Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

A Generative Foundation Model for Chest Radiography

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ChexGen generates training-ready chest X-rays from text, masks, and bounding boxes, and shows that synthetic data can substitute for much of the real labeled data in downstream diagnosis, segmentation, detection, and fairness evaluation.

desk verdict ChexGen is a serious empirical contribution—new largest CXR dataset and broad downstream evaluation—but the label-fidelity numbers (Pearson 0.59, radiologist mean 0.49) make the 'precise control' claims softer than the tone. read the letter →

arxiv 2509.03903 v1 pith:M7WLNT4V submitted 2025-09-04 cs.CV

Yuanfeng Ji , Dan Lin , Xiyue Wang , Lu Zhang , Wenhui Zhou , Chongjian Ge , Ruihang Chu , Xiaoli Yang
show 7 more authors
Junhan Zhao Junsong Chen Xiangde Luo Sen Yang Jin Fang Ping Luo Ruijiang Li
This is my paper
classification cs.CV
keywords chestradiographygenerativefoundationmodellatentdiffusiontransformertext-to-imageControlNetdataaugmentationsyntheticpretrainingbiasmitigation
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 presents ChexGen, a generative foundation model for chest radiographs that synthesizes X-rays from text descriptions, segmentation masks, or bounding boxes. It was pretrained on OpenChest, a curated dataset of about 960,000 radiograph-report pairs assembled from eight public datasets, then aligned on expert-written reports at higher resolution. The paper's central claim is that ChexGen produces clinically realistic images with precise textual and spatial control, outperforming prior CXR generation models, and that the synthetic images are useful as training data for classification, segmentation, detection, severity scoring, and as controlled test cohorts for detecting and mitigating demographic bias. If true, this offers a practical route around the scarcity of well-annotated medical images: generate the data you need, then train or audit models on it.

What carries the argument

The load-bearing components are OpenChest, a curated corpus of about 960,000 radiograph-text pairs created by converting structured annotations from eight public datasets into standardized radiological descriptions via a two-pass LLM pipeline; a latent diffusion transformer with a frozen variational autoencoder, a frozen text encoder, and a DiT-style transformer backbone trained with classifier-free guidance; and a ControlNet-style adapter that encodes spatial annotations into latent conditioning features, allowing pathologies to appear at specified locations. The two-stage curriculum first learns broad vision-language alignment on large noisy descriptions, then fine-grained alignment on exp

What would settle it

Take a set of real chest X-rays, generate synthetic counterparts from their ground-truth finding labels, then train a standard classifier to distinguish synthetic images whose condition was present from those where it was absent. If ChexGen's label transfer is faithful, a classifier trained on the synthetic images should achieve AUC close to that of a classifier trained on real images with the same labels; a large gap would indicate that part of the synthetic training signal is noise, casting doubt on the augmentation and pretraining gains when applied to unseen real data.

Watch

Extended reading notes

Core claim

ChexGen is a latent diffusion transformer whose denoising backbone is conditioned by a frozen text encoder, trained in two stages: first on roughly 960,000 machine-generated image-text pairs at 256x256 resolution, then fine-tuned on 45,000 expert-written PA-view reports at 512x512 resolution. A lightweight ControlNet-style adapter adds mask- and bounding-box-conditioned synthesis while keeping the backbone frozen. On 3,500 held-out PA-view chest X-rays, the paper reports substantially lower FID than prior CXR text-to-image models (0.89 vs 4.42 and 9.63 using CLIP features), higher diversity (SSIM 0.41 vs 0.46 and 0.57), and higher factual correctness (Pearson correlation 0.59 vs 0.45 and 0.2

Load-bearing premise

The argument assumes that the disease labels and spatial annotations used as conditioning inputs transfer to the generated image accurately enough for downstream learning to pick up real signal; the paper's own measurements, such as a Pearson correlation of 0.59 between real and synthetic disease scores and a radiologist agreement mean of 0.49 on a -2 to +2 scale, show that a meaningful share of generated findings are imperfectly preserved.

Editorial extensions

If this is right

  • Synthetic data can substantially lower the amount of real labeled data needed: models pretrained on ChexGen-generated images and fine-tuned on 1% of real training data matched or exceeded models trained on the full real dataset.
  • Augmentation with ChexGen-generated images improves not only classification but also tasks requiring spatial precision, such as pneumothorax and rib segmentation, foreign-object detection, and opacity severity regression, suggesting the synthetic images carry usable structural and pathological information.
  • ChexGen-generated test cohorts can reveal demographic performance gaps that mirror real-world disparities, offering a lower-cost way to audit AI models for fairness before deployment.
  • Re-balancing underrepresented demographic groups with synthetic images narrows best-to-worst performance gaps by up to 60%, pointing toward a practical bias-mitigation strategy for clinical AI.
  • A single generative backbone, extended with a lightweight adapter, can serve multiple conditioning modalities, meaning task-specific synthetic datasets can be produced without retraining the foundation model.

Reading between the lines

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

  • A natural next test is to check whether ChexGen's downstream gains concentrate in findings where the generator is most faithful: stratifying augmentation and pretraining improvements by per-pathology label-transfer fidelity (e.g., the per-class Pearson correlation between real and synthetic disease scores) would separate true signal from generic regularization effects.
  • The high correlation between real and synthetic cohort performance (r = 0.83 to 1.0) suggests ChexGen could be used as a synthetic patient population for auditing models on protected attributes beyond the three axes examined, including intersections of race, age, sex, and socioeconomic proxies, without collecting new private real-world data.
  • The two-stage recipe--large-scale noisy pretraining, expert-report alignment, then frozen-backbone spatial adaptation--is portable to other imaging modalities; if it generalizes, synthetic pretraining could become a standard tool for label-hungry medical imaging tasks beyond chest radiography.
  • Because the paper's augmentation, pretraining, and bias results all rely on the same generated images, a single systematic failure mode--for example, silent generation of findings that were not in the conditioning annotation--could affect all three applications; directly measuring the false-positive and false-negative rates of generated findings per condition would sharpen the practical limits of
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

3 major / 6 minor

Summary. The paper introduces ChexGen, a latent-diffusion-transformer generative foundation model for chest radiographs, trained on a newly curated OpenChest dataset of ~960k image-text pairs and aligned on 45k MIMIC-CXR report pairs. It supports text-, mask-, and bounding-box-conditioned synthesis, and the authors evaluate image quality with FID/SSIM/Pearson metrics and two board-certified radiologists. The paper then reports three downstream applications: data augmentation (classification, segmentation, detection, severity regression), supervised pretraining followed by fine-tuning on limited real data, and synthetic test cohorts for demographic bias detection and mitigation. Headline results include FID 0.89 (CLIP) versus 4.42 for RoentGen, average AUC improvement from 78.0 to 80.1 with 2x synthetic augmentation, macro AUC 78.6 at 1% real data after 5x synthetic pretraining versus 78.1 for ImageNet-pretrained full-data training, and high correlations between real and synthetic test-cohort subgroup performance (r up to 0.995).

Significance. If the central claims hold, this is a substantial contribution: it provides a large publicly derived CXR image-text dataset, a scalable DiT-based generator with a unified conditioning interface, and a broad downstream evaluation spanning classification, segmentation, detection, regression, and fairness. The use of a held-out MIMIC-CXR test set, external datasets, and two radiologist raters are genuine strengths, as is the reported code/weight availability. However, the causal claims about synthetic data utility rest on two assumptions that are not fully tested: that disease labels transfer faithfully from conditioning inputs to generated images, and that the gains come from the synthetic nature of the data rather than from added data volume or domain-appropriate pretraining. These are fixable with additional controls and audits, but they are load-bearing for the paper's strongest conclusions.

major comments (3)
  1. [Results: Synthesizing data for training data augmentation; Methods: Supervised pretraining experiments] The downstream utility experiments assume that a generated image actually contains the pathology specified by the conditioning input. In the augmentation experiments 'each generated image was paired with its corresponding condition annotations' (Results), and in the pretraining experiments 'each synthetic image inherits the disease labels from its corresponding source case' (Methods, Supervised pretraining experiments). No per-image verification of this transfer is reported. The paper's own quality evidence shows only moderate factual fidelity: Pearson r = 0.59 between disease scores on real and synthetic images (Figure 2a) and radiologist mean scores of 1.41 and 0.49 on a -2..2 scale (Figure 2c). For rare classes, even a small systematic drop/add rate can corrupt the training signal. The observed downstream gains (e.g., AUC 78.0 to 80.1 in Figure 3c) may therefore partly reflect added d
  2. [Results: Synthesizing data for supervised pretraining; Methods: Supervised pretraining experiments] The augmentation and pretraining experiments lack a real-data scaling control. In Figure 3, adding 1x and 2x synthetic data is compared only to the real-data baseline; no experiment increases the real training set by the same factor. In the pretraining experiments (Figure 4), ChexGen-synthetic pretraining is compared only with ImageNet initialization; there is no baseline that pretrains on real chest radiographs with the same labels and the same protocol. The headline result—5x synthetic pretraining plus 1% real data reaching 78.6% macro AUC versus 78.1% for ImageNet-pretrained full-data training—could therefore be explained by the synthetic images being domain-appropriate X-rays and by the large number of pretraining examples, rather than by the synthetic data per se. Adding a real-image pretraining control at matched scale (e.g., using the 45k source MIMIC images, or a larger real CXR
  3. [Results: Synthesizing data for model bias detection and mitigation; Methods: Model bias detection and mitigation experim] The validation of synthetic test cohorts for bias detection uses Pearson correlations over very few demographic groups: gender has only two groups (making r = 1 uninformative), and the Methods text is internally inconsistent ('three racial groups (White and Black)'). With two or three points, a high correlation is weak evidence that synthetic cohorts 'closely mirror' real test data. In addition, the synthetic test set is generated from the same real impression reports and demographic attributes used for real evaluation, so the high correlations may partly reflect the conditioning variables rather than an independent ability to reveal subgroup performance. Please report per-group sample sizes, confidence intervals, and a more demanding validation (e.g., per-disease subgroup error rates or calibration) before claiming synthetic cohorts can reliably detect model bias.
minor comments (6)
  1. [Methods: Quality Assessment of Generated Images] The statistical testing for FID/SSIM/Pearson is not specified. FID and SSIM on a fixed test set are point estimates; please state the resampling or permutation procedure that produces the reported p < 0.001 values, and report confidence intervals.
  2. [Methods: Quality Assessment of Generated Images; Figure 2a] The diversity metric is average pairwise SSIM among four images from the same prompt. This measures within-prompt variability, not coverage of the real CXR distribution; a prompt-ignoring model could score well. Please label it as 'within-prompt variability' or add a distribution-coverage measure.
  3. [Methods: Dataset Curation] The OpenChest captions are generated by GPT-4 with Qwen-2.5 quality control, but no radiologist audit of the final text is reported. The authors note in Results that Stage I descriptions 'may contain noise or hallucinations'; a small expert audit of OpenChest captions would help calibrate the remaining noise.
  4. [Results: Quality assessment of ChexGen-generated images] The two radiologists' mean scores differ substantially (1.41 vs 0.49). Reporting inter-rater agreement (e.g., Cohen's kappa) and per-case score distributions would help interpret the subjective evaluation.
  5. [Methods: Quality Assessment of Generated Images] The factual-correctness metric description ('Pearson correlation coefficient between the prediction scores and the ground truths') is ambiguous; Results clarifies that it is between scores on synthetic and real images. Please align the Methods wording and report per-class correlations, since the aggregate r = 0.59 may hide large variation across the 14 diseases.
  6. [Extended Data Figure 2] Typo: 'ChestGen' should be 'ChexGen'; also 'pertaining' should be 'pretraining'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the paper's central claims are validated against held-out real datasets and external benchmarks.

full rationale

The paper's derivation chain is: curate OpenChest from public datasets, train ChexGen with a two-stage text-to-image objective, evaluate generation quality on a held-out MIMIC-CXR test set versus published baselines, and then test downstream utility on real test sets (MIMIC-CXR, VinDr-CXR, MedFMC, SIIM-ACR-PTX, Candix-PTX, VinDr-RibCXR, Object-CXR, ChestX-Det, RALO). The quality metrics (FID, MS-SSIM, Pearson correlation, radiologist scores) are computed between synthetic images and real reference images, and the radiologist evaluation is blinded and uses real impression reports. In the augmentation and pretraining experiments, synthetic images inherit labels from the conditioning annotations or source cases, but these labels are used as training targets, not as measurements that the paper claims to predict; the downstream models are evaluated exclusively on real labeled test data. The bias-detection experiment computes correlations between model performance on real and synthetic test cohorts; although the synthetic cohorts are generated from real impressions and demographic attributes, the correlation is an empirical validation rather than a definitional identity. No parameter is fitted to a subset and then reported as a prediction on the same subset, no load-bearing self-citation chain exists, and no uniqueness claim is imported from the authors' prior work. The paper's own limitations—rare-disease validation and further investigation of synthetic data for model evaluation—are generalizability concerns, not circularity. Accordingly, no circular steps were identified.

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

No fitted free parameters in the classical sense; the model is trained end-to-end and hyperparameters are standard. The central claims rest on domain assumptions about latent fidelity, label transfer, and the validity of automatic text and demographic labels. No new particles, forces, or physical entities are introduced.

free parameters (2)
  • Classifier-free guidance scale = 4.0
    Hand-chosen for text-to-image alignment; affects fidelity/diversity tradeoff.
  • Stage II token length threshold = 120
    Filter that selects MIMIC-CXR reports with impression length < 120 tokens; a manual curation choice.
assumptions (5)
  • domain assumption The Stable Diffusion VAE latent representation preserves clinically relevant image details.
    The generator is built on a frozen SD VAE; if it discards fine medical detail, generated CXRs cannot be clinically accurate.
  • domain assumption Disease labels transfer from conditioning text/annotations to generated images.
    All downstream augmentation and pretraining experiments assume the label of a synthetic image equals the label of the prompt; the paper's reported Pearson correlation of 0.59 indicates imperfect transfer.
  • domain assumption GPT-4-generated descriptions used in Stage I are accurate enough to train text-image alignment.
    Stage I uses 960k auto-generated reports; the paper itself notes these may contain noise or hallucinations.
  • domain assumption The FID/SSIM/Pearson metrics and radiologist Likert scores are valid proxies for clinical fidelity.
    The paper's quality claims rest on these metrics; if they miss clinically important distortions, the fidelity claims are overstated.
  • domain assumption Demographic labels (race, age) obtained through MIMIC-IV linkage are correct.
    Used for bias detection and mitigation; linkage errors would confound fairness conclusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generative Foundation Model for Chest Radiography." pith.science (2026). https://pith.science/paper/M7WLNT4V

@misc{pith2026250903903,
  author       = {Pith},
  title        = {Pith review of: A Generative Foundation Model for Chest Radiography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M7WLNT4V}},
  note         = {Machine review of arXiv:2509.03903}
}
read the original abstract

The scarcity of well-annotated diverse medical images is a major hurdle for developing reliable AI models in healthcare. Substantial technical advances have been made in generative foundation models for natural images. Here we develop `ChexGen', a generative vision-language foundation model that introduces a unified framework for text-, mask-, and bounding box-guided synthesis of chest radiographs. Built upon the latent diffusion transformer architecture, ChexGen was pretrained on the largest curated chest X-ray dataset to date, consisting of 960,000 radiograph-report pairs. ChexGen achieves accurate synthesis of radiographs through expert evaluations and quantitative metrics. We demonstrate the utility of ChexGen for training data augmentation and supervised pretraining, which led to performance improvements across disease classification, detection, and segmentation tasks using a small fraction of training data. Further, our model enables the creation of diverse patient cohorts that enhance model fairness by detecting and mitigating demographic biases. Our study supports the transformative role of generative foundation models in building more accurate, data-efficient, and equitable medical AI systems.

Figures

Figures reproduced from arXiv: 2509.03903 by the authors.

Figure 4
Figure 4. Supervised pretraining using synthetic data. ChexGen-generated data can be used for super￾vised pretraining, which improves model performance on downstream tasks including disease classification. a, Classification performance (macro AUC) for models with varying amounts of ChexGen-synthesized data (from 1× to 5× the size of task-specific training data) for pretraining and varying amounts of real data for training. Mo… view at source ↗
Figure 5
Figure 5. Bias detection and mitigation using synthetic data. ChexGen-generated synthetic data can used to supplement real-world data for model evaluation. a, Conceptual framework to illustrate that ChexGen￾generated data can be utilized as test data to evaluate model performance (e.g., accuracy, fairness), offering an alternative to costly and difficult-to-obtain real-world data. b, Demographic distribution of the MIMIC data… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Revisiting 2D Foundation Models for Scalable 3D Medical Image Classification

    cs.CV 2025-12 conditional novelty 5.0 of 10

    A frozen 2D vision foundation model with lightweight LoRA adapters and attention-based slice fusion achieves state-of-the-art 3D medical image classification across 12 tasks with about 1M trainable parameters per task.

Reference graph

Works this paper leans on

54 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature 616, 259–265 (2023)

    2023

  2. [2]

    Rao, V . M.et al. Multimodal generative ai for medical image interpretation. Nature 639, 888–896 (2025)

    work page 2025

  3. [3]

    Wang, J. et al. Self-improving generative foundation model for synthetic medical image generation and clinical applications. Nature Medicine 1–9 (2024)

    work page 2024

  4. [4]

    Ktena, I. et al. Generative models improve fairness of medical classifiers under distribution shifts. Nature Medicine 30, 1166–1173 (2024)

    work page 2024

  5. [5]

    Sun, Y . et al. A data-efficient strategy for building high-performing medical foundation models. Nature Biomedical Engineering 1–13 (2025)

    work page 2025

  6. [6]

    & Abbeel, P

    Ho, J., Jain, A. & Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020)

    work page 2020

  7. [7]

    Ramesh, A. et al. Zero-shot text-to-image generation. In International conference on machine learning, 8821–8831 (Pmlr, 2021)

    work page 2021

  8. [8]

    Saharia, C. et al. Photorealistic text-to-image diffusion models with deep language understanding. Ad- vances in neural information processing systems 35, 36479–36494 (2022)

    work page 2022

Show all 54 references
  1. [9]

    & Ommer, B

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P. & Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684–10695 (2022)

    2022

  2. [10]

    & Agrawala, M

    Zhang, L., Rao, A. & Agrawala, M. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, 3836–3847 (2023)

    2023

  3. [11]

    & Nichol, A

    Dhariwal, P. & Nichol, A. Diffusion models beat gans on image synthesis.Advances in neural information processing systems 34, 8780–8794 (2021)

    2021

  4. [12]

    Croitoru, F.-A., Hondru, V ., Ionescu, R. T. & Shah, M. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence45, 10850–10869 (2023)

    2023

  5. [13]

    Bluethgen, C. et al. A vision–language foundation model for the generation of realistic chest x-ray images. Nature Biomedical Engineering 1–13 (2024)

    2024

  6. [14]

    Carrillo-Perez, F. et al. Generation of synthetic whole-slide image tiles of tumours from rna-sequencing data via cascaded diffusion models. Nature Biomedical Engineering 9, 320–332 (2025)

    2025

  7. [15]

    Wang, X. et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2097–2106 (2017)

  8. [16]

    Irvin, J. et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. vol. 33, 590–597 (2019)

    2019

  9. [17]

    Vayá, M. D. L. I. 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

  10. [18]

    H., Tran, T

    Pham, H. H., Tran, T. T. & Nguyen, H. Q. Vindr-pcxr: An open, large-scale pediatric chest x-ray dataset for interpretation of common thoracic diseases. PhysioNet (version 1.0. 0) 10 (2022)

    2022

  11. [19]

    Reis, E. P. et al. Brax, brazilian labeled chest x-ray dataset. Scientific Data 9, 487 (2022). 50

    2022

  12. [20]

    Chen, J. et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthe- sis. arXiv preprint arXiv:2310.00426 (2023)

    2023 arXiv

  13. [21]

    Esser, P. et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning (2024)

    2024

  14. [22]

    & Hochreiter, S

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  15. [23]

    Wang, Z., Simoncelli, E. P. & Bovik, A. C. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003 , vol. 2, 1398–1402 (Ieee, 2003)

    2003

  16. [25]

    Cohen, J. P. et al. TorchXRayVision: A library of chest X-ray datasets and models. In Medical Imaging with Deep Learning (2022). URL https://github.com/mlmed/torchxrayvision

    2022

  17. [26]

    Johnson, A. E. et al. Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs.arXiv preprint arXiv:1901.07042 (2019)

    1901 arXiv

  18. [27]

    Siim-acr pneumothorax segmentation 2019 (2019)

    American College of Radiology. Siim-acr pneumothorax segmentation 2019 (2019). URL https: //www.kaggle.com/competitions/siim-acr-pneumothorax-segmentation/data

    2019

  19. [28]

    C., Le, T

    Nguyen, H. C., Le, T. T., Pham, H. H. & Nguyen, H. Q. 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

  20. [29]

    Kufel, J. et al. Chest x-ray foreign objects detection using artificial intelligence. Journal of Clinical Medicine 12, 5841 (2023)

    2023

  21. [30]

    Liu, J., Lian, J. & Yu, Y . Chestx-det10: Chest x-ray dataset on detection of thoracic abnormalities (2020). 2006.10550v3

    2020 arXiv

  22. [31]

    Cohen, J. P. et al. Radiographic assessment of lung opacity score dataset. Zenodo, 2021b. URL https://doi. org/10.5281/zenodo 4633999 (2021)

    2021 doi

  23. [32]

    Shumailov, I. et al. Ai models collapse when trained on recursively generated data. Nature 631, 755–759 (2024)

    2024

  24. [33]

    Seddik, M. E. A., Chen, S.-W., Hayou, S., Youssef, P. & Debbah, M. How bad is training on synthetic data? a statistical analysis of language model collapse. arXiv preprint arXiv:2404.05090 (2024)

    2024 arXiv

  25. [34]

    Selvaraju, R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, 618–626 (2017)

    2017

  26. [35]

    Chen, Z. et al. Chexagent: Towards a foundation model for chest x-ray interpretation. arXiv preprint arXiv:2401.12208 (2024)

    2024 arXiv

  27. [36]

    Pérez-García, F. et al. Exploring scalable medical image encoders beyond text supervision. Nature Ma- chine Intelligence 1–12 (2025)

    2025

  28. [37]

    & De La Iglesia-Vaya, M

    Bustos, A., Pertusa, A., Salinas, J.-M. & De La Iglesia-Vaya, M. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis 66, 101797 (2020)

    2020

  29. [38]

    Tang, J. S. et al. Clip, catheter and line position dataset. Scientific Data 8, 285 (2021)

    2021

  30. [39]

    Achiam, J. et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023). 51

    2023 arXiv

  31. [40]

    Yang, A. et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024)

    2024 arXiv

  32. [41]

    Betker, J. et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf 2, 8 (2023)

    2023

  33. [42]

    & Xie, S

    Peebles, W. & Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4195–4205 (2023)

    2023

  34. [43]

    & Salimans, T

    Ho, J. & Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    2022 arXiv

  35. [44]

    & Ermon, S

    Song, J., Meng, C. & Ermon, S. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2010 arXiv

  36. [45]

    Radford, A. et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, 8748–8763 (PMLR, 2021)

    2021

  37. [46]

    & Weinberger, K

    Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 4700–4708 (2017)

    2017

  38. [47]

    Nguyen, H. Q. et al. Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations. Scientific Data 9, 429 (2022)

    2022

  39. [48]

    & Brox, T

    Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmenta- tion. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, 23...

    2015

  40. [49]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    2016

  41. [50]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Ren, S. Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497 (2015)

    2015 arXiv

  42. [51]

    Feature pyramid networks for object detection

    Lin, T.-Y .et al. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2117–2125 (2017)

    2017

  43. [52]

    Object-cxr - automatic detection of foreign objects on chest x-rays (2019)

    JF Healthcare. Object-cxr - automatic detection of foreign objects on chest x-rays (2019). URL https: //jfhealthcare.github.io/object-CXR/

    2019

  44. [53]

    & Xie, W

    Wu, C., Zhang, X., Zhang, Y ., Wang, Y . & Xie, W. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463 (2023)

    2023 arXiv

  45. [54]

    & Winzeck, S

    Glocker, B., Jones, C., Bernhardt, M. & Winzeck, S. Algorithmic encoding of protected characteristics in chest x-ray disease detection models. EBioMedicine 89 (2023)

    2023

  46. [55]

    Wang, D. et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data 10, 574 (2023). 52

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.