Pith. sign in

REVIEW 4 major objections 6 minor 43 references

ModAn-MulSupCon: Modality-and Anatomy-Aware Multi-Label Supervised Contrastive Pretraining for Medical Imaging

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Encoding each image's modality and anatomy as a multi-hot label and training with a Jaccard-weighted supervised contrastive loss beats SimCLR, ImageNet, and supervised baselines after fine-tuning.

desk verdict A modest, clearly written empirical study whose practical claim is plausible but whose statistics and missing ablation keep it from being solid; worth a revision. read the letter →

arxiv 2508.18613 v1 pith:NTRV7TAA submitted 2025-08-26 eess.IV cs.LG

Eichi Takaya , Ryusei Inamori This is my paper · ORCID
classification eess.IVcs.LG
keywords multi-labelsupervisedcontrastivelearningmedicalimagepretrainingmodalityandanatomymetadataJaccardsimilaritytransferfine-tuninglinearprobingradiology
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 sets out to show that two pieces of metadata almost every medical image already carries—imaging modality and anatomical region—can replace expert disease labels as a pretraining signal. The authors encode these attributes as a multi-hot vector and train a ResNet-18 with a Jaccard-weighted multi-label supervised contrastive loss on 16,222 radiologic images, then transfer to three binary classification tasks. After full fine-tuning, the method reaches AUC 0.964 on ACL tear and 0.763 on thyroid nodule malignancy, above all baselines with $p<0.05$, and 0.926 on breast ultrasound, statistically tied with the best baseline. With the encoder frozen, SimCLR and ImageNet win, so the paper's practical claim is conditional: metadata-supervised pretraining is a strong initialization when fine-tuning is feasible, not a universal replacement for instance-level or generic pretraining. If correct, this matters because modality and anatomy labels can be harvested automatically from hospital archives without expert annotation.

What carries the argument

The load-bearing object is the multi-label supervised contrastive loss in Eq. (3): a multi-hot vector $y_i\in\{0,1\}^k$ is formed by concatenating one-hot modality and one-hot anatomy encodings, and for each anchor $a$ the positive set is $\mathcal{P}_\tau(a)=\{p : \mathrm{Jaccard}(y_a,y_p)\ge\tau\}$, with each positive contributing a weight equal to its Jaccard similarity. This turns partially matching metadata into soft positives, so the embedding space is organized by modality-specific imaging physics and anatomical structure rather than by instance identity or disease label. The threshold $\tau=0.3$ and temperature $T=0.07$ control how strict those soft positives are, and this is what allows a single encoder to learn hierarchical proximity between, for example, a CT of the abdomen and an MR of the abdomen. It is the reason the paper can claim fine-tuning efficiency: the geometry built during pretraining is shaped into task boundaries rather than learned from scratch.

What would settle it

Re-run the ACL-tear and thyroid comparisons using ten fresh random 5-fold partitions of the same datasets, or with a short per-initialization learning-rate sweep for fine-tuning, and check whether the proposed encoder still beats the SimCLR baseline with $p<0.05$; if the advantage does not survive split-selection variance or a tuned schedule, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that coarse metadata supervision creates a representation space that is unusually easy to specialize. Instead of pulling two augmented views of the same image together (SimCLR) or pulling same-class images together (supervised contrastive learning), the objective pulls images together in proportion to how much of their modality–anatomy label set they share, so a knee MR and a brain MR are weakly attracted while two ultrasound images of the thyroid are strongly attracted. After fine-tuning, the resulting encoder gives AUC 0.964 on ACL tear and 0.763 on thyroid malignancy, surpassing every baseline at $p<0.05$, and 0.926 on breast malignancy, statistically tied with SimCLR's 0.940. Under linear probing the same encoder trails SimCLR and ImageNet, which the paper reads as evidence that the learned features are less linearly separable but more plastic under task adaptation. The paper frames this as a practical division of labor: use modality- and anatomy-aware pretraining when fine-tuning is possible, and use SimCLR or ImageNet for frozen-encoder deployment.

Load-bearing premise

The significance claim depends on a fixed 5-fold split of each downstream dataset, with the ten repeats varying only training stochasticity, and on a single shared fine-tuning recipe; if split selection or a method-specific schedule changes the ranking, the reported $p<0.05$ advantages could disappear.

Editorial extensions

If this is right

  • On ACL-tear and thyroid-nodule classification, initializing from the metadata-supervised encoder is preferable to SimCLR, ImageNet, supervised multi-class pretraining, or random initialization when full fine-tuning is affordable.
  • In lightweight deployments where the encoder must stay frozen, SimCLR or ImageNet initializations remain the better choice, since linear-probe AUC favors them on all three downstream tasks.
  • Because modality and anatomy labels require no expert annotation, the pretraining signal can be generated automatically for nearly any image in a hospital archive, making scale possible at low labeling cost.
  • The claimed advantage is measured under a single shared fine-tuning recipe, so the practical recommendation carries a condition: it holds for fine-tuning setups comparable to the one tested.

Reading between the lines

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

  • Because the loss operates purely on embeddings and multi-hot labels, the mechanism is not ResNet-specific; a similar fine-tuning advantage would likely appear for vision-transformer or 3D encoders, which the paper lists as future work but does not test.
  • A hybrid objective that adds an instance-contrastive term to the Jaccard-weighted loss could plausibly recover frozen-encoder linear-probe performance while keeping the fine-tuning gains; the paper mentions this hybridization as a future direction but does not investigate it.
  • The Breast ultrasound result suggests that coarse modality and anatomy labels may be insufficient in domains dominated by acquisition heterogeneity; adding metadata bits such as device or probe type is a direct, testable extension of the same multi-hot mechanism.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces ModAn-MulSupCon, a pretraining method that encodes modality and anatomical-region metadata as a multi-hot label and trains a ResNet-18 with a Jaccard-weighted multi-label supervised contrastive loss (Eq. 3) on a 16,222-image subset of RadImageNet (miniRIN). The pretrained encoder is transferred to three binary classification tasks (MRNet-ACL, breast ultrasound, thyroid ultrasound) under both full fine-tuning and linear probing, and compared against SimCLR pretrained on miniRIN, a 165-class supervised baseline, ImageNet initialization, and random initialization. The authors report that under fine-tuning ModAn-MulSupCon achieves the best AUC on ACL and Thyroid, and ranks second on Breast, while under linear probing SimCLR/ImageNet are superior; they conclude that the modality/anatomy metadata signal is most useful when fine-tuning is feasible.

Significance. If the causal attribution holds, the contribution is practically useful: it turns cheap, ubiquitous metadata into a pretraining signal for label-scarce medical imaging, and it clearly separates the fine-tuning and frozen-encoder regimes. The study has clear strengths: it uses public datasets, compares several meaningful baselines, fixes the downstream protocol across methods, reports repeated runs over training stochasticity, and explicitly acknowledges its limitations. However, the paper's central claim—that the modality/anatomy label content, rather than multi-label supervised contrastive pretraining on in-domain data, drives the gains—is not tested by any ablation, and the p-value reporting is inconsistent with the stated two-sided test. The method's novelty is also limited to the choice of label targets, since Eq. (3) is explicitly taken from prior work [38].

major comments (4)
  1. [Section 5, Tables 4-5] The paper attributes the fine-tuning gains to modality- and anatomy-aware supervision, but this attribution is not tested. Equation (3) is the multi-label supervised contrastive loss of [38]; the only proposed change is the label signal. A control experiment that keeps Eq. (3), T=0.07, tau=0.3, the optimizer, and the downstream protocol fixed while replacing the modality/anatomy multi-hot targets with (a) the 165 disease classes already in miniRIN and (b) random or permuted multi-hot codes would determine whether the metadata content, rather than multi-label supervised contrastive pretraining on in-domain data, drives the improvements. Without such a control, the conclusion in Section 5 is overclaimed.
  2. [Section 4.2, Table 4] The p-value reporting is internally inconsistent. Section 4.2 states a two-sided Wilcoxon signed-rank test, but Table 4 reports p=1.000 for miniRINSimCLR on Breast even though its AUC (0.940) exceeds ModAn-MulSupCon (0.926), and Table 5 reports p=1.000 or p=0.999 in several rows where the baseline is numerically better. Under a two-sided test, a large gap in either direction should produce a small p-value; p=1.000 suggests a one-sided test in favor of the proposed method was used. Please clarify the test direction, apply a uniform two-sided test, and report effect sizes or confidence intervals alongside p-values. This matters for the abstract's claim 'surpassing all baselines (p < 0.05)'.
  3. [Section 4.2, Table 3] The repeated runs use the same fixed 5-fold partition and therefore measure only training stochasticity, not split-selection variance. On datasets as small as 349 images (Thyroid), the fixed split can determine the ranking. The claim that ModAn-MulSupCon 'surpasses all baselines (p<0.05)' should be supported by repeating the whole procedure over multiple random 5-fold partitions (e.g., several different 5-fold seeds) or by reporting bootstrap intervals over split-induced variation. Additionally, the single fine-tuning recipe in Table 3 (10 epochs, Adam 1e-4, StepLR at epoch 5) is assumed equally fair to all initializations; a sensitivity check with a longer schedule or a small learning-rate sweep would strengthen the conclusion.
  4. [Abstract, Section 4.1.1] The abstract and conclusion describe the signal as 'scalable', but the evidence is limited to miniRIN (16,222 images), and the limitations paragraph acknowledges this. Either temper the scalability claim or add a small-scale study (e.g., pretraining on 25%, 50%, and 100% of miniRIN) to show that the benefit is not an artifact of the subset size.
minor comments (6)
  1. [Section 3.1] Please specify the number of modality and anatomy classes in miniRIN and how exactly the two one-hot vectors are concatenated; the current description leaves k and the label vocabulary undefined.
  2. [Equation (3)] Please state how empty positive sets Pτ(a) are handled (e.g., skipping the anchor or falling back to instance-level positives), because with a Jaccard threshold of 0.3 some anchors may have no in-batch positive.
  3. [Table 2] The miniRIN165 baseline uses Adam while the contrastive methods use SGD; this confounds the comparison. Please use the same optimizer or justify the difference.
  4. [Figure 2] The claim that ModAn-MulSupCon clusters are 'more entangled' is qualitative; adding a quantitative measure (e.g., silhouette score or k-NN accuracy per metadata class) would make the interpretation more rigorous.
  5. [General] Please include a code/data availability statement; the paper provides detailed hyperparameters but no information about whether the pretraining and evaluation code will be released.
  6. [General] The abstract says 'independent one-hot vectors' but the method in Section 3.1 describes a single multi-hot vector; please unify the terminology for consistency.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: pretraining labels are external to downstream tasks, the contrastive loss is cited from prior work, and the only self-citation is non-load-bearing.

full rationale

The paper's derivation chain is self-contained rather than circular. The proposed method encodes modality and anatomy as a multi-hot target and optimizes a Jaccard-weighted multi-label supervised contrastive loss (Eq. 3), explicitly attributed to external prior work [37, 38]. The pretraining labels are metadata external to the downstream binary tasks (ACL tear, breast lesion malignancy, thyroid nodule malignancy), so no downstream target is used to define the pretraining signal. No fitted parameter is later renamed as a prediction: the pretraining hyperparameters (T=0.07, tau=0.3, optimizer, epochs) are fixed before downstream evaluation, and the downstream results are measured against external baselines (SimCLR, ImageNet, miniRIN165, Scratch) on fixed 5-fold splits. The only self-citation is reference [3] in the related-work survey of pretext tasks ('jigsaw puzzles [3]'), which is not load-bearing for the central claim. The Discussion's interpretation that modality/anatomy metadata encode 'hierarchical proximity' and 'modality-specific physics' is an untested causal attribution, and the absence of an ablation varying the label target is a completeness limitation, not a circularity: the paper's empirical claim concerns the specific proposed initialization compared with external baselines, and that comparison does not reduce to its own inputs. The paper also explicitly acknowledges its limitations (miniRIN scale, restricted modality diversity, ResNet-18 backbone only), further indicating that the central empirical result is not forced by definition or by a self-citation chain.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method introduces no invented entities. The main free knobs are the Jaccard threshold and temperature; the effectiveness of the metadata pretraining signal and the representativeness of the miniRIN sample are domain assumptions. The statistical design uses fixed splits, which is an additional premise behind the significance claims.

free parameters (2)
  • Jaccard similarity threshold tau = 0.3
    Samples with Jaccard similarity at least 0.3 are treated as positives in Eq. 3; this value is chosen by hand and directly controls the positive set, and it is not swept in the paper.
  • Contrastive temperature T = 0.07 for ModAn-MulSupCon, 0.10 for miniRINSimCLR
    The temperature in Eq. 3 and NT-Xent is fixed per method and not tuned jointly; differing temperatures across methods complicate the comparison.
assumptions (3)
  • domain assumption Modality and anatomy metadata provide a sufficient and useful pretext signal for downstream medical classification.
    Section 3.1 encodes the metadata as the multi-hot target; if these metadata are not informative for the downstream tasks, the pretraining could be neutral or harmful, and the linear-probing results suggest weaker linear separability.
  • domain assumption The miniRIN random subset, capped at 100 images per class, preserves enough of RadImageNet's diversity for a fair pretraining comparison.
    Section 4.1.1 describes sampling up to 100 images per class from 165 classes, yielding 16,222 images; no seed or class distribution is reported, and this is a small fraction of the 1.3M-image dataset.
  • domain assumption A single fine-tuning recipe (10 epochs, Adam 1e-4, StepLR at epoch 5) is equally fair across all initializations.
    Section 4.3 and Table 3 apply one recipe to all methods; the authors did not tune per method, but neither did they test whether ModAn-MulSupCon would still lead under other schedules.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ModAn-MulSupCon: Modality-and Anatomy-Aware Multi-Label Supervised Contrastive Pretraining for Medical Imaging." pith.science (2026). https://pith.science/paper/NTRV7TAA

@misc{pith2026250818613,
  author       = {Pith},
  title        = {Pith review of: ModAn-MulSupCon: Modality-and Anatomy-Aware Multi-Label Supervised Contrastive Pretraining for Medical Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTRV7TAA}},
  note         = {Machine review of arXiv:2508.18613}
}
abstract

Background and objective: Expert annotations limit large-scale supervised pretraining in medical imaging, while ubiquitous metadata (modality, anatomical region) remain underused. We introduce ModAn-MulSupCon, a modality- and anatomy-aware multi-label supervised contrastive pretraining method that leverages such metadata to learn transferable representations. Method: Each image's modality and anatomy are encoded as a multi-hot vector. A ResNet-18 encoder is pretrained on a mini subset of RadImageNet (miniRIN, 16,222 images) with a Jaccard-weighted multi-label supervised contrastive loss, and then evaluated by fine-tuning and linear probing on three binary classification tasks--ACL tear (knee MRI), lesion malignancy (breast ultrasound), and nodule malignancy (thyroid ultrasound). Result: With fine-tuning, ModAn-MulSupCon achieved the best AUC on MRNet-ACL (0.964) and Thyroid (0.763), surpassing all baselines ($p<0.05$), and ranked second on Breast (0.926) behind SimCLR (0.940; not significant). With the encoder frozen, SimCLR/ImageNet were superior, indicating that ModAn-MulSupCon representations benefit most from task adaptation rather than linear separability. Conclusion: Encoding readily available modality/anatomy metadata as multi-label targets provides a practical, scalable pretraining signal that improves downstream accuracy when fine-tuning is feasible. ModAn-MulSupCon is a strong initialization for label-scarce clinical settings, whereas SimCLR/ImageNet remain preferable for frozen-encoder deployments.

Figures

Figures reproduced from arXiv: 2508.18613 by the authors.

Figure 1
Figure 1. ModAn-MulSupCon: a shared encoder maps images to embeddings trained with a Jaccard-weighted [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. UMAP projections of penultimate-layer embeddings from five initializations— ModAn-MulSupCon, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [38]

    Contrastive learning for multi-label classification

    Vladimir Zaigrajew and Maciej Zieba. Contrastive learning for multi-label classification. In Proceedings of Conference on Neural Information Processing Systems, New Orleans, pages 1–8, 2022

    work page 2022

  2. [1]

    A survey on self-supervised methods for visual representation learning

    Tobias Uelwer, Jan Robine, Stefan Sylvius Wagner, Marc H¨oftmann, Eric Upschulte, Sebastian Konietzny, Maike Behrendt, and Stefan Harmeling. A survey on self-supervised methods for visual representation learning. Ma- 7 E. Takaya and R. Inamori chine Learning, 114(4):1–56, 2025. Publisher: Springer

    work page 2025

  3. [2]

    Self-supervised learning for medical image analysis: a comprehensive review

    Veenu Rani, Munish Kumar, Aastha Gupta, Monika Sachdeva, Ajay Mittal, and Krishan Kumar. Self-supervised learning for medical image analysis: a comprehensive review. Evolving Systems, 15(4):1607–1633, 2024. Pub- lisher: Springer

    work page 2024

  4. [3]

    Breast cancer classification based on breast tissue structures using the Jigsaw puzzle task in self-supervised learning

    Keisuke Sugawara, Eichi Takaya, Ryusei Inamori, Yuma Konaka, Jumpei Sato, Yuta Shiratori, Fumihito Hario, Tomoya Kobayashi, Takuya Ueda, and Yoshikazu Okamoto. Breast cancer classification based on breast tissue structures using the Jigsaw puzzle task in self-supervised learning. Radiological Physics and Technology, pages 1–10, 2025. Publisher: Springer

    work page 2025

  5. [4]

    Distributed contrastive learning for medical image segmentation

    Yawen Wu, Dewen Zeng, Zhepeng Wang, Yiyu Shi, and Jingtong Hu. Distributed contrastive learning for medical image segmentation. Medical Image Analysis, 81:102564, 2022. Publisher: Elsevier

    work page 2022

  6. [5]

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

    Jiaxin Zhuang, Linshan Wu, Qiong Wang, Peng Fei, Varut Vardhanabhuti, Lin Luo, and Hao Chen. Mim: Mask in mask self-supervised pre-training for 3d medical image analysis. IEEE Transactions on Medical Imaging ,

    work page

  7. [6]

    Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E

    Xueyan Mei, Zelong Liu, Philip M. Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E. Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A. Fayad, and Yang Yang. RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning. Radiology: Artificial Intelligence, 4(5):e2...

    2022

  8. [7]

    Vis-mae: An efficient self-supervised learning approach on medical image segmentation and classification

    Zelong Liu, Andrew Tieu, Nikhil Patel, George Soultanidis, Louisa Deyer, Ying Wang, Sean Huver, Alexander Zhou, Yunhao Mei, Zahi A Fayad, et al. Vis-mae: An efficient self-supervised learning approach on medical image segmentation and classification. In International Workshop on Machine Learning in Medical Imaging , pages 95–107. Springer, 2024

    work page 2024

Show all 43 references
  1. [8]

    Rotation- oriented collaborative self-supervised learning for retinal disease diagnosis

    Xiaomeng Li, Xiaowei Hu, Xiaojuan Qi, Lequan Yu, Wei Zhao, Pheng-Ann Heng, and Lei Xing. Rotation- oriented collaborative self-supervised learning for retinal disease diagnosis. IEEE Transactions on Medical Imaging, 40(9):2284–2294, 2021. Publisher: IEEE

    2021

  2. [9]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  3. [10]

    Momentum contrast for unsupervised visual representation learning

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

    2020

  4. [11]

    UKSSL: Underlying knowledge based semi- supervised learning for medical image classification

    Zeyu Ren, Xiangyu Kong, Yudong Zhang, and Shuihua Wang. UKSSL: Underlying knowledge based semi- supervised learning for medical image classification. IEEE Open Journal of Engineering in Medicine and Biol- ogy, 5:459–466, 2023. Publisher: IEEE

    2023

  5. [12]

    Moco pretraining improves representation and transferability of chest x-ray models

    Hari Sowrirajan, Jingbo Yang, Andrew Y Ng, and Pranav Rajpurkar. Moco pretraining improves representation and transferability of chest x-ray models. InMedical Imaging with Deep Learning, pages 728–744. PMLR, 2021

    2021

  6. [13]

    Gloria: A multimodal global-local rep- resentation learning framework for label-efficient medical image recognition

    Shih-Cheng Huang, Liyue Shen, Matthew P Lungren, and Serena Yeung. Gloria: A multimodal global-local rep- resentation learning framework for label-efficient medical image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3942–3951, 2021

    2021

  7. [14]

    Making the most of text semantics to improve biomedical vision–language processing

    Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, and others. Making the most of text semantics to improve biomedical vision–language processing. In Eur...

    2022

  8. [15]

    Self pre-training with masked autoencoders for medical image classification and segmentation

    Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image classification and segmentation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1–6. IEEE, 2023

    2023

  9. [16]

    Advancing V olumetric Medical Image Segmentation via Global-Local Masked Autoencoders

    Jiaxin Zhuang, Luyang Luo, Qiong Wang, Mingxiang Wu, Lin Luo, and Hao Chen. Advancing V olumetric Medical Image Segmentation via Global-Local Masked Autoencoders. IEEE Transactions on Medical Imaging,

  10. [17]

    Swin MAE: masked autoencoders for small datasets

    Zi’an Xu, Yin Dai, Fayu Liu, Weibing Chen, Yue Liu, Lifu Shi, Sheng Liu, and Yuhang Zhou. Swin MAE: masked autoencoders for small datasets. Computers in biology and medicine , 161:107037, 2023. Publisher: Elsevier

    2023

  11. [18]

    Unest: local spatial representation learning with hierarchical transformer for efficient medical segmentation

    Xin Yu, Qi Yang, Yinchi Zhou, Leon Y Cai, Riqiang Gao, Ho Hin Lee, Thomas Li, Shunxing Bao, Zhoubing Xu, Thomas A Lasko, and others. Unest: local spatial representation learning with hierarchical transformer for efficient medical segmentation. Medical Image Analysis, 90:102939...

    2023

  12. [19]

    Video-CT MAE: Self-supervised Video-CT Domain Adaptation for Vertebral Fracture Diagnosis

    Lukas Buess, Marijn F Stollenga, David Schinz, Benedikt Wiestler, Jan Kirschke, Andreas Maier, Nassir Navab, and Matthias Keicher. Video-CT MAE: Self-supervised Video-CT Domain Adaptation for Vertebral Fracture Diagnosis. In Medical Imaging with Deep Learning, 2024

    2024

  13. [20]

    Self-supervised pre-training of swin transformers 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 transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognit...

    2022

  14. [21]

    Mining multi-label data

    Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vlahavas. Mining multi-label data. Data mining and knowl- edge discovery handbook, pages 667–685, 2010. Publisher: Springer

    2010

  15. [22]

    Learning multi-label scene classifica- tion

    Matthew R Boutell, Jiebo Luo, Xipeng Shen, and Christopher M Brown. Learning multi-label scene classifica- tion. Pattern recognition, 37(9):1757–1771, 2004. Publisher: Elsevier

    2004

  16. [23]

    ML-KNN: A lazy learning approach to multi-label learning

    Min-Ling Zhang and Zhi-Hua Zhou. ML-KNN: A lazy learning approach to multi-label learning. Pattern recognition, 40(7):2038–2048, 2007. Publisher: Elsevier

    2007

  17. [24]

    A kernel method for multi-labelled classification

    Andr ´e Elisseeff and Jason Weston. A kernel method for multi-labelled classification. Advances in neural infor- mation processing systems, 14, 2001

    2001

  18. [25]

    Cnn-rnn: A unified framework for multi-label image classification

    Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2285–2294, 2016

    2016

  19. [26]

    Learning spatial regularization with image-level supervisions for multi-label image classification

    Feng Zhu, Hongsheng Li, Wanli Ouyang, Nenghai Yu, and Xiaogang Wang. Learning spatial regularization with image-level supervisions for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5513–5522, 2017

    2017

  20. [27]

    Multi-label image recognition with graph con- volutional networks

    Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph con- volutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5177–5186, 2019

    2019

  21. [28]

    Attentionxml: Label tree-based attention-aware deep model for high-performance extreme multi-label text classification

    Ronghui You, Zihan Zhang, Ziye Wang, Suyang Dai, Hiroshi Mamitsuka, and Shanfeng Zhu. Attentionxml: Label tree-based attention-aware deep model for high-performance extreme multi-label text classification. Ad- vances in neural information processing systems, 32, 2019

    2019

  22. [29]

    Use All The Labels: A Hierarchical Multi-Label Contrastive Learning Framework

    Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use All The Labels: A Hierarchical Multi-Label Contrastive Learning Framework. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16639–16648, New Orleans, LA, USA, June 2022. IEEE

    2022

  23. [30]

    Hex: Hierarchical emergence exploitation in self-supervised algorithms

    Kiran Kokilepersaud, Seulgi Kim, Mohit Prabhushankar, and Ghassan AlRegib. Hex: Hierarchical emergence exploitation in self-supervised algorithms. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1111–1121. IEEE, 2025

    2025

  24. [31]

    Hierarchical multi-granular multi-label contrastive learning

    Haixiang Li, Min Fang, Xiao Li, Bo Chen, and Guizhi Wang. Hierarchical multi-granular multi-label contrastive learning. Pattern Recognition, 164:111567, August 2025

    2025

  25. [32]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, and others. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the ...

    2019

  26. [33]

    Mimic-cxr database

    Alistair Johnson, Tom Pollard, Roger Mark, Seth Berkowitz, and Steven Horng. Mimic-cxr database. Phys- ioNet10, 13026:C2JT1Q, 2024

    2024

  27. [34]

    Imagegcn: Multi-relational image graph convolutional networks for disease identification with chest x-rays

    Chengsheng Mao, Liang Yao, and Yuan Luo. Imagegcn: Multi-relational image graph convolutional networks for disease identification with chest x-rays. IEEE transactions on medical imaging , 41(8):1990–2003, 2022. Publisher: IEEE

    1990

  28. [35]

    Scalp- supervised contrastive learning for cardiopulmonary disease classification and localization in chest x-rays using patient metadata

    Ajay Jaiswal, Tianhao Li, Cyprian Zander, Yan Han, Justin F Rousseau, Yifan Peng, and Ying Ding. Scalp- supervised contrastive learning for cardiopulmonary disease classification and localization in chest x-rays using patient metadata. In 2021 IEEE International Conference on ...

    2021

  29. [36]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  30. [37]

    Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673, 2020

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673, 2020. 9 E. Takaya and R. Inamori

    2020

  31. [39]

    An open access thyroid ultrasound image database

    Lina Pedraza, Carlos Vargas, Fabi ´an Narv ´aez, Oscar Dur ´an, Emma Mu ˜noz, and Eduardo Romero. An open access thyroid ultrasound image database. In 10th International symposium on medical information processing and analysis, volume 9287, pages 188–193. SPIE, 2015

    2015

  32. [40]

    Dataset of breast ultrasound images

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. Publisher: Elsevier

    2020

  33. [41]

    Deep-learning-assisted diagnosis for knee mag- netic resonance imaging: development and retrospective validation of MRNet.PLoS medicine, 15(11):e1002699,

    Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, and others. Deep-learning-assisted diagnosis for knee mag- netic resonance imaging: development and retrospective validatio...

  34. [42]

    Umap: Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 10

    2018 arXiv

  35. [2018]

    Publisher: Public Library of Science San Francisco, CA USA

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.