Pith. sign in

REVIEW 5 major objections 4 minor 48 references

A chest X-ray encoder trained on unlabeled images alone can beat text-supervised and distillation-based encoders.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 08:43 UTC pith:KAYR6NL6

load-bearing objection RadJEPA is a straightforward I-JEPA port to chest X-rays with plausible but not airtight results; the MIMIC-CXR leakage risk and abstract/full-text mismatch justify conditional acceptance. the 5 major comments →

arxiv 2601.15891 v3 pith:KAYR6NL6 submitted 2026-01-22 cs.CV

RadJEPA: Radiology Encoder for Chest X-Rays via Joint Embedding Predictive Architecture

Anas Anwarul Haq Khan , Mariam Husain , Pratik Jalan , Kshitij Jadhav This is my paper
classification cs.CV
keywords chest X-rayradiology encodingself-supervised learningjoint embedding predictive architecturelatent predictionreport generationrepresentation learningvision-language pretraining
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether a radiology encoder needs language supervision at all. It introduces RadJEPA, trained on 839,364 unlabeled chest X-rays, to predict latent representations of masked image regions from visible context. Across disease classification, semantic segmentation, and radiology report generation, the frozen encoder matches or exceeds the strongest vision-language and distillation-based baselines, despite a smaller backbone and lower input resolution. The authors conclude that predictive latent modeling is a viable and perhaps better route to general-purpose radiology representations than image-text alignment or view-invariance approaches.

Core claim

The central claim is that a Joint Embedding Predictive Architecture, trained purely on unlabelled chest X-rays, learns representations that transfer better to downstream radiology tasks than representations learned from paired image-text data or from self-distillation. The objective minimizes the squared distance between a predictor's output on a visible context region and a momentum-updated target encoder's embedding of a masked target region. The authors report that RadJEPA exceeds the strongest baselines on disease classification, semantic segmentation, and report generation, with particularly large gains on subtle findings like pulmonary fibrosis and on fine-grained anatomical segmentati

What carries the argument

The central mechanism is the JEPA (Joint Embedding Predictive Architecture) objective: a context encoder and a momentum-updated teacher encode non-overlapping image regions, and a small predictor network must reconstruct the teacher's latent representation of a masked target region, with no pixel reconstruction, no contrastive negatives, and no augmentations. This forces the encoder to capture semantic content that explains the masked region from context, which the paper argues is the right inductive bias for chest X-rays.

Load-bearing premise

The pretraining data does not overlap with the patients used in evaluation, especially the ~90,000 MIMIC-CXR lateral images added to balance viewpoints; if those patients appear in the MIMIC-CXR report test set, the flagship report-generation gains could come from data leakage rather than the predictive objective.

What would settle it

Re-run the report-generation experiment after excluding every pretraining image whose subject identifier appears in the MIMIC-CXR test split, or retrain RadJEPA without the added MIMIC lateral images; if the ROUGE-L/BLEU-4 advantage over the strongest baseline shrinks to near zero, the central claim that predictive modeling alone drives the gains would be refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If RadJEPA's claim holds, hospitals with large unlabeled chest X-ray archives can build strong encoders without written reports, which are costly to curate and biased toward positive findings.
  • Report generation with a frozen RadJEPA encoder beats text-supervised encoders, suggesting that detailed clinical narratives may not be necessary to teach visual semantics.
  • The gains occur at 224x224 resolution with an 86M-parameter backbone, implying that compute and resolution can be reduced without sacrificing representation quality if the pretraining objective is right.
  • The largest improvements on subtle and fine-grained targets indicate that predictive latent modeling preserves diagnostic cues that contrastive or text-aligned models tend to suppress.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If predictive latent modeling is the reason for the gains, then combining RadJEPA with even modest image-text data could push report generation further; the paper does not test this hybrid.
  • The ~90,000 lateral-view images added from MIMIC-CXR to balance the frontal-to-lateral ratio come from the same corpus used for the report-generation evaluation; a strict subject-exclusion audit would settle whether part of the MIMIC report gains comes from pretraining on the same patients.
  • Since the encoder beats distillation-based methods at a lower resolution, predictive objectives might also benefit other dense medical imaging tasks such as CT or MRI, where fine-grained anatomy matters more than global views.
  • The success at 224x224 suggests that the current trend toward higher input resolutions may matter less than the pretraining objective itself for representation quality.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes RadJEPA, a Joint Embedding Predictive Architecture pretrained on 839,364 unlabeled chest X-rays with a ViT-B/14 encoder. The encoder is frozen and evaluated on disease classification (linear probing), semantic segmentation (UPerNet/linear decoders), and report generation (LLaVA-style projector with Vicuna-7B). The central claim is that this purely self-supervised, prediction-based encoder matches or exceeds stronger image-only and vision-language baselines without paired image-text data, and that the improvements are statistically significant. The authors also state that code and trained weights are publicly released.

Significance. If the results are valid, RadJEPA would be a significant contribution: it would show that latent predictive modeling can outperform both CLIP-style text supervision and DINO-style distillation in radiology, with a smaller backbone and lower resolution, and would support the broader thesis that radiology encoders need not depend on paired reports. The paper covers three distinct downstream tasks and a wide baseline set, and the public release of code/weights is a strength. However, the current manuscript has several internal inconsistencies — abstract claims for experiments not in the full text, an unexplained MIMIC-CXR data row, impossible token counts at 224×224, and missing error bars — that as written prevent verification of the core claims. These are addressable but require substantial revision.

major comments (5)
  1. [§4, Table 1, §6.3.1] The pretraining data accounting is internally inconsistent. Section 4 states that only 'approximately 90k lateral chest X-rays from MIMIC-CXR' are added, yet Table 1 lists 300,491 MIMIC-CXR images (210,491 frontal + 90,000 lateral). The footnote about avoiding overlap with evaluation sets does not state whether the frontal MIMIC images are from subjects excluded from the MIMIC-CXR report-generation test split (§6.3.1, 2,461 studies). If any pretraining images come from subjects in that test set, the ROUGE-L, BLEU-4, and RGER gains in Table 5 could reflect subject-level leakage rather than better representations. This is load-bearing for the central claim, and a precise subject-level exclusion statement is essential.
  2. [Abstract vs. full text] The abstract promises that the encoder is 'additionally substitute[d] into four widely used vision-language backbones (MedLLaVA, Qwen-2.5, BLIP-2, and Phi-4).' The full text contains no such experiments or analysis. This is an unsupported claim in the abstract and must be either implemented or removed.
  3. [Table 5] The token counts for RadJEPA and I-JEPA at 224×224 are listed as 1369. For a ViT-B/14 backbone, a 224×224 input yields 16×16=256 patches, not 1369 (which corresponds to 37×37 patches, i.e., 518×518 input). This contradicts the stated resolution and undermines the 'lower input resolution' argument in §6.3.3. The table entries must be corrected and the comparison with RAD-DINO (which legitimately uses 518×518) re-evaluated.
  4. [§6.1.2, Table 3, Table 5, Table 6] The text claims results are reported as mean±standard deviation over 5 folds (§6.1.2) and as median with 95% bootstrap confidence intervals (Table 5 caption), but Tables 3 and 5 show only point estimates. Table 6 reports statistical significance without p-values or effect sizes. Without error bars or confidence intervals, the claimed 'statistically significant' improvements cannot be independently assessed.
  5. [§4] The pretraining configuration is underspecified. The text says 'follows the Meta I-JEPA protocol' but omits the EMA momentum τ, masking ratio/target block size, optimizer, batch size, number of epochs, learning rate schedule, and any regularization. These hyperparameters are necessary for reproducibility, especially since the paper's contribution is the pretraining method itself.
minor comments (4)
  1. [Table 4] The row for 'NN-UNet' and the column headers (e.g., '# Features', '# Params') are not aligned with the other rows; please reformat.
  2. [Table 5] The column header 'RG ER' is unclear; presumably it should be 'RGER'. Also, define 'Macro-F1-14' in the caption.
  3. [References] The reference list has incomplete author entries ('Assran and 1 others', 'Liang and 1 others') and duplicate RAD-DINO entries (Pérez-García et al., 2024 and 2025). Please clean up.
  4. [§2 / §4] The related work cites 'DINO-style self-distillation' but does not discuss relevant medical-imaging JEPA or MIM works beyond BrainJEPA; adding one or two medical-imaging MIM references would help contextualize the contribution.

Circularity Check

0 steps flagged

No significant circularity; RadJEPA is a domain application of I-JEPA with external benchmarks and no self-citation chain.

full rationale

RadJEPA's predictive objective is explicitly inherited from the externally published Meta I-JEPA protocol (Assran et al., 2023a), and the paper does not claim to derive that objective from its own results. The downstream evaluations use task-specific heads trained on frozen encoder features, with external benchmarks (VinDr-CXR, RSNA, IU-Xray) and controlled variants that restrict pretraining data. The abstract's claim is supported by empirical comparisons rather than by a parameter fitted to the reported metrics. The only potentially load-bearing concern is whether MIMIC-CXR pretraining includes subjects from the MIMIC report-generation test split; the paper explicitly asserts subject-level exclusion in the Table 1 footnote ('only a subset of subjects is included to avoid overlap with evaluation sets'), and no passage in the manuscript admits a violation. If that exclusion failed, the MIMIC report-generation numbers would be invalid as an evaluation, but that would be data leakage, not a circular derivation in which a prediction is defined as its own training target. There are no self-citations by the present authors, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. Therefore the correct circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard SSL assumptions plus unreported pretraining hyperparameters; the main independent-support gap is the MIMIC-CXR overlap and the missing pretraining configuration.

free parameters (4)
  • EMA momentum τ = not reported
    Momentum coefficient for target encoder EMA; inherited from I-JEPA but not specified; controls stability of pretraining.
  • Masking ratio / target block size = not reported
    Proportion and scale of masked target regions in the latent prediction objective; follows I-JEPA but exact values absent.
  • Pretraining input resolution = 224×224
    Choice of resolution; differs from RAD-DINO and I-JEPA baselines at 518/448, yet report-generation table lists 1,369 tokens, an unresolved inconsistency.
  • Projector residual coefficient λ = not reported
    Appears in adapter equation v + λW2σ(W1v) but λ is never defined.
axioms (5)
  • domain assumption I-JEPA predictive latent objective is an effective pretraining signal for chest X-rays
    The whole paper assumes that predicting masked latent regions yields useful representations for radiology; this is the hypothesis under test, adopted from Assran et al.
  • domain assumption Frozen ViT-B/14 representations transfer to classification, segmentation, and report generation with lightweight heads
    The evaluation protocol assumes linear/decoder heads can extract task-relevant information from the frozen encoder.
  • domain assumption Pretraining and evaluation subject sets do not overlap
    Section 4 claims a subset of MIMIC-CXR subjects is excluded to avoid overlap, but Table 1 lists 300,491 MIMIC images; the exact overlap with the MIMIC report-generation test set (Section 6.3.1) is unverified.
  • domain assumption Evaluation metrics (ROUGE-L, BLEU-4, RGER, Macro-F1-14) are reliable proxies for report quality
    No human evaluation or clinical validity check; standard n-gram/lexical metrics can be gamed.
  • ad hoc to paper No data augmentation during pretraining is beneficial or at least not harmful
    RadJEPA removes augmentations from I-JEPA protocol; no ablation justifies this choice.

pith-pipeline@v1.3.0-alltime-deepseek · 12740 in / 15292 out tokens · 130783 ms · 2026-08-03T08:43:16.568909+00:00 · methodology

0 comments
read the original abstract

Vision-language pretraining has driven much of the recent progress in medical image representation learning, but this paradigm is constrained by the availability of paired image-text data and by the reporting bias of clinical narratives. We ask whether competitive radiology encoders can be learned without any language supervision. We introduce RadJEPA, a self-supervised framework built on a Joint Embedding Predictive Architecture and pretrained on approximately 840K unlabeled chest X-ray images. The model learns to predict latent representations of masked target regions from a visible context region, an objective that differs from both image-text contrastive pretraining and DINO-style self-distillation by explicitly modelling conditional structure in representation space. We evaluate RadJEPA primarily on radiology report generation with a frozen Vicuna-7B decoder, and additionally substitute its encoder into four widely used vision-language backbones (MedLLaVA, Qwen-2.5, BLIP-2, and Phi-4). For completeness we also report disease classification and semantic segmentation results. Across two datasets and four metrics, RadJEPA matches or exceeds the strongest image-only and vision-language baselines while using a ViT-B/14 backbone at 224 x 224 resolution.

Figures

Figures reproduced from arXiv: 2601.15891 by Anas Anwarul Haq Khan, Kshitij Jadhav, Mariam Husain, Pratik Jalan.

Figure 1
Figure 1. Figure 1: RadJEPA frozen encoder with trainable task [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 14 linked inside Pith

  1. [1]

    MD Anouk Stein, Carol Wu, Chris Carr, George Shih, Jamie Dulkowski, kalpathy, Leon Chen, Luciano Prevedello, MD Marc Kohli, Mark McDonald, Peter, Phil Culliton, Safwan Halabi MD, and Tian Xia. 2018. Rsna pneumonia detection challenge. https://kaggle.com/competitions/rsna-pneumonia-detection-challenge. Kaggle

    2018

  2. [2]

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. 2023 a . Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619--15629

    2023

  3. [3]

    Mahmoud Assran and 1 others. 2023 b . Self-supervised learning from images with a joint-embedding predictive architecture. In IEEE Conference on Computer Vision and Pattern Recognition

    2023

  4. [4]

    Shruthi Bannur, Maximilian Ilse, Oladimeji Farri, Gaspard Pons, and Jean-Baptiste Schiratti. 2023. Learning to exploit temporal structure for biomedical vision--language pretraining. arXiv preprint arXiv:2301.12018

    Pith/arXiv arXiv 2023

  5. [5]

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. 2022. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations

    2022

  6. [6]

    Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, and 1 others. 2022. Making the most of text semantics to improve biomedical vision--language processing. In European conference on computer vision, pages 1--21. Springer

    2022

  7. [7]

    Salinas, and Mar \' a de la Iglesia-Vay \'a

    Aurelia Bustos, Antonio Pertusa, Jose M. Salinas, and Mar \' a de la Iglesia-Vay \'a . 2020. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical Image Analysis, 66:101797

    2020

  8. [8]

    Erdi C all , Ecem Sogancioglu, Bram Van Ginneken, Kicky G van Leeuwen, and Keelin Murphy. 2021. Deep learning for chest x-ray analysis: A survey. Medical image analysis, 72:102125

    2021

  9. [9]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650--9660

    2021

  10. [10]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning

    2020

  11. [11]

    Jordan, Joseph E

    Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Stefano Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael I. Jordan, Joseph E. Gonzalez, and Ion Stoica. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ arXiv preprint arXiv:2303.10130

    Pith/arXiv arXiv 2023

  12. [12]

    Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, and 1 others. 2023. Scaling vision transformers to 22 billion parameters. In International conference on machine learning, pages 7480--7512. PMLR

    2023

  13. [13]

    Karan Desai and Justin Johnson. 2021. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162--11173

    2021

  14. [14]

    Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, and 1 others. 2023. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10995--11005

    2023

  15. [15]

    Zijian Dong, Ruilin Li, Yilei Wu, Thuan Tinh Nguyen, Joanna Chong, Fang Ji, Nathanael Tong, Christopher Chen, and Juan Helen Zhou. 2024. Brain-jepa: Brain dynamics foundation model with gradient positioning and spatiotemporal masking. Advances in Neural Information Processing Systems, 37:86048--86073

    2024

  16. [16]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \'a r, and Ross Girshick. 2022. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition

    2022

  17. [17]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition

    2020

  18. [18]

    Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. 2023. Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2506--2517

    2023

  19. [19]

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Mihaela Ciurea-Ilinca, Chris Chute, Henrik Marklund, Babak Haghgoo, Robyn Ball, Katie Shpanskaya, and 1 others. 2019. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 590--597

    2019

  20. [20]

    Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Koehler, Tobias Norajitra, Sebastian Wirkert, and Klaus H

    Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F. Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Koehler, Tobias Norajitra, Sebastian Wirkert, and Klaus H. Maier-Hein. 2018. nnu-net: Self-adapting framework for u-net-based medical image segmentation. arXiv preprint arXiv:1809.10486

    Pith/arXiv arXiv 2018

  21. [21]

    Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317

    2019

  22. [22]

    Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. 2024. A causal perspective on dataset bias in machine learning for medical imaging. Nature Machine Intelligence, 6(2):138--146

    2024

  23. [23]

    Yann LeCun. 2022. A path towards autonomous machine intelligence. arXiv preprint arXiv:2207.09215

    Pith/arXiv arXiv 2022

  24. [24]

    Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. 2022. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision (ECCV)

    2022

  25. [25]

    Paul Pu Liang and 1 others. 2022. Mind the gap: Understanding the modality gap in vision--language models. In Advances in Neural Information Processing Systems

    2022

  26. [26]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023 a . Improved baselines with visual instruction tuning. arXiv preprint arXiv:2306.06202

    Pith/arXiv arXiv 2023

  27. [27]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023 b . Visual instruction tuning. arXiv preprint arXiv:2304.08485

    Pith/arXiv arXiv 2023

  28. [28]

    Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, and 1 others. 2022. Vindr-cxr: An open dataset of chest x-rays with radiologist’s annotations. Scientific Data, 9(1):429

    2022

  29. [29]

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

    Pith/arXiv arXiv 2021

  30. [30]

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

    Pith/arXiv arXiv 2018

  31. [31]

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, and 1 others. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

    Pith/arXiv arXiv 2023

  32. [32]

    Fernando P \'e rez-Garc \' a, Harshita Sharma, Sam Bond-Taylor, Kenza Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, and 1 others. 2024. Rad-dino: Exploring scalable medical image encoders beyond text supervision. arXiv preprint arXiv:2401.10815

    Pith/arXiv arXiv 2024

  33. [33]

    Fernando Perez-Garcia, Harshita Sharma, Sam Bond-Taylor, Kenza Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, and 1 others. 2025. Exploring scalable medical image encoders beyond text supervision. Nature Machine Intelligence, 7(1):119--130

    2025

  34. [34]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748--8763. PmLR

    2021

  35. [35]

    Reis, Maur \' cio J

    Eduardo P. Reis, Maur \' cio J. Cardoso, Filipe Cordeiro, Jo \ a o F. Teixeira, Victor Alves, Hugo Silva, and 1 others. 2022. Brax: A large-scale radiology dataset for benchmarking chest x-ray algorithms. Scientific Data, 9(1):1--10

    2022

  36. [36]

    Mingxing Tan and Quoc V. Le. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. International Conference on Machine Learning (ICML)

    2019

  37. [37]

    Yucheng Tang and 1 others. 2022. Self-supervised learning of medical image representations via relation-driven masking. IEEE Transactions on Medical Imaging

    2022

  38. [38]

    Lungren, and Andrew Y

    Ethan Tiu, Yifan Peng, Michael Zhou, Pranav Rajpurkar, Matthew P. Lungren, and Andrew Y. Ng. 2022. Expert-level zero-shot learning for chest x-ray interpretation. arXiv preprint arXiv:2206.12068

    Pith/arXiv arXiv 2022

  39. [39]

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. 2017. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3462--3471

    2017

  40. [40]

    Joy T Wu, Nkechinyere N Agu, Ismini Lourentzou, Arjun Sharma, Joseph A Paguio, Jasper S Yao, Edward C Dee, William Mitchell, Satyananda Kashyap, Andrea Giovannini, and 1 others. 2021. Chest imagenome dataset for clinical reasoning. arXiv preprint arXiv:2108.00316

    Pith/arXiv arXiv 2021

  41. [41]

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. 2018. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision (ECCV)

    2018

  42. [42]

    Xiaohua Zhai and 1 others. 2022. Lit: Zero-shot transfer with locked-image text tuning. In IEEE Conference on Computer Vision and Pattern Recognition

    2022

  43. [43]

    Yuhao Zhang, Yu Wang, Yuhui Zhang, Zhe Li, and Yifan Xu. 2023. Large-scale domain-specific vision--language pretraining for biomedical applications. arXiv preprint arXiv:2303.00915

    Pith/arXiv arXiv 2023

  44. [44]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36:46595--46623

    2023

  45. [45]

    Hong-Yu Zhou, Chenyu Lian, Liansheng Wang, and Yizhou Yu. 2023. Advancing radiograph representation learning with masked record modeling. arXiv preprint arXiv:2301.13155

    Pith/arXiv arXiv 2023

  46. [46]

    Zongwei Zhou and 1 others. 2019. Models genesis: Generic autodidactic models for 3d medical image analysis. In International Conference on Medical Image Computing and Computer-Assisted Intervention

    2019

  47. [47]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  48. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...