Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Chest X-ray Foundation Model with Global and Local Representations Integration

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

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

Pith's one-line read CheXFound, a self-supervised ViT-Large model pretrained on 987,733 chest X-rays, claims to outperform previous CXR foundation models on classifying 40 disease findings and to generalize to new tasks like mortality prediction.

desk verdict A worthwhile, well-built CXR encoder with a genuinely useful head, but the headline SOTA claim is over-strong because resolution and architecture are not fully controlled. read the letter →

arxiv 2502.05142 v2 pith:AKWAT3PU submitted 2025-02-07 eess.IV cs.CV

Zefan Yang , Xuanang Xu , Jiajin Zhang , Ge Wang , Mannudeep K. Kalra , Pingkun Yan This is my paper · ORCID
classification eess.IVcs.CV
keywords chestx-rayfoundationmodelself-supervisedlearningDINOv2multilabelclassificationlong-tailedglobal-localrepresentationstransfer
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

CheXFound is a vision foundation model for chest X-rays built by pretraining a ViT-Large encoder with DINOv2-style self-distillation on roughly 987,000 unique X-rays drawn from 12 public datasets. The paper's central claim is that this combination of a larger model and broader pretraining data produces representations that transfer better than previous CXR foundation models, including RAD-DINO, EVA-X, and CheXzero. On the long-tailed CXR-LT 24 benchmark of 40 disease findings, CheXFound reaches an AUROC of 0.840 when paired with a new prediction head, GLoRI, and it also leads on datasets with little training data, such as Shenzhen, Montgomery, and JSRT. The paper further reports gains on out-of-distribution tasks: cardiovascular-disease risk and all-cause mortality prediction on PLCO, malpositioned tube detection, and rib and lung segmentation. The reason this matters is that a single pretrained CXR encoder with good label efficiency could replace task-specific models across many clinical uses.

What carries the argument

The carrying mechanisms are two. First, DINOv2-style self-distillation pretraining: a teacher-student pair of ViT-Large networks trained with two losses, a [CLS]-token alignment loss and a masked image modeling loss, over 987,733 chest X-rays from 12 sources, with a low-resolution warm-up so the model can be pretrained at 512x512 resolution. Second, the GLoRI head, which runs on frozen patch tokens: it projects patch tokens, initializes one disease query per finding, and uses multi-head cross-attention whose temperature is adapted per disease to sharpen or smooth attention, plus a pyramid patch merging module that pools 8x8, 4x4, and 2x2 neighborhoods into multi-scale coarse features; the resulting fine and coarse local features are concatenated with the global [CLS] token before a linear classifier. GLoRI's contribution is showing that these local features carry disease-specific signal that a pure [CLS]-token linear probe misses.

What would settle it

Retrain a comparison baseline such as RAD-DINO or Medical MAE on CXR-987K with a ViT-Large backbone, the same low-resolution warm-up, the same masked-image-modeling loss weight of 3, and the same GLoRI head; if such a model reaches or exceeds 0.840 AUROC on CXR-LT 24, the paper's conclusion that its pretraining approach is superior to prior methods would be falsified, leaving architecture and tuning as the real cause.

Watch

Extended reading notes

Core claim

The central discovery is that scaling a self-distilled vision transformer to large capacity and pretraining it on a large, multi-source corpus of chest X-rays yields a frozen encoder whose linear-probe features already beat prior CXR foundation models, and that adding local patch-level information to the global [CLS] token further improves multilabel classification. On CXR-LT 24, CheXFound with the GLoRI head achieves AUROC 0.840 and AUPRC 0.265, outperforming the next best method by clear margins, and it outperforms comparison models across all 40 disease findings, including low-prevalence ones. The same frozen backbone also improves out-of-distribution tasks: AUROC 0.749 for CVD risk and 0.786 for all-cause mortality on PLCO, better tube-position classification on RANZCR-CLiP, and higher Dice on rib segmentation. The paper attributes these gains to three ingredients: large-scale diverse pretraining data, higher pretraining resolution with a low-resolution warm-up, and an increased masked-image-modeling loss weight tuned for medical images.

Load-bearing premise

The claim that CheXFound's pretraining recipe beats prior methods rests on the comparison in Section V-E being fair, but the baseline models there use smaller backbones and were not tuned in the same way, so part of the headline advantage could come from model size and tuning rather than the method.

Editorial extensions

If this is right

  • A single frozen CheXFound encoder can serve many downstream CXR tasks without per-task backbone fine-tuning, since GLoRI heads trained on top of the frozen features already outperform end-to-end fine-tuned models.
  • On long-tail and out-of-distribution benchmarks, larger-scale self-supervised pretraining on diverse X-ray sources appears to improve classification of rare and unseen-in-training findings, which matters for clinical deployment.
  • The GLoRI head's disease queries produce attention maps that localize abnormalities, offering interpretability alongside classification.
  • Higher pretraining resolution (up to 512x512) and a heavier masked-image-modeling weight yield measurable AUROC gains across all evaluated datasets, so resolution and loss balance are load-bearing design choices, not incidental details.
  • CheXFound's success on opportunistic tasks suggests routine CXRs could be reused for CVD risk estimation and mortality screening without new imaging.

Reading between the lines

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

  • If confirmed by a controlled comparison where every baseline uses the same ViT-Large backbone and the same resolution warm-up, the paper's recipe would suggest that data diversity and resolution matter more than the specific self-supervised loss family for chest X-ray representations.
  • The GLoRI head's attention maps, which the paper shows align with disease regions but only partially cover abnormality areas, could be turned into weakly supervised localization or segmentation seeds.
  • The authors did not run a ViT-giant variant, and they note it would likely improve further but require more data and compute; the scaling table suggests returns continue beyond 987K images, so larger-scale CXR pretraining is a plausible next step.
  • Because CheXFound was pretrained on both frontal and lateral views while many evaluations use frontal-only inputs, a multiview-aware adaptation head might recover additional signal for tasks like CVD risk, where prior work exploits multi-view information.
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 / 4 minor

Summary. This paper presents CheXFound, a ViT-Large chest X-ray foundation model pretrained with a DINOv2-style self-distillation objective on 987,733 unique CXRs curated from 12 public datasets. The authors add a low-resolution warm-up schedule and an increased masked image modeling loss weight relative to the original DINOv2 recipe. For downstream tasks, they propose GLoRI, an attention-based head that combines fine-grained and coarse-grained local patch features with the global [CLS] token. The paper reports linear-probe and GLoRI results across in-distribution (CXR-LT 24, CheXpert) and out-of-distribution datasets (VinDr-CXR, CXR-Pneumonia, Shenzhen, Montgomery, JSRT, RANZCR-CLiP, PLCO), plus segmentation on VinDr-RibCXR and Montgomery. CheXFound with GLoRI achieves the best AUROC on CXR-LT 24 (0.840) and shows gains on CVD risk and mortality prediction.

Significance. Assuming the results are reproducible, this is a strong empirical contribution: it is among the largest self-supervised pretraining efforts for CXR, with careful test-set contamination control (only training splits of CXR-LT 24 and CheXpert were used for pretraining), bootstrapped 95% CIs and permutation tests throughout, and a public code release. The GLoRI head provides a modest but consistent improvement over linear probing and attention pooling across several tasks, and the ablation of resolution, data scale, and model scale is informative. The main uncertainty is whether the headline gains over released baselines are due to the proposed representation quality or to uncontrolled differences in pretraining resolution, model capacity, and hyperparameter selection.

major comments (4)
  1. [Section V-A, Tables III/IV] The paper never states the input resolution used for downstream evaluation of the released baselines. CheXFound is pretrained at up to 512x512 (Section IV-A.1 and Fig. 6), while RAD-DINO, EVA-X, and CheXzero are public checkpoints typically pretrained at 224x224. Fig. 6 shows that increasing CheXFound's pretraining resolution from 224 to 512 alone improves CXR-LT 24 AUROC by 3.3% (p<0.001), which is a substantial fraction of the reported 11.7% gap over RAD-DINO under GLoRI (0.840 vs 0.723). If all models were evaluated at 512, the baselines may be handicapped by positional-embedding interpolation; if each model was evaluated at its native resolution, CheXFound received more input pixels. Please report the downstream resolution for every model and provide a matched-resolution comparison (e.g., all models at 224 and 512) or explicitly discount the resolution component of the headline advantage.
  2. [Section V-E, Table IX] The pretraining-method benchmark confounds method and architecture for three of the four baselines: Medical MAE, EVA-X, and RAD-DINO use ViT-B (86M), while CheXFound uses ViT-L (307M). The comparison to iBOT is architecture-controlled (both ViT-L) and supports the claim about the low-resolution warm-up and MIM loss weight, but the larger gains over Medical MAE, EVA-X, and RAD-DINO cannot be attributed solely to the proposed method. Please include ViT-L runs of these baselines on CXR-987K, or restrict the headline method claim to the iBOT comparison and clearly label the others as architecture+method comparisons.
  3. [Section V-H and Fig. 7] The values of the MIM loss weight (3) and the low-resolution warm-up were selected by studying AUROC on CXR-LT 24, the same benchmark used for the headline comparison in Tables III-V. This is selection on the evaluation set (or at least on the same task), so the 4.7% gain over 'original DINOv2 settings' is in-sample. Please validate the chosen hyperparameters on at least one task not used for selection, or use a nested validation scheme, to demonstrate that the tuning did not inflate the main results.
  4. [Section IV.A.2 and Tables IV/V] The abstract and Section VI.B claim 'superior label efficiency' and 'generalization to limited training data,' but the supporting experiments are only full-data trainings on small datasets (Shenzhen, Montgomery, JSRT). No controlled low-label protocol (e.g., subsampling the training set to 1%, 10%, 50%) is presented for CheXFound or the baselines, so the label-efficiency claim is not directly substantiated. Either add such experiments or revise the claim.
minor comments (4)
  1. [Table III] The CheXpert AUPRC for CheXFound linear probe is reported as 0.620(.630-.727), an internally inconsistent interval; the point estimate likely should be 0.679. The RAD-DINO linear-probe AUROC row also shows 0.557(.570-.583), which has the same issue.
  2. [Abstract and Section I] The abstract says 'CXR-1M' and 'over one million unique CXRs,' while the body consistently reports CXR-987K with 987,733 images; please make the numbers consistent.
  3. [Section V-A, V-E, IV.A.2] There are several typos: 'CXT-LT 24' (Section V-A), 'benckmarked' (Section V-E), and 'Mongomery' (Section IV.A.2) should be corrected.
  4. [Fig. 6] The caption does not specify whether downstream evaluation uses the same input resolution as pretraining for each setting; please clarify this methodological detail in the caption or text.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity: CheXFound is empirical model-building with externally benchmarked comparisons; only a non-load-bearing self-citation and a selection-on-benchmark caveat keep the score at 2.

full rationale

The paper's central claims are empirical measurements on held-out test splits of CXR-LT 24, CheXpert, VinDr-CXR, Shenzhen, Montgomery, JSRT, RANZCR-CLiP, PLCO, and VinDr-RibCXR, compared against released checkpoints of RAD-DINO, EVA-X, CheXzero, BiomedCLIP, PubMedCLIP, and ConvNeXt. No equation defines a predicted quantity in terms of the measured outcome, so there is no self-definitional or constructional circularity. GLoRI is a supervised attention head trained with binary cross-entropy loss; its attention maps are learned from the same disease labels, which is standard fitting rather than a circular derivation. The pretraining hyperparameters (MIM loss weight 3 and low-resolution warm-up) were selected using CXR-LT 24 downstream AUROC in Fig. 7, so the headline CXR-LT 24 number is not fully independent evidence for the pretraining recipe; this is benchmark selection rather than constructional circularity, and the paper does disclose that only the CXR-LT 24 training set was used for pretraining. The manuscript itself acknowledges limitations: attention maps cover only partial abnormality regions (Sec. VI-A) and ViT-g was not evaluated (Sec. VI-B). There is one self-citation (ref. [3], the authors' Bi-Mamba CVD paper) in the introduction as an example task, but it is not load-bearing for any result. A citation-consistency error exists in Table VI, where EVA-X and RAD-DINO are cited as [20] instead of [11]/[12], but this does not affect the derivation. Overall, the comparisons are externally benchmarked and the central claims retain independent content, so no significant circularity is found.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central contribution is empirical; no new physical entities are introduced. Free parameters are standard hyperparameters, with the MIM loss weight and resolution schedule explicitly tuned on the CXR-LT 24 benchmark. The key assumptions are data quality and deduplication, and the transferability of self-supervised features.

free parameters (4)
  • MIM loss weight = 3
    Selected via ablations on CXR-LT 24 downstream AUROC (Section V-H); the original DINOv2 uses a weight of 1.
  • Low-resolution warm-up schedule = 224^2 to 512^2
    Introduced because direct high-resolution pretraining failed (Section V-G), tuned on downstream performance.
  • GLoRI embedding dimension = 768
    Chosen as a hyperparameter for the attention head (Section IV-A2).
  • Number of disease queries = M (number of findings)
    Set to the number of diseases in each downstream task; a design choice, not a fitted constant.
assumptions (3)
  • domain assumption DINOv2 self-supervised pretraining produces transferable representations for medical images.
    The entire method assumes that self-distillation and masked image modeling on CXRs learn features that transfer to downstream tasks; this is supported by prior work but not proven in the paper.
  • domain assumption The CXR-987K dataset is deduplicated and contains 987,733 unique images.
    The paper calls the images 'unique' but does not describe a deduplication procedure across the 12 public sources, which may contain overlapping patients.
  • domain assumption Labels in public downstream datasets (CXR-LT 24, CheXpert, VinDr-CXR, etc.) are sufficiently accurate to support the performance comparisons.
    The evaluation uses these labels as ground truth; label noise is known in CXR datasets and could affect small test sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chest X-ray Foundation Model with Global and Local Representations Integration." pith.science (2026). https://pith.science/paper/AKWAT3PU

@misc{pith2026250205142,
  author       = {Pith},
  title        = {Pith review of: Chest X-ray Foundation Model with Global and Local Representations Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKWAT3PU}},
  note         = {Machine review of arXiv:2502.05142}
}
read the original abstract

Chest X-ray (CXR) is the most frequently ordered imaging test, supporting diverse clinical tasks from thoracic disease detection to postoperative monitoring. However, task-specific classification models are limited in scope, require costly labeled data, and lack generalizability to out-of-distribution datasets. To address these challenges, we introduce CheXFound, a self-supervised vision foundation model that learns robust CXR representations and generalizes effectively across a wide range of downstream tasks. We pretrain CheXFound on a curated CXR-1M dataset, comprising over one million unique CXRs from publicly available sources. We propose a Global and Local Representations Integration (GLoRI) module for downstream adaptations, by incorporating disease-specific local features with global image features for enhanced performance in multilabel classification. Our experimental results show that CheXFound outperforms state-of-the-art models in classifying 40 disease findings across different prevalence levels on the CXR-LT 24 dataset and exhibits superior label efficiency on downstream tasks with limited training data. Additionally, CheXFound achieved significant improvements on new tasks with out-of-distribution datasets, including opportunistic cardiovascular disease risk estimation and mortality prediction. These results highlight CheXFound's strong generalization capabilities, enabling diverse adaptations with improved label efficiency. The project source code is publicly available at https://github.com/RPIDIAL/CheXFound.

Figures

Figures reproduced from arXiv: 2502.05142 by the authors.

Figure 1
Figure 1. Overview of self-supervised pretraining of CheXFound, using publicly available CXRs from multiple institutions with a masked image modeling objective and a [CLS] token alignment objective. 2) Encoding Coarse-grained Local Features with Pyramid Patch Merging: To effectively summarize contextual infor￾mation, we incorporate a pyramid patch merging module to integrate multi-scale patch embeddings at low resolutions to … view at source ↗
Figure 2
Figure 2. Global and Local Representations Integration (GLoRI) for evaluating CheXFound on downstream tasks. GLoRI is trained on top of the frozen CheXFound backbone. GLoRI incorporates adaptive temperatures and pyramid patch merging into the attention mechanism to encode fine-grained and coarse-grained local features respectively and integrates the global image features from the [CLS] token to predict final multilabel classi… view at source ↗
Figure 3
Figure 3. Detailed Performance for 40 disease findings in AUROC on the CXR-LT 24 dataset. Our CheXFound is compared with the vision￾centric foundation models (EVA-X and RAD-DINO), the vision-language pretrained foundation models (CheXzero, BiomedCLIP, and PubMed￾CLIP), and the end-to-end trained model (ConvNeXt) with ImageNet￾22K pretraining. the linear probe performance across seven datasets (CXR￾LT 24, CheXpert, VinDr-CXR, … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Model performance under high, medium and low disease prevalence. a, The number of labels for the 40 disease findings on the CXR-LT 24 dataset [21]. b, Model performance in AUPRC stratified by high, medium and low disease prevalence. Error bars indicate the 95% confiden…
Figure 5
Figure 5. Figure 5: Kaplan-Meier curves for low-risk and high-risk groups of all￾cause mortality on the PLCO dataset. The number of subjects in the test set is 10,509. The survival rates for low-risk and high-risk groups are significantly different (p<0.0001, log-rank test). Shaded areas …
Figure 6
Figure 6. Figure 6: Evaluation of CheXFound model on the CXR-LT 24, CheXpert, Shenzhen, Montgomery, and JSRT datasets across a range of pretrain￾ing image resolutions. Error bars indicate 95% confidence intervals. data from CXR-207K to CXR-987K with the ViT-L backbone leads to significant…
Figure 7
Figure 7. Figure 7: Ablation studies on the impact of MIM loss weights and param￾eter initialization strategies on downstream classification. We evaluated the classification performance on the CXR-LT 24 dataset. Error bars indicate 95% confidence intervals. TABLE XI ABLATION STUDIES ON TH…
Figure 8
Figure 8. Figure 8: Detailed results comparing GLoRI and linear probing for classifying different diseases. Diseases are sorted in descending order regarding their frequency on the dataset. We computed the mean and 95% confidence intervals of average precision using 1,000 bootstrapped sam…
Figure 9
Figure 9. Figure 9: GLoRI attention maps for disease findings of atelectasis, cardiomegaly, consolidation, edema, and pleural effusion, respectively. Each subfigure contains 25 CXRs with their GLoRI attention maps overlaid (bottom) and an anchor CXR with a global attention map overlaid (t…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Xray2Xray: World Model from Chest X-rays with Volumetric Context

    eess.IV 2025-06 conditional novelty 6.0 of 10

    Xray2Xray, a world-model-style transformer, generates a sequence of predicted chest X-ray projections from one or two input views, and these latent predictions improve downstream disease prediction and enable tomograp...

Reference graph

Works this paper leans on

58 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Interpretation of plain chest roentgenogram,

    S. Raoof, D. Feigin, A. Sung, S. Raoof, L. Irugulpati, and E. C. Rosenow III, “Interpretation of plain chest roentgenogram,”Chest, vol. 141, no. 2, pp. 545–558, 2012

    work page 2012

  2. [2]

    Deep learning to estimate cardiovascular risk from chest radiographs,

    J. Weiss, V . K. Raghu, K. Paruchuri, A. Zinzuwadia, P. Natarajan, H. J. Aerts, and M. T. Lu, “Deep learning to estimate cardiovascular risk from chest radiographs,”Annals of Internal Medicine, vol. 177, no. 4, pp. 409–417, 2024

    work page 2024

  3. [3]

    Cardiovascular disease detection from multi-view chest x-rays with bi-mamba,

    Z. Yang, J. Zhang, G. Wang, M. K. Kalra, and P. Yan, “Cardiovascular disease detection from multi-view chest x-rays with bi-mamba,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 134–144

    work page 2024

  4. [4]

    Deep learning to estimate lung disease mortality from chest radiographs,

    J. Weiss, V . K. Raghu, D. Bontempi, D. C. Christiani, R. H. Mak, M. T. Lu, and H. J. Aerts, “Deep learning to estimate lung disease mortality from chest radiographs,”Nature Communications, vol. 14, no. 1, p. 2797, 2023

    work page 2023

  5. [5]

    Opportunistic detection of type 2 diabetes using deep learning from frontal chest radiographs,

    A. Pyrros, S. M. Borstelmann, R. Mantravadi, Z. Zaiman, K. Thomas, B. Price, E. Greenstein, N. Siddiqui, M. Willis, I. Shulhanet al., “Opportunistic detection of type 2 diabetes using deep learning from frontal chest radiographs,”Nature communications, vol. 14, no. 1, p. 4039, 2023

    work page 2023

  6. [6]

    An empirical study of training self- supervised vision transformers,

    X. Chen, S. Xie, and K. He, “An empirical study of training self- supervised vision transformers,” inProceedings of the IEEE/CVF in- ternational conference on computer vision, 2021, pp. 9640–9649

    2021

  7. [7]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 000–16 009

    2022

  8. [8]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

Show all 58 references
  1. [9]

    ibot: Image bert pre-training with online tokenizer,

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong, “ibot: Image bert pre-training with online tokenizer,”arXiv preprint 13 arXiv:2111.07832, 2021

    2021 arXiv

  2. [10]

    Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging,

    S. Azizi, L. Culp, J. Freyberg, B. Mustafa, S. Baur, S. Kornblith, T. Chen, N. Tomasev, J. Mitrovi ´c, P. Strachanet al., “Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging,”Nature Biomedical Engineering, vol. 7, no. 6, pp. 756...

    2023

  3. [11]

    Eva-x: A foundation model for general chest x-ray analysis with self- supervised learning,

    J. Yao, X. Wang, Y . Song, H. Zhao, J. Ma, Y . Chen, W. Liu, and B. Wang, “Eva-x: A foundation model for general chest x-ray analysis with self- supervised learning,”arXiv preprint arXiv:2405.05237, 2024

    2024 arXiv

  4. [12]

    Exploring scalable medical image encoders beyond text supervision,

    F. P ´erez-Garc´ıa, H. Sharma, S. Bond-Taylor, K. Bouzid, V . Salvatelli, M. Ilse, S. Bannur, D. C. Castro, A. Schwaighofer, M. P. Lungrenet al., “Exploring scalable medical image encoders beyond text supervision,” Nature Machine Intelligence, pp. 1–12, 2025

    2025

  5. [13]

    Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning,

    E. Tiu, E. Talius, P. Patel, C. P. Langlotz, A. Y . Ng, and P. Rajpurkar, “Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning,”Nature Biomedical Engineering, vol. 6, no. 12, pp. 1399–1406, 2022

    2022

  6. [14]

    Mimic-cxr, a de- identified publicly available database of chest radiographs with free-text reports,

    A. E. Johnson, T. J. Pollard, S. J. Berkowitz, N. R. Greenbaum, M. P. Lungren, C.-y. Deng, R. G. Mark, and S. Horng, “Mimic-cxr, a de- identified publicly available database of chest radiographs with free-text reports,”Scientific data, vol. 6, no. 1, p. 317, 2019

    2019

  7. [15]

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

    J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Mark- lund, B. Haghgoo, R. Ball, K. Shpanskayaet al., “Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison,” inProceedings of the AAAI conference on artificial intelligenc...

    2019

  8. [16]

    Padchest: A large chest x-ray image dataset with multi-label annotated reports,

    A. Bustos, A. Pertusa, J.-M. Salinas, and M. De La Iglesia-Vaya, “Padchest: A large chest x-ray image dataset with multi-label annotated reports,”Medical image analysis, vol. 66, p. 101797, 2020

    2020

  9. [17]

    Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,

    X. Wang, Y . Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,” inProceedings of the IEEE conference on computer vision and pattern...

    2017

  10. [18]

    Brax, brazilian labeled chest x-ray dataset,

    E. P. Reis, J. P. De Paiva, M. C. Da Silva, G. A. Ribeiro, V . F. Paiva, L. Bulgarelli, H. M. Lee, P. V . Santos, V . M. Brito, L. T. Amaralet al., “Brax, brazilian labeled chest x-ray dataset,”Scientific Data, vol. 9, no. 1, p. 487, 2022

    2022

  11. [19]

    Curation of the candid-ptx dataset with free-text reports,

    S. Feng, D. Azzollini, J. S. Kim, C.-K. Jin, S. P. Gordon, J. Yeoh, E. Kim, M. Han, A. Lee, A. Patelet al., “Curation of the candid-ptx dataset with free-text reports,”Radiology: Artificial Intelligence, vol. 3, no. 6, p. e210136, 2021

    2021

  12. [20]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  13. [21]

    CXR-LT 2024: Long-tailed, multi-label, and zero- shot classification on chest X-rays,

    Y . Peng, M. Lin, G. Holste, S. Wang, Y . Zhou, H. Chen, A. Wang, A. E. Flanders, L. A. Celi, Z. Lu, G. Shih, and R. M. Summers, “CXR-LT 2024: Long-tailed, multi-label, and zero- shot classification on chest X-rays,” Apr. 2024. [Online]. Available: https://doi.org/10.5281/zeno...

    2024 doi

  14. [22]

    Lung cancer screening in the randomized prostate, lung, colorectal, and ovarian (plco) cancer screening trial,

    W. G. Hocking, P. Hu, M. M. Oken, S. D. Winslow, P. A. Kvale, P. C. Prorok, L. R. Ragard, J. Commins, D. A. Lynch, G. L. Andrioleet al., “Lung cancer screening in the randomized prostate, lung, colorectal, and ovarian (plco) cancer screening trial,”JNCI: Journal of the Nationa...

    2010

  15. [23]

    Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations,

    H. Q. Nguyen, K. Lam, L. T. Le, H. H. Pham, D. Q. Tran, D. B. Nguyen, D. D. Le, C. M. Pham, H. T. Tong, D. H. Dinhet al., “Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations,”Scientific Data, vol. 9, no. 1, p. 429, 2022

    2022

  16. [24]

    Ranzcr clip - catheter and line position challenge,

    J. Seah, Jen, Maggie, M. Law, P. Culliton, and S. Dowd, “Ranzcr clip - catheter and line position challenge,” https://kaggle.com/competitions/ ranzcr-clip-catheter-line-classification, 2020, kaggle

    2020

  17. [25]

    Vindr-ribcxr: A benchmark dataset for automatic segmentation and labeling of individual ribs on chest x-rays,

    H. C. Nguyen, T. T. Le, H. H. Pham, and H. Q. Nguyen, “Vindr-ribcxr: A benchmark dataset for automatic segmentation and labeling of individual ribs on chest x-rays,”arXiv preprint arXiv:2107.01327, 2021

    2021 arXiv

  18. [26]

    Beit: Bert pre-training of image transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “Beit: Bert pre-training of image transformers,”arXiv preprint arXiv:2106.08254, 2021

    2021 arXiv

  19. [27]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inInternational conference on machine learning. PMLR, 2020, pp. 1597–1607

    2020

  20. [28]

    Towards a general-purpose foundation model for computational pathology,

    R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, A. H. Song, B. Chen, A. Zhang, D. Shao, M. Shabanet al., “Towards a general-purpose foundation model for computational pathology,”Nature Medicine, vol. 30, no. 3, pp. 850–862, 2024

    2024

  21. [29]

    Bootstrap your own latent-a new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azaret al., “Bootstrap your own latent-a new approach to self-supervised learning,” Advances in neural information processing systems, vol. 33, pp. 21...

    2020

  22. [30]

    Chexnet: Radiologist- level pneumonia detection on chest x-rays with deep learning,

    P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskayaet al., “Chexnet: Radiologist- level pneumonia detection on chest x-rays with deep learning,”arXiv preprint arXiv:1711.05225, 2017

    2017 arXiv

  23. [31]

    Chex- transfer: performance and parameter efficiency of imagenet models for chest x-ray interpretation,

    A. Ke, W. Ellsworth, O. Banerjee, A. Y . Ng, and P. Rajpurkar, “Chex- transfer: performance and parameter efficiency of imagenet models for chest x-ray interpretation,” inProceedings of the conference on health, inference, and learning, 2021, pp. 116–124

    2021

  24. [32]

    Medaug: Contrastive learning leveraging patient metadata improves representations for chest x-ray interpretation,

    Y . N. T. Vu, R. Wang, N. Balachandar, C. Liu, A. Y . Ng, and P. Rajpurkar, “Medaug: Contrastive learning leveraging patient metadata improves representations for chest x-ray interpretation,” inMachine Learning for Healthcare Conference. PMLR, 2021, pp. 755–769

    2021

  25. [33]

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

    H. Sowrirajan, J. Yang, A. Y . Ng, and P. Rajpurkar, “Moco pretraining improves representation and transferability of chest x-ray models,” in Medical Imaging with Deep Learning. PMLR, 2021, pp. 728–744

    2021

  26. [34]

    Advancing human-centric ai for robust x-ray analysis through holistic self-supervised learning,

    T. Moutakanni, P. Bojanowski, G. Chassagnon, C. Hudelot, A. Joulin, Y . LeCun, M. Muckley, M. Oquab, M.-P. Revel, and M. Vakalopoulou, “Advancing human-centric ai for robust x-ray analysis through holistic self-supervised learning,”arXiv preprint arXiv:2405.01469, 2024

    2024 arXiv

  27. [35]

    Delving into masked autoen- coders for multi-label thorax disease classification,

    J. Xiao, Y . Bai, A. Yuille, and Z. Zhou, “Delving into masked autoen- coders for multi-label thorax disease classification,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 3588–3600

    2023

  28. [36]

    Foundation ark: Accruing and reusing knowledge for superior and robust performance,

    D. Ma, J. Pang, M. B. Gotway, and J. Liang, “Foundation ark: Accruing and reusing knowledge for superior and robust performance,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 651–662

    2023

  29. [37]

    Towards foundation models learned from anatomy in medical imaging via self- supervision,

    M. R. Hosseinzadeh Taher, M. B. Gotway, and J. Liang, “Towards foundation models learned from anatomy in medical imaging via self- supervision,” inMICCAI Workshop on Domain Adaptation and Repre- sentation Transfer. Springer, 2023, pp. 94–104

    2023

  30. [38]

    Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs,

    S. Zhang, Y . Xu, N. Usuyama, H. Xu, J. Bagga, R. Tinn, S. Preston, R. Rao, M. Wei, N. Valluriet al., “Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs,”arXiv preprint arXiv:2303.00915, 2023

    2023 arXiv

  31. [39]

    Does clip benefit visual question answering in the medical domain as much as it does in the general domain?

    S. Eslami, G. de Melo, and C. Meinel, “Does clip benefit visual question answering in the medical domain as much as it does in the general domain?”arXiv preprint arXiv:2112.13906, 2021

    2021 arXiv

  32. [40]

    Chexagent: Towards a foundation model for chest x-ray interpretation,

    Z. Chen, M. Varma, J.-B. Delbrouck, M. Paschali, L. Blankemeier, D. Van Veen, J. M. J. Valanarasu, A. Youssef, J. P. Cohen, E. P. Reiset al., “Chexagent: Towards a foundation model for chest x-ray interpretation,”arXiv preprint arXiv:2401.12208, 2024

    2024 arXiv

  33. [41]

    Llava-med: Training a large language-and-vision assistant for biomedicine in one day,

    C. Li, C. Wong, S. Zhang, N. Usuyama, H. Liu, J. Yang, T. Naumann, H. Poon, and J. Gao, “Llava-med: Training a large language-and-vision assistant for biomedicine in one day,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  34. [42]

    Advancing multi- modal medical capabilities of gemini,

    L. Yang, S. Xu, A. Sellergren, T. Kohlberger, Y . Zhou, I. Ktena, A. Ki- raly, F. Ahmed, F. Hormozdiari, T. Jaroensriet al., “Advancing multi- modal medical capabilities of gemini,”arXiv preprint arXiv:2405.03162, 2024

    2024 arXiv

  35. [43]

    Siim-acr pneumothorax seg- mentation,

    A. Zawacki, C. Wu, G. Shih, J. Elliott, M. Fomitchev, M. Hussain, ParasLakhani, P. Culliton, and S. Bao, “Siim-acr pneumothorax seg- mentation,” 2019

    2019

  36. [44]

    Object-cxr - automatic detection of foreign objects on chest x-rays,

    “Object-cxr - automatic detection of foreign objects on chest x-rays,” JF Healthcare, 2020

    2020

  37. [45]

    The 2021 siim-fisabio-rsna machine learning covid-19 challenge: An- notation and standard exam classification of covid-19 chest radiographs,

    P. Lakhani, J. Mongan, C. Singhal, Q. Zhou, K. P. Andriole, W. F. Auffermann, P. Prasanna, T. X. Pham, M. Peterson, P. J. Bergquistet al., “The 2021 siim-fisabio-rsna machine learning covid-19 challenge: An- notation and standard exam classification of covid-19 chest radiograp...

    2021

  38. [46]

    Covidx cxr-4: An expanded multi-institutional open-source benchmark dataset for chest x- ray image-based computer-aided covid-19 diagnostics,

    Y . Wu, H. Gunraj, C.-e. A. Tai, and A. Wong, “Covidx cxr-4: An expanded multi-institutional open-source benchmark dataset for chest x- ray image-based computer-aided covid-19 diagnostics,”arXiv preprint arXiv:2311.17677, 2023

    2023 arXiv

  39. [47]

    Midrc covidx challenge,

    “Midrc covidx challenge,” Medical Imaging and Data Resource Center, 2022

    2022

  40. [48]

    Bimcv covid-19+: a large annotated dataset of rx and ct images from covid-19 patients,

    M. D. L. I. Vay ´a, J. M. Saborit, J. A. Montell, A. Pertusa, A. Bustos, M. Cazorla, J. Galant, X. Barber, D. Orozco-Beltr ´an, F. Garc ´ıa-Garc´ıa et al., “Bimcv covid-19+: a large annotated dataset of rx and ct images from covid-19 patients,”arXiv preprint arXiv:2006.01174, 2020

    2006 arXiv

  41. [49]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763. 14

    2021

  42. [50]

    Revisiting feature prediction for learning visual representations from video,

    A. Bardes, Q. Garrido, J. Ponce, X. Chen, M. Rabbat, Y . LeCun, M. Assran, and N. Ballas, “Revisiting feature prediction for learning visual representations from video,”arXiv preprint arXiv:2404.08471, 2024

    2024 arXiv

  43. [51]

    Gloria: A multimodal global-local representation learning framework for label- efficient medical image recognition,

    S.-C. Huang, L. Shen, M. P. Lungren, and S. Yeung, “Gloria: A multimodal global-local representation learning framework for label- efficient medical image recognition,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3942–3951

    2021

  44. [52]

    Identifying medical diagnoses and treatable diseases by image-based deep learning,

    D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, F. Yanet al., “Identifying medical diagnoses and treatable diseases by image-based deep learning,” cell, vol. 172, no. 5, pp. 1122–1131, 2018

    2018

  45. [53]

    Two public chest x-ray datasets for computer-aided screen- ing of pulmonary diseases,

    S. Jaeger, S. Candemir, S. Antani, Y .-X. J. W ´ang, P.-X. Lu, and G. Thoma, “Two public chest x-ray datasets for computer-aided screen- ing of pulmonary diseases,”Quantitative imaging in medicine and surgery, vol. 4, no. 6, p. 475, 2014

    2014

  46. [54]

    Shiraishi, S

    J. Shiraishi, S. Katsuragawa, J. Ikezoe, T. Matsumoto, T. Kobayashi, K.- i. Komatsu, M. Matsui, H. Fujita, Y . Kodera, and K. Doi, “Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radio...

    2000

  47. [55]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 976–11 986

    2022

  48. [56]

    Eva: Exploring the limits of masked visual representation learning at scale,

    Y . Fang, W. Wang, B. Xie, Q. Sun, L. Wu, X. Wang, T. Huang, X. Wang, and Y . Cao, “Eva: Exploring the limits of masked visual representation learning at scale,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 358–19 369

    2023

  49. [57]

    Screening by chest radiograph and lung cancer mortality: the prostate, lung, colorectal, and ovarian (plco) randomized trial,

    M. M. Oken, W. G. Hocking, P. A. Kvale, G. L. Andriole, S. S. Buys, T. R. Church, E. D. Crawford, M. N. Fouad, C. Isaacs, D. J. Reding et al., “Screening by chest radiograph and lung cancer mortality: the prostate, lung, colorectal, and ovarian (plco) randomized trial,”Jama, v...

    2011

  50. [58]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 418–434

    2018

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.