reconstruction uses different noise, n ∼ N (0, I), as described in Equation (7). The samples generated with ConceptVAE remain within the orig…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    3d self-supervised methods for medical imaging

    Taleb, A.; Loetzsch, W.; Danz, N.; Severin, J.; Gaertner, T.; Bergner, B.; Lippert, C. 3d self-supervised methods for medical imaging. Adv. Neural Inf. Process. Syst. 2020, 33, 18158–18172

    work page 2020

  2. [2]

    Big self-supervised models advance medical image classification

    Azizi, S.; Mustafa, B.; Ryan, F.; Beaver, Z.; Freyberg, J.; Deaton, J.; Loh, A.; Karthikesalingam, A.; Kornblith, S.; Chen, T.; et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, BC, Canada, 11–17 October 2021; pp. 3478–3488

    work page 2021

  3. [3]

    Systematic comparison of semi-supervised and self-supervised learning for medical image classification

    Huang, Z.; Jiang, R.; Aeron, S.; Hughes, M.C. Systematic comparison of semi-supervised and self-supervised learning for medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, W A, USA, 16–22 June 2024; pp. 22282–22293

    work page 2024

  4. [4]

    A Cookbook of Self-Supervised Learning.arXiv 2023

    Balestriero, R.; Ibrahim, M.; Sobal, V .; Morcos, A.; Shekhar, S.; Goldstein, T.; Bordes, F.; Bardes, A.; Mialon, G.; Tian, Y .; et al. A Cookbook of Self-Supervised Learning.arXiv 2023. http://arxiv.org/abs/2304.12210

    arXiv 2023

  5. [5]

    The ssl interplay: Augmentations, inductive bias, and generalization

    Cabannes, V .; Kiani, B.; Balestriero, R.; LeCun, Y .; Bietti, A. The ssl interplay: Augmentations, inductive bias, and generalization. In Proceedings of the International Conference on Machine Learning, PMLR, Honolulu, HI, USA, 23–29 July 2023; pp. 3252–3298

    work page 2023

  6. [6]

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

    Wu, L.; Zhuang, J.; Chen, H. 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, Seattle, W A, USA, 16–22 June 2024; pp. 22873–22882

    work page 2024

  7. [7]

    Data2vec: A general framework for self-supervised learning in speech, vision and language

    Baevski, A.; Hsu, W.N.; Xu, Q.; Babu, A.; Gu, J.; Auli, M. Data2vec: A general framework for self-supervised learning in speech, vision and language. In Proceedings of the International Conference on Machine Learning, PMLR, Baltimore, MD, USA, 25-27 July 2022; pp. 1298–1312

    work page 2022

  8. [8]

    Swinmm: Masked multi-view with swin transformers for 3d medical image segmentation

    Wang, Y .; Li, Z.; Mei, J.; Wei, Z.; Liu, L.; Wang, C.; Sang, S.; Yuille, A.L.; Xie, C.; Zhou, Y . Swinmm: Masked multi-view with swin transformers for 3d medical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Vancouver, BA, Canada, 8–12 October 2023; Springer: Berlin/Heide...

    work page 2023

Show all 49 references
  1. [9]

    Benchmarking Supervised and Self-Supervised Learning Methods in A Large Ultrasound Multi-task Images Dataset

    Liu, P.; Zhang, J.; Wu, X.; Liu, S.; Wang, Y .; Feng, L.; Diao, Y .; Liu, Z.; Lyu, G.; Chen, Y . Benchmarking Supervised and Self-Supervised Learning Methods in A Large Ultrasound Multi-task Images Dataset. IEEE J. Biomed. Health Inform. 2024, Early Access

    2024

  2. [10]

    Efficient deep learning-based automated diagnosis from echocardiography with contrastive self-supervised learning

    Holste, G.; Oikonomou, E.K.; Mortazavi, B.J.; Wang, Z.; Khera, R. Efficient deep learning-based automated diagnosis from echocardiography with contrastive self-supervised learning. Commun. Med. 2024, 4, 133

    2024

  3. [11]

    InsCLR: Improving Instance Retrieval with Self-Supervision

    Deng, Z.; Zhong, Y .; Guo, S.; Huang, W. InsCLR: Improving Instance Retrieval with Self-Supervision. Proc. AAAI Conf. Artif. Intell. 2022, 36, 516–524. https://doi.org/10.1609/aaai.v36i1.19930. 17 ConceptV AE: Self-Supervised Fine-Grained Concept Disentanglement from 2D Echoca...

    2022 doi

  4. [12]

    Deep Learning for Instance Retrieval: A Survey

    Chen, W.; Liu, Y .; Wang, W.; Bakker, E.M.; Georgiou, T.; Fieguth, P.; Liu, L.; Lew, M.S. Deep Learning for Instance Retrieval: A Survey. IEEE Trans. Pattern Anal. Mach. Intell. 2023, 45, 7270–7292. https: //doi.org/10.1109/TPAMI.2022.3218591

    2023

  5. [13]

    Understanding Human Object Vision: A Picture Is Worth a Thousand Rep- resentations

    Bracci, S.; Op de Beeck, H. Understanding Human Object Vision: A Picture Is Worth a Thousand Rep- resentations. Annu. Rev. Psychol. 2023, 74, 113–135. https://doi.org/https://doi.org/10.1146/ annurev-psych-032720-041031

    2023

  6. [14]

    How does the brain solve visual object recognition? Neuron 2012, 73, 415–434

    DiCarlo, J.J.; Zoccolan, D.; Rust, N.C. How does the brain solve visual object recognition? Neuron 2012, 73, 415–434

    2012

  7. [15]

    Recent advances in understanding object recognition in the human brain: Deep neural networks, temporal dynamics, and context

    Wardle, S.G.; Baker, C.I. Recent advances in understanding object recognition in the human brain: Deep neural networks, temporal dynamics, and context. F1000Research 2020, 9, 590

    2020

  8. [16]

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

    Zhang, C.; Zheng, H.; Gu, Y . Dive into the details of self-supervised learning for medical image analysis. Med Image Anal. 2023, 89, 102879

    2023

  9. [17]

    Unsupervised learning of disentangled representation via auto-encoding: A survey

    Eddahmani, I.; Pham, C.H.; Napol ´eon, T.; Badoc, I.; Fouefack, J.R.; El-Bouz, M. Unsupervised learning of disentangled representation via auto-encoding: A survey. Sensors 2023, 23, 2362

    2023

  10. [18]

    A simple framework for contrastive learning of visual repre- sentations

    Chen, T.; Kornblith, S.; Norouzi, M.; Hinton, G. A simple framework for contrastive learning of visual repre- sentations. In Proceedings of the 37th International Conference on Machine Learning, ICML’20, Virtual, 13–18 July 2020

    2020

  11. [19]

    Momentum Contrast for Unsupervised Visual Representation Learning

    He, K.; Fan, H.; Wu, Y .; Xie, S.; Girshick, R. Momentum Contrast for Unsupervised Visual Representation Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, W A, USA, 13–19 June 2020; pp. 9726–9735.https://doi.org/10....

    2020

  12. [20]

    VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning

    Bardes, A.; Ponce, J.; LeCun, Y . VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning. In Proceedings of the 10th International Conference on Learning Representations, Virtual, 25–29 April 2022

    2022

  13. [21]

    Masked Autoencoders Are Scalable Vision Learners

    He, K.; Chen, X.; Xie, S.; Li, Y .; Doll´ar, P.; Girshick, R. Masked Autoencoders Are Scalable Vision Learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 18–24 June 2022; pp. 15979–15988. https://doi.org/...

    2022

  14. [22]

    Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language

    Baevski, A.; Babu, A.; Hsu, W.N.; Auli, M. Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language. In Proceedings of the 40th International Conference on Machine Learning, ICML’23, Honolulu, HI, USA, 23–29 July 2023

    2023

  15. [23]

    On the duality between contrastive and non-contrastive self-supervised learning

    Garrido, Q.; Chen, Y .; Bardes, A.; Najman, L.; LeCun, Y . On the duality between contrastive and non-contrastive self-supervised learning. In Proceedings of the 11th International Conference on Learning Representations, Kigali, Rwanda, 1–5 May 2023

    2023

  16. [24]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the 9th International Conference on ...

    2021

  17. [25]

    Emerging Properties in Self-Supervised Vision Transformers

    Caron, M.; Touvron, H.; Misra, I.; Jegou, H.; Mairal, J.; Bojanowski, P.; Joulin, A. Emerging Properties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, BC, Canada, 11–17 October 2021; pp. 963...

    2021

  18. [26]

    Learning vision from models rivals learning vision from data

    Tian, Y .; Fan, L.; Chen, K.; Katabi, D.; Krishnan, D.; Isola, P. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, W A, USA, 16–22 June 2024; pp. 15887–15898

    2024

  19. [27]

    ACSeg: Adaptive Conceptualization for Unsupervised Semantic Segmentation

    Li, K.; Wang, Z.; Cheng, Z.; Yu, R.; Zhao, Y .; Song, G.; Liu, C.; Yuan, L.; Chen, J. ACSeg: Adaptive Conceptualization for Unsupervised Semantic Segmentation. In Proceedings of the IEEE/CVF Conf. on Com- puter Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 17–2...

    2023

  20. [28]

    Learning disentangled representations in the imaging domain

    Liu, X.; Sanchez, P.; Thermos, S.; O’Neil, A.Q.; Tsaftaris, S.A. Learning disentangled representations in the imaging domain. Med Image Anal. 2022, 80, 102516. https://doi.org/https://doi.org/10.1016/j. media.2022.102516

    2022

  21. [29]

    Disentangled Representation Learning

    Wang, X.; Chen, H.; Tang, S.; Wu, Z.; Zhu, W. Disentangled Representation Learning. arXiv 2023. http://arxiv.org/abs/2211.11695

    2023 arXiv

  22. [30]

    Challenging common assumptions in the unsupervised learning of disentangled representations

    Locatello, F.; Bauer, S.; Lucic, M.; Raetsch, G.; Gelly, S.; Sch ¨olkopf, B.; Bachem, O. Challenging common assumptions in the unsupervised learning of disentangled representations. In Proceedings of the International Conference on Machine Learning, PMLR, Long Beach, CA, USA, ...

    2019

  23. [31]

    ClusterGAN: Latent Space Clustering in Generative Adversarial Networks

    Mukherjee, S.; Asnani, H.; Lin, E.; Kannan, S. ClusterGAN: Latent Space Clustering in Generative Adversarial Networks. Proc. AAAI Conf. Artif. Intell. 2019, 33, 4610–4617. https://doi.org/10.1609/aaai.v33i01. 33014610

    2019 doi

  24. [32]

    Simple disentanglement of style and content in visual representations

    Ngweta, L.; Maity, S.; Gittens, A.; Sun, Y .; Yurochkin, M. Simple disentanglement of style and content in visual representations. In Proceedings of the 40th International Conferenceon Machine Learning, ICML’23, Honolulu, HI, USA, 23–29 July 2023

    2023

  25. [33]

    Generating diverse high-fidelity images with VQ-V AE-2

    Razavi, A.; van den Oord, A.; Vinyals, O. Generating diverse high-fidelity images with VQ-V AE-2. In Proceed- ings of the 33rd International Conference on Neural Information Processing Systems, Vancouver, BC, Canada, 8–14 December 2019; Curran Associates Inc.: Red Hook, NY , USA, 2019

    2019

  26. [34]

    Taming Transformers for High-Resolution Image Synthesis

    Esser, P.; Rombach, R.; Ommer, B. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 20–25 June 2021; pp. 12868–12878. https://doi.org/10.1109/CVPR46437.2021.01268

    2021

  27. [35]

    Disentangled representation learning in cardiac image analysis

    Chartsias, A.; Joyce, T.; Papanastasiou, G.; Semple, S.; Williams, M.; Newby, D.E.; Dharmakumar, R.; Tsaftaris, S.A. Disentangled representation learning in cardiac image analysis. Med Image Anal. 2019, 58, 101535. https://doi.org/https://doi.org/10.1016/j.media.2019.101535

    2019

  28. [36]

    RetCCL: Clustering- guided contrastive learning for whole-slide image retrieval

    Wang, X.; Du, Y .; Yang, S.; Zhang, J.; Wang, M.; Zhang, J.; Yang, W.; Huang, J.; Han, X. RetCCL: Clustering- guided contrastive learning for whole-slide image retrieval. Med Image Anal. 2023, 83, 102645

    2023

  29. [37]

    Self-supervised contrastive learning with random walks for medical image segmentation with limited annotations

    Fischer, M.; Hepp, T.; Gatidis, S.; Yang, B. Self-supervised contrastive learning with random walks for medical image segmentation with limited annotations. Comput. Med Imaging Graph. 2023, 104, 102174

    2023

  30. [38]

    Categorical Reparameterization with Gumbel-Softmax

    Jang, E.; Gu, S.; Poole, B. Categorical Reparameterization with Gumbel-Softmax. In Proceedings of the 5th International Conference on Learning Representations, Toulon, France, 24–26 April 2017

    2017

  31. [39]

    Semantic Image Synthesis with Spatially-Adaptive Normalization

    Park, T.; Liu, M.; Wang, T.; Zhu, J. Semantic Image Synthesis with Spatially-Adaptive Normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 15–20 June 2019; pp. 2332–2341. https://doi.org/10.1109/CVPR.2...

    2019

  32. [40]

    Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Med Image Anal

    Haghighi, F.; Taher, M.R.H.; Gotway, M.B.; Liang, J. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Med Image Anal. 2024, 94, 103086

    2024

  33. [41]

    Light-weight refinenet for real-time semantic segmentation

    Nekrasov, V .; Shen, C.; Reid, I. Light-weight refinenet for real-time semantic segmentation. arXiv 2018. http: //arxiv.org/abs/1810.03272

    2018 arXiv

  34. [42]

    Sketch-based semantic retrieval of medical images

    Kobayashi, K.; Gu, L.; Hataya, R.; Mizuno, T.; Miyake, M.; Watanabe, H.; Takahashi, M.; Takamizawa, Y .; Yoshida, Y .; Nakamura, S.; et al. Sketch-based semantic retrieval of medical images. Med Image Anal. 2024, 92, 103060

    2024

  35. [43]

    Large-scale retrieval for medical image analytics: A comprehensive review

    Li, Z.; Zhang, X.; M ¨uller, H.; Zhang, S. Large-scale retrieval for medical image analytics: A comprehensive review. Med Image Anal. 2018, 43, 66–84

    2018

  36. [44]

    A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection

    Ren, J.; Fort, S.; Liu, J.; Roy, A.G.; Padhy, S.; Lakshminarayanan, B. A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection. arXiv 2021. http://arxiv.org/abs/2106.09022

    2021 arXiv

  37. [45]

    Back to the Basics: Revisiting Out-of-Distribution Detection Baselines

    Kuan, J.; Mueller, J. Back to the Basics: Revisiting Out-of-Distribution Detection Baselines. arXiv 2022. http://arxiv.org/abs/2207.03061

    2022 arXiv

  38. [46]

    Normalizing Flows: An Introduction and Review of Current Methods

    Kobyzev, I.; Prince, S.J.; Brubaker, M.A. Normalizing Flows: An Introduction and Review of Current Methods. IEEE Trans. Pattern Anal. Mach. Intell. 2021, 43, 3964–3979. https://doi.org/10.1109/TPAMI.2020. 2992934

    2021 doi

  39. [47]

    On Feature Collapse and Deep Kernel Learning for Single Forward Pass Uncertainty

    van Amersfoort, J.; Smith, L.; Jesson, A.; Key, O.; Gal, Y . On Feature Collapse and Deep Kernel Learning for Single Forward Pass Uncertainty. arXiv 2022. http://arxiv.org/abs/2102.11409

    2022 arXiv

  40. [48]

    Fast r-cnn

    Girshick, R. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Santiago, Chile, 7–13 December 2015; pp. 1440–1448

    2015

  41. [49]

    Variability of echocardiographic measures of left ventricular diastolic function

    Letnes, J.M.; Eriksen-V olnes, T.; Nes, B.; Wisløff, U.; Salvesen, O.; Dalen, H. Variability of echocardiographic measures of left ventricular diastolic function. The HUNT study.Echocardiography 2021, 38, 901–908. https: //doi.org/https://doi.org/10.1111/echo.15073. 19

    2021 doi

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.