Pith. sign in

REVIEW 3 major objections 5 minor 66 references

Understanding Dataset Bias in Medical Imaging: A Case Study on Chest X-rays

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that four large open-source chest X-ray datasets carry a learnable origin signal that survives stripping away anatomical structure, and that pixel intensity and texture, not anatomy, drive the bias.

desk verdict Useful empirical warning, but the main claim about pixel intensity/texture over anatomical structure is contradicted by their own contour results. read the letter →

arxiv 2507.07722 v4 pith:A2UJ3QVU submitted 2025-07-10 cs.CV

Ethan Dack , Chengliang Dai This is my paper · ORCID
classification cs.CV
keywords datasetbiaschestX-rayNameThatoriginclassificationpixelintensityandtexturesegmentationmasksmedicalimaging
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

This paper asks whether the well-known "Name That Dataset" effect—where classifiers can tell which dataset an image came from—also holds for chest X-rays. Testing four widely used open-source collections with three network architectures, it finds that dataset origin is almost perfectly detectable: ResNet50 reaches 99.28% F1 on cropped images and 98.96% when only the lungs and heart remain. Even images reduced to one-pixel organ contours stay classifiable at 88.62% average F1. From this and from shuffling ablations, the paper concludes that pixel intensity and texture, not anatomical structure, carry the dataset-specific signal. If that is right, diagnostic models trained on these datasets may be learning which source produced the image rather than the pathology it contains, with direct consequences for cross-dataset generalization and foundation models.

What carries the argument

The paper's diagnostic instrument is a family of four image transformations—crop around the lungs, semantic segmentation maps, one-pixel contour outlines, and lung-and-heart-only composites—all produced by a single segmentation model, together with ablation experiments that shuffle image patches or individual pixels. Each transformation strips a different kind of information: contours remove almost all texture and intensity, lung-and-heart images remove background and bone, and pixel shuffling destroys local texture while preserving global layout. The pattern of classification accuracy across these variants is what the claim about pixel intensity and texture rests on.

What would settle it

Re-run the four-way origin classifier on semantic, contour, and lung/heart images generated by an independently trained segmentation model rather than the single model used in the paper; if the high F1 scores collapse, the anatomical-structure images were carrying a fingerprint of the segmenter, not of the X-ray acquisition, while if they persist, the paper's texture-intensity conclusion is confirmed.

Watch

Extended reading notes

Core claim

The central discovery is that the source of a chest X-ray can be identified with high accuracy even after nearly all anatomical information is removed. Using one segmentation model to prepare four image variants—crops around the lungs, semantic segmentation maps, one-pixel contour outlines, and lung-and-heart-only composites—the paper reports ResNet50 F1 scores of 99.28%, 86.59%, 88.62%, and 98.96% for the four-way origin task. Shuffling individual pixels drops the scores to 62.52% and 57.81% on cropped and lung-and-heart images, while shuffling larger patches leaves them near intact at 97.81% and 96.13%; the paper reads this as direct evidence that local pixel intensity and texture, not global anatomical arrangement, carry the bias. The paper's conclusion is that, unlike natural images where semantic maps and structures dominate dataset bias, chest X-ray bias survives structural stripping and is primarily a texture-and-intensity phenomenon.

Load-bearing premise

The conclusion that anatomy does not drive the bias rests on transformations all generated by the same segmentation model; if that model has itself learned dataset-specific shape or intensity patterns, the "anatomy-only" images still carry a dataset fingerprint.

Editorial extensions

If this is right

  • Diagnostic models trained on these datasets may be learning dataset origin alongside or instead of pathology, so their cross-dataset performance could be poorer than within-dataset benchmarks suggest.
  • Foundation and self-supervised models trained on merged open-source chest X-ray collections can inherit this origin signal, making reported gains look larger than what would transfer to new clinical sites.
  • Intensity- and texture-based data augmentation only mildly reduces origin classification (from 99.28% to about 95.96% for combined transforms), so standard augmentation is unlikely to erase the bias.
  • The persistence of high accuracy at 32x32 resolution (93.67% F1) implies the origin signal survives extreme downsampling and is located in coarse intensity statistics rather than fine detail.

Reading between the lines

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

  • A direct test of the paper's mechanism would be to redo the semantic, contour, and lung/heart variants with masks from an independently trained segmenter; a large drop in accuracy would mean part of the "texture" signal is actually a fingerprint of the segmentation model shared by all variants.
  • If the intensity/texture origin signal is general, the same probe could screen other radiology modalities (CT, mammography, brain MRI) and multi-site versions of one scanner type to measure how much acquisition-specific signal survives before datasets are released for diagnostic training.
  • The near-null effect of patch shuffling relative to pixel shuffling suggests a cheap pre-release test for new X-ray collections: run an origin classifier on pixel-shuffled images, and if accuracy stays near chance the collection is comparatively clean of local texture bias; if not, texture bias remains.
  • The paper's finding may have unequal consequences across disease classes: classes with narrow intensity/texture ranges could be more susceptible to origin memorization, which would make long-tailed disease benchmarks especially vulnerable.
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 / 5 minor

Summary. The paper revisits the "Name That Dataset" task for four large open-source chest X-ray datasets (NIH, CheXpert, MIMIC-CXR, PadChest), training AlexNet, ResNet50, and ViT to classify dataset origin under several image transformations: cropped images, semantic segmentation maps, contour-only images, and lung-and-heart-only images. The authors report very high F1 scores (e.g., ResNet50 achieves 99.28% on cropped, 98.96% on lung-and-heart, 88.62% on contours) and, based on augmentation and ablation experiments (patch shuffling versus pixel shuffling), conclude that dataset bias is primarily driven by pixel intensity and texture rather than anatomical structure. The paper includes patient-level train/test splits, a comparison across architectures, and heatmap and segmentation-statistic analyses to explain the bias.

Significance. If the central claim were fully supported, the paper would provide an important cautionary result for the medical imaging community: popular open-source chest X-ray datasets carry a learnable origin signal that survives substantial structural stripping, and models may rely on low-level cues rather than clinically relevant anatomy. The manuscript is commendable for using multiple architectures, providing public code, and performing patient-level splits. However, the key interpretive claim is undermined by the paper's own evidence: contour-only images, which remove intensity and texture by construction, still achieve 88.62% F1, and the paper's segmentation statistics explicitly explain the semantic and contour results via anatomical shape differences. The conclusion that structure is "largely irrelevant" overreaches the experiments, which show only that texture is sufficient, not that structure is not also a strong driver. The lack of repeated runs and confidence intervals further weakens the comparative claims. With substantial revision, the work could become a useful empirical case study, but in its current form the conclusion is not convincingly supported.

major comments (3)
  1. [Section 5 and Table 1] The conclusion that "dataset bias is primarily driven by pixel intensity and texture rather than anatomical structure" is not supported by the presented evidence. The contour-only images, which preserve anatomy while removing pixel intensity and local texture by construction, yield 88.62% F1 with ResNet50 (Table 1), far above the 25% chance level. Moreover, Section 4.2 states that the mask-area statistics in Table 2 "can explain how models can classify so well within the semantic and contour tasks." If coarse anatomical proportions (background, left/right lung, facies diaphragmatica) explain the semantic and contour results, then anatomical structure is a genuine and substantial driver of dataset-origin classification, not a residual nuisance. The patch/pixel shuffle ablations (Table 3) demonstrate that local texture is sufficient to retain high accuracy (patch shuffle: 97.81%), but they do not show that structure is irrelevant, nor do they establish which cue is "primarily" responsible. To support the primacy claim, the authors would need a condition that removes anatomy while preserving texture (or an explicit attribution analysis); as it stands, the conclusion overreaches the experiments.
  2. [Section 3.3] All four transformed image types (cropped, semantic, contour, lung-and-heart) are generated with the same TorchXRayVision/PSNet segmentation model, but the paper does not disclose PSNet's training data or its segmentation behavior across the four datasets. If PSNet has been trained on some of these datasets, or if it exhibits systematic dataset-specific segmentation biases (e.g., differing organ-size estimates or edge artifacts due to input resolution or source distribution), then the semantic and contour images inherit that fingerprint. In that case, the reported high F1 on contour images may partly measure the segmentation model rather than the original X-ray datasets. This is load-bearing for the claim that anatomical structure is not the primary driver. The authors should report PSNet's training distribution, and ideally validate the transformations on a subset with manual or independent segmentations, or use multiple segmentation models to show robustness.
  3. [Tables 1 and 3] All reported F1 scores are single-run point estimates without confidence intervals, repeated seeds, or multiple patient-level splits. The interpretation of the paper hinges on comparisons between conditions—for example, semantic (86.59%) versus contour (88.62%) versus cropped (99.28%) in Table 1, and patch-shuffle versus pixel-shuffle in Table 3. Without an estimate of run-to-run variability, it is impossible to judge whether the differences between conditions are meaningful, and the single patient split leaves open the possibility that the results are specific to one data partition. The authors should report mean and standard deviation over at least several training runs (and ideally multiple patient splits), and where appropriate perform statistical tests on the differences. This is particularly important because the paper uses these comparisons to argue that intensity/texture, rather than structure, are the primary bias drivers.
minor comments (5)
  1. [Section 1] The sentence "The distributions are very similar in Figure 1, showing that models should struggle to predict the dataset origin, Will this make the task more difficult?" is a run-on with an unnecessary capitalization; please rephrase for clarity.
  2. [Section 3.1] The text states that PadChest images were converted from DICOM by TorchXRayVision, but it is unclear whether the other datasets were already in a standard raster format or whether any additional preprocessing (e.g., windowing, orientation correction) was applied. Please specify the exact preprocessing steps for each dataset to improve reproducibility.
  3. [Figure 5] The y-axis label reads "Percentage of Pixels (%)", but the caption and Section 4.2 refer to the "mean pixel percentage per class". Please clarify that this is the fraction of image pixels occupied by each segmentation-mask class, not a class-label frequency.
  4. [Table 3] The column headers "Transform Cropped Result LH Result" are ambiguous; please label them explicitly as "Cropped Images" and "Lung & Heart Images" so the table is self-contained.
  5. [Abstract] The GitHub URL in the abstract contains spaces ("x ray ds bias") and will not hyperlink correctly; it should be "https://github.com/eedack01/x_ray_ds_bias".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports empirical classification benchmarks against external datasets, and its interpretive conclusions follow from ablation experiments rather than from any fitted parameter or self-citation stated as the result.

full rationale

The paper's contribution is an empirical measurement campaign, not a derivation. The central numbers in Table 1 (e.g., ResNet50 at 99.28% F1 on cropped images, 88.62% on contours, 98.96% on lung-and-heart images) are obtained by training standard classifiers on a four-way dataset-origin task with patient-level train/test splits and reporting held-out F1 scores. There is no fitted parameter that is later renamed as a prediction, and no quantity is defined in terms of another quantity it is then said to predict. The interpretive claim that 'dataset bias is primarily driven by pixel intensity and texture rather than anatomical structure' is an inference drawn from ablation results (Section 4.4, Table 3), where patch shuffling retains 97.81% F1 and pixel shuffling drops to 62.52%; this is a measured contrast, not a conclusion built into the training objective. The use of a single TorchXRayVision/PSNet segmentation model to generate all transformed inputs is a potential validity limitation, because the segmentation model could itself encode dataset-specific signatures; however, this is a confound or correctness concern, not circular reasoning in the sense of the analysis. The paper also does not rely on any load-bearing self-citation: its references to prior work, including [35] and [61], motivate the experimental design but do not substitute for the reported measurements. The mask-statistics analysis (Table 2, Figure 5) is descriptive and used to suggest explanations, not to derive the accuracies. Accordingly, no circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The central claim is empirical, so no fitted constants or invented entities are needed. The main unstated assumptions are about metadata reliability, the neutrality of the segmentation tool, and the sufficiency of preprocessing. These are all ordinary domain assumptions for this kind of study, but the segmentation-tool assumption deserves explicit testing.

assumptions (3)
  • domain assumption Patient identifiers in the metadata are reliable, so the 80:20 patient-level split prevents the same patient from appearing in both train and test.
    Invoked in Section 3.3 Experiment settings. If patient IDs are duplicated or error-prone, model accuracy could reflect patient identity rather than dataset origin.
  • domain assumption The PSNet/TorchXRayVision segmentation model produces masks that capture anatomy without adding dataset-specific signatures.
    Invoked in Section 3.3; all semantic, contour, and lung/heart transformations are generated with this model, so its behavior is baked into the stimuli.
  • domain assumption Z-score normalization per image plus resizing to (512,512), followed by training at (224,224), removes trivial format and resolution differences between datasets.
    Invoked in Sections 3.1 and 3.3; Figure 1 shows distributions become similar after resizing, but residual intensity differences are exactly what the ablation later identifies as the bias signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding Dataset Bias in Medical Imaging: A Case Study on Chest X-rays." pith.science (2026). https://pith.science/paper/A2UJ3QVU

@misc{pith2026250707722,
  author       = {Pith},
  title        = {Pith review of: Understanding Dataset Bias in Medical Imaging: A Case Study on Chest X-rays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2UJ3QVU}},
  note         = {Machine review of arXiv:2507.07722}
}
read the original abstract

Recent works have revisited the infamous task ``Name That Dataset'', demonstrating that non-medical datasets contain underlying biases and that the dataset origin task can be solved with high accuracy. In this work, we revisit the same task applied to popular open-source chest X-ray datasets. Medical images are naturally more difficult to release for open-source due to their sensitive nature, which has led to certain open-source datasets being extremely popular for research purposes. By performing the same task, we wish to explore whether dataset bias also exists in these datasets. To extend our work, we apply simple transformations to the datasets, repeat the same task, and perform an analysis to identify and explain any detected biases. Given the importance of AI applications in medical imaging, it's vital to establish whether modern methods are taking shortcuts or are focused on the relevant pathology. We implement a range of different network architectures on the datasets: NIH, CheXpert, MIMIC-CXR and PadChest. We hope this work will encourage more explainable research being performed in medical imaging and the creation of more open-source datasets in the medical domain. Our code can be found here: https://github.com/eedack01/x_ray_ds_bias.

Figures

Figures reproduced from arXiv: 2507.07722 by the authors.

Figure 1
Figure 1. We computed the distributions of selected datasets be [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Choosing a random image from each dataset, we can [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. As in [61], we apply various transformations to the sampled images, selecting two random images from each dataset for each transformation. The first row shows the cropped images with black patches; the second row displays the corresponding semantic maps; the third row shows the contours; and the final row presents the combined lung and heart images alone. (grayscale, e.g., for X-ray images) instead of the original 3… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Using the same images from Figure [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: We computed the mean pixel percentage per class in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Likewise, as in the first heatmap figure, we pass the same images through their corresponding best-performing model and generate [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 65 canonical work pages

  1. [1]

    Arias-Garz ´on, R

    D. Arias-Garz ´on, R. Tabares-Soto, J. Bernal-Salcedo, et al. Biases associated with database structure for covid-19 detec- tion in x-ray images. Scientific Reports, 13:3477, 2023. 3

    work page 2023

  2. [2]

    The effects of regularization and data augmentation are class de- pendent, 2022

    Randall Balestriero, Leon Bottou, and Yann LeCun. The effects of regularization and data augmentation are class de- pendent, 2022. 8

    work page 2022

  3. [3]

    Deep Learning Models May Spuriously Classify Covid-19 from X-ray Images Based on Confounders

    Kaoutar Ben Ahmed, Lawrence O. Hall, Dmitry B. Goldgof, Gregory M. Goldgof, and Rahul Paul. Deep learning models may spuriously classify covid-19 from x-ray images based on confounders. arXiv preprint arXiv:2102.04300, 2021. 3

    work page Pith review arXiv 2021

  4. [4]

    "Name that manufacturer". Relating image acquisition bias with task complexity when training deep learning models: experiments on head CT

    Giorgio Pietro Biondetti and et al. ”name that manufacturer”: Relating image acquisition bias with task complexity when training deep learning models: experiments on head ct.arXiv preprint arXiv:2008.08525, 2020. 2

    work page Pith review arXiv 2008

  5. [5]

    G. Bradski. The OpenCV Library. Dr. Dobb’s Journal of Software Tools, 2000. 5

    work page 2000

  6. [6]

    Aurelia et al. Bustos. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical Image Analysis, 66:101797, 2020. 1, 2

    work page 2020

  7. [7]

    Jorge Cardoso and Wenqi Li et al

    M. Jorge Cardoso and Wenqi Li et al. Monai: An open- source framework for deep learning in healthcare, 2022. 4, 8

    work page 2022

  8. [8]

    O. D. T. Catala, I. S. Igual, F. J. Perez-Benito, and et al. Bias analysis on public x-ray image datasets of pneumonia and covid-19 patients. IEEE Access, 9:42370–42383, 2021. 3

    work page 2021

Show all 66 references
  1. [9]

    Sources of bias in artificial in- telligence that perpetuate healthcare disparities—a global re- view

    Leo Anthony Celi and et al. Sources of bias in artificial in- telligence that perpetuate healthcare disparities—a global re- view. PLOS Digital Health, 1(3):e0000022, 2022. 2

    2022

  2. [10]

    Viviano, Paul Bertin, Paul Morrison, Parsa Torabian, Matteo Guarrera, Matthew P Lungren, Akshay Chaudhari, Rupert Brooks, Mohammad Hashir, and Hadrien Bertrand

    Joseph Paul Cohen, Joseph D. Viviano, Paul Bertin, Paul Morrison, Parsa Torabian, Matteo Guarrera, Matthew P Lungren, Akshay Chaudhari, Rupert Brooks, Mohammad Hashir, and Hadrien Bertrand. TorchXRayVision: A library of chest X-ray datasets and models. InMedical Imaging with D...

    2022

  3. [11]

    Joseph Paul et al. Cohen. Radiographic assessment of lung opacity score dataset, 2021. 2

    2021

  4. [12]

    Re- producibility in machine learning for medical imaging

    Olivier Colliot, Elina Thibeau-Sutre, and Ninon Burgos. Re- producibility in machine learning for medical imaging. In Machine Learning for Brain Disorders. 2023. 1

    2023

  5. [13]

    Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V

    Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V . Le. Randaugment: Practical automated data augmenta- tion with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 702–703, 2020. 7

    2020

  6. [14]

    Omar Adil Deheyab and et al

    A. Omar Adil Deheyab and et al. Future networks and dis- tributed systems: Innovations and challenges. In Proceed- ings of the 6th International Conference on Future Networks and Distributed Systems (ICFNDS ’22). ACM, 2022. 1

    2022

  7. [15]

    Kohli, Marc B

    Dina Demner-Fushman, Marc D. Kohli, Marc B. Rosen- man, Sonya E. Shooshan, Laritza Rodriguez, Sameer Antani, George R. Thoma, and Clement J. McDonald. Preparing a collection of radiology examinations for distribution and re- trieval. Journal of the American Medical Informatics...

    2016

  8. [16]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3

    2009

  9. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  10. [18]

    A survey on bias in visual datasets

    Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding , 223: 103552, 2022. 2

    2022

  11. [19]

    Torchcam: class activation explorer

    Franc ¸ois-Guillaume Fernandez. Torchcam: class activation explorer. https://github.com/frgfm/torch- cam, 2020. 7

    2020

  12. [20]

    Ross W. et al. Filice. Crowdsourcing pneumothorax anno- tations using machine learning annotations on the nih chest x-ray dataset. Journal of Digital Imaging, 2020. 2

    2020

  13. [21]

    Public covid-19 x-ray datasets and their impact on model bias – a systematic review of a significant problem

    Beatriz Garcia Santa Cruz, Mat ´ıas Nicol´as Bossa, Jan S¨olter, and Andreas Dominik Husch. Public covid-19 x-ray datasets and their impact on model bias – a systematic review of a significant problem. Medical Image Analysis , 74:102225,

  14. [22]

    Wichmann, and Wieland Brendel

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 8

    2022

  15. [23]

    Risk of bias in chest radiography deep learning foundation models

    Ben Glocker and et al. Risk of bias in chest radiography deep learning foundation models. Radiology: Artificial In- telligence, 5(6):e230060, 2023. 2

    2023

  16. [24]

    Curran, and Brian Mac Namee

    Misgina Tsighe Hagos, Kathleen M. Curran, and Brian Mac Namee. Unlearning spurious correlations in chest x-ray clas- sification. In Discovery Science, pages 361–375. Springer Nature Switzerland, 2023. 3

    2023

  17. [25]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2015. 4

    2015

  18. [26]

    Shen, George Shih, Ronald M

    Gregory Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax dis- eases on chest x-ray: A new benchmark study. In Medi- cal Image Computing and Computer-Assisted Intervention – MICCA...

    2022

  19. [27]

    Jeremy et al. Irvin. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI Conference on Artificial Intelli- gence, 2019. 1, 2

    2019

  20. [28]

    W´ang, Pu-Xuan Lu, and George Thoma

    Stefan Jaeger, Sema Candemir, Sameer Antani, Y `ı-Xi´ang J. W´ang, Pu-Xuan Lu, and George Thoma. Two public chest x- ray datasets for computer-aided screening of pulmonary dis- eases. Quantitative Imaging in Medicine and Surgery, 2014. Jaeger et al. 2

    2014

  21. [29]

    Alistair E. W. et al. Johnson. MIMIC-CXR: A De-identified Publicly Available Database of Chest Radiographs with Free-text Reports. Scientific Data, 6:317, 2019. 1, 2

    2019

  22. [30]

    Castro, Filipe De Sousa Ribeiro, et al

    Christopher Jones, Daniel C. Castro, Filipe De Sousa Ribeiro, et al. A causal perspective on dataset bias in machine learning for medical imaging. Nature Machine Intelligence, 6:138–146, 2024. 3

    2024

  23. [31]

    Efros, Antonio Torralba, et al

    Aditya Khosla, Tinghui Zhou, Tomasz Malisiewicz, Alexei A. Efros, Antonio Torralba, et al. Undoing the dam- age of dataset bias. In Computer Vision – ECCV 2012, pages 158–171, Berlin, Heidelberg, 2012. Springer Berlin Heidel- berg. 2

    2012

  24. [32]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. In Advances in Neural Information Processing Sys- tems. Curran Associates, Inc., 2012. 3

    2012

  25. [33]

    Falet, Sotirios Tsaftaris, and Tal Arbel

    Amar Kumar, Nima Fathi, Raghav Mehta, Brennan Nichy- poruk, Jean-Pierre R. Falet, Sotirios Tsaftaris, and Tal Arbel. Debiasing counterfactuals in the presence of spurious corre- lations. In Clinical Image-Based Procedures, Fairness of AI in Medical Imaging, and Ethical and Phi...

    2023

  26. [34]

    A Structure-Aware Rela- tion Network for Thoracic Diseases Detection and Segmen- tation

    Jie Lian, Jingyu Liu, Shu Zhang, Kai Gao, Xiaoqing Liu, Dingwen Zhang, and Yizhou Yu. A Structure-Aware Rela- tion Network for Thoracic Diseases Detection and Segmen- tation. IEEE Transactions on Medical Imaging, 2021. 5

    2021

  27. [35]

    A decade’s battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632,

    Zhuang Liu and Kaiming He. A decade’s battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632,

    arXiv

  28. [36]

    Steiner, Joshua J

    Anna Majkowska, Sid Mittal, David F. Steiner, Joshua J. Reicher, Scott Mayer McKinney, Gavin E. Duggan, Kr- ish Eswaran, Po-Hsuan Cameron Chen, Yun Liu, Sreeni- vasa Raju Kalidindi, Alexander Ding, Greg S. Corrado, Daniel Tse, and Shravya Shetty. Chest radiograph in- terpretat...

    2019

  29. [37]

    Mosquera and et al

    C. Mosquera and et al. Class imbalance on medical image classification: Towards better evaluation practices for dis- crimination and calibration performance. European Radi- ology, 34:7895–7903, 2024. 2

    2024

  30. [38]

    Intriguing properties of vision transformers, 2021

    Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers, 2021. 8

    2021

  31. [39]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis- sacco, Bo Wu, and Andrew Y . Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Work- shop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 2

    2011

  32. [40]

    Nguyen, Khanh Lam, Linh T

    Ha Q. Nguyen, Khanh Lam, Linh T. Le, Hieu H. Pham, Dat Q. Tran, Dung B. Nguyen, Dung D. Le, Chi M. Pham, Hang T. T. Tong, Diep H. Dinh, Cuong D. Do, Luu T. Doan, Cuong N. Nguyen, Binh T. Nguyen, Que V . Nguyen, Au D. Hoang, Hien N. Phan, Anh T. Nguyen, Phuong H. Ho, Dat T. Ngo...

    2020

  33. [41]

    How do vision transformers work?, 2022

    Namuk Park and Songkuk Kim. How do vision transformers work?, 2022. 8

    2022

  34. [42]

    Hauglid, Marzia Fumagalli, Lena Petersson, Anagha P

    Filippo Pesapane, Mathias K. Hauglid, Marzia Fumagalli, Lena Petersson, Anagha P. Parkar, Enrico Cassano, and De- nis Horgan. The translation of in-house imaging ai research into a medical device ensuring ethical and regulatory in- tegrity. European Journal of Radiology, 182:1...

    2025

  35. [43]

    Ritor´e and et al

    ´A. Ritor´e and et al. The role of open access data in democ- ratizing healthcare ai: A pathway to research enhancement, patient well-being and treatment equity in andalusia, spain. PLOS Digital Health, 3(9):e0000599, 2024. 1

    2024

  36. [44]

    Common pitfalls and recommen- dations for using machine learning to detect and prognosti- cate for covid-19 using chest radiographs and ct scans

    Michael Roberts and et al. Common pitfalls and recommen- dations for using machine learning to detect and prognosti- cate for covid-19 using chest radiographs and ct scans. Na- ture Machine Intelligence, 3(3):199–217, 2021. 1

    2021

  37. [45]

    Mitigating bias in radiology ma- chine learning: 1

    Parsa Rouzrokh and et al. Mitigating bias in radiology ma- chine learning: 1. data handling. Radiology: Artificial Intel- ligence, 4(5):e210290, 2022. 2

    2022

  38. [46]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE In- ternational Conference on Computer Vision (ICCV) , page...

    2017

  39. [47]

    George et al. Shih. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence ,

  40. [48]

    Herbert A. Simon. Spurious correlation: A causal interpreta- tion. In Models of Discovery: And Other Topics in the Meth- ods of Science , pages 93–106. Springer Netherlands, Dor- drecht, 1977. 3

    1977

  41. [49]

    A. M. Speets, Y . van der Graaf, A. W. Hoes, et al. Chest radiography in general practice: indications, diagnostic yield and consequences for patient management. British Journal of General Practice, 56(529):574–578, 2006. 6

    2006

  42. [50]

    How to train your vit? data, augmentation, and regularization in vision transformers, 2022

    Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers, 2022. 4, 6

    2022

  43. [51]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1– ...

    2015

  44. [52]

    Expert-level de- tection of pathologies from unannotated chest x-ray images via self-supervised learning

    Edward Tiu, Evan Talius, Pranav Patel, et al. Expert-level de- tection of pathologies from unannotated chest x-ray images via self-supervised learning. Nature Biomedical Engineer- ing, 2022. 2, 4

    2022

  45. [53]

    A Deeper Look at Dataset Bias, pages 37–

    Tatiana Tommasi, Novi Patricia, Barbara Caputo, Tinne Tuytelaars, et al. A Deeper Look at Dataset Bias, pages 37–

  46. [54]

    Antonio Torralba and Alexei A. Efros. Unbiased look at dataset bias. In CVPR 2011 , pages 1521–1528, 2011. 1, 2

    2011

  47. [55]

    Springer International Publishing, Cham, 2017. 2

    2017

  48. [56]

    Chen, Drew F

    Abhishek Vaidya, Richard J. Chen, Drew F. K. Williamson, et al. Demographic bias in misdiagnosis by computational pathology models. Nature Medicine, 30:1174–1190, 2024. 3

    2024

  49. [57]

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

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mo- hammadhadi Bagheri, and Ronald M Summers. Chestx- ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of com- mon thorax diseases. In Proceedings of the IEEE Conference ...

    2017

  50. [58]

    Delving into masked autoencoders for multi-label thorax dis- ease classification

    Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax dis- ease classification. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 3588–3600, 2023. 4

    2023

  51. [59]

    Demographic bias of expert- level vision-language foundation models in medical imaging

    Yuzhe Yang, Yujia Liu, Xin Liu, Avanti Gulhane, Domenico Mastrodicasa, Wei Wu, Edward J Wang, Dushyant W Sa- hani, and Shwetak Patel. Demographic bias of expert- level vision-language foundation models in medical imaging. arXiv preprint arXiv:2402.14815, 2024. 3

    2024 arXiv

  52. [60]

    Gichoya, et al

    Yufeng Yang, Hao Zhang, Judy W. Gichoya, et al. The lim- its of fair medical imaging ai in real-world generalization. Nature Medicine, 30:2838–2848, 2024. 3

    2024

  53. [61]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE International Confer- ence on Computer Vision (ICCV), 2019. 7

    2019

  54. [62]

    Understanding bias in large-scale visual datasets

    Boya Zeng, Yida Yin, and Zhuang Liu. Understanding bias in large-scale visual datasets. In The Thirty-eighth An- nual Conference on Neural Information Processing Systems,

  55. [63]

    Dauphin, and David Lopez-Paz

    Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. In International Conference on Learning Representa- tions (ICLR), 2018. 7

    2018

  56. [64]

    Learning deep features for discrimi- native localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discrimi- native localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921–2929, 2016. 7 Understanding Dataset Bias ...

    2016

  57. [65]

    As a comparison, we now generate heatmaps by averaging the Grad-CAMs computed across all model layers of the best-performing model

    Average Heatmaps In the previous heatmap figure, the Grad-CAM was gener- ated using only the final activation layer. As a comparison, we now generate heatmaps by averaging the Grad-CAMs computed across all model layers of the best-performing model. Looking at Figure 6, while t...

  58. [66]

    Our experiments evaluate these transformations using two distinct application probabilities: P = 0.2 (conservative) and P = 0.5 (aggressive)

    Transformations We utilise the MONAI library to implement 13 care- fully selected augmentation transformations designed to en- hance model robustness through controlled variations in pixel intensity and texture. Our experiments evaluate these transformations using two distinct...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.