CF-Seg: Counterfactuals meet Segmentation

Ainkaran Santhirasekaram; Ben Glocker; Dominic C. Marshall; Fabio De Sousa Ribeiro; Melanie Roschewitz; Raghav Mehta; Tian Xia

REVIEW 4 major objections 5 minor 35 references

CF-Seg: Counterfactuals meet Segmentation

T0 review · 4 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 The paper claims that generating counterfactual pseudo-healthy chest X-rays before segmentation improves lung segmentation substantially in the presence of pleural effusion, without any change to the segmentation model.

desk verdict A clean empirical application of counterfactual pre-processing to lung segmentation, with real expert evaluation, but the core anatomy-preservation assumption of the CF generator is unverified and the large-scale proxy is too weak to carry the causal claim. read the letter →

arxiv 2506.16213 v1 pith:JVRVBSKM submitted 2025-06-19 eess.IV cs.AIcs.CV

Raghav Mehta , Fabio De Sousa Ribeiro , Tian Xia , Melanie Roschewitz , Ainkaran Santhirasekaram , Dominic C. Marshall , Ben Glocker This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords counterfactualimagesanatomicalsegmentationchestX-raypleuraleffusiondeepstructuralcausalmodelspseudo-healthyimagegenerationlunginference-timepreprocessing

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 proposes that instead of segmenting a diseased chest X-ray directly, one first generates a counterfactual pseudo-healthy version of the image, showing the same lungs as if the pleural effusion were absent, and then runs an off-the-shelf segmenter on that cleaned image. It claims this improves anatomical lung segmentation substantially in the presence of pleural effusion while leaving the segmentation model untouched. On two real clinical chest X-ray datasets, radiologists preferred the counterfactual-based segmentation over the silver-standard masks for more than 70% of effusion cases, and expert-annotated ground truth confirms higher Dice scores. The significance is that disease-induced opacity is treated as a removable nuisance rather than something the segmenter must learn to see through.

What carries the argument

The central machinery is a deep structural causal model (DSCM), a causal graph whose structural equations are neural networks, implemented here as a hierarchical variational autoencoder. The generator abducts the latent noise from the observed image, performs the do-intervention that sets disease status to healthy, and predicts the hypothetical image that would result, thereby removing effusion while keeping the same anatomy. This component carries the argument because it is what supposedly produces a clearer anatomical representation; the U-Net segmenter is unchanged and is applied only to the counterfactual image.

What would settle it

Compare the pseudo-healthy counterfactual of a healthy patient with the original healthy image; if the generator changes lung boundaries or volumes even when no disease is present, then the counterfactual does not preserve anatomy as claimed and the segmentation improvement cannot be attributed to disease removal.

Watch

Extended reading notes

Core claim

The central discovery is that a deep structural causal model's counterfactual intervention, setting the disease variable to healthy, removes pleural effusion from a chest X-ray while preserving the underlying lung anatomy, and feeding this pseudo-healthy image to an existing U-Net yields lung segmentations closer to radiologist ground truth than segmenting the original image. The improvement is concentrated in the right lung and in cases where the silver-standard masks undersegment, consistent with effusion obscuring lung boundaries. The paper also reports population-level evidence: lung volume distributions for healthy and effusion patients overlap when measured from CF-Seg segmentations but not from direct segmentations, indicating that the counterfactual pipeline recovers anatomy rather than merely shifting volumes.

Load-bearing premise

The counterfactual generator removes pleural effusion while preserving the exact lung boundaries; if the pseudo-healthy image subtly changes lung shape or size, the reported segmentation gains could be an artifact of altered anatomy rather than clearer anatomy.

Editorial extensions

If this is right

For images with pleural effusion, segmenting the counterfactual instead of the original image improves Dice against expert ground truth, with the largest gain on undersegmented right lungs.
Radiologists prefer CF-Seg segmentations over silver-standard masks for more than 70% of effusion images, while preferences are roughly balanced for healthy images.
Population-level lung volume distributions of healthy and effusion patients overlap when measured with CF-Seg, suggesting the method recovers true anatomy rather than randomly rescaling volumes.
No retraining or architectural change is needed, so the counterfactual step is a drop-in preprocessing module that can be combined with any pretrained segmenter.
The authors note that other opacity-causing conditions, such as edema, ARDS, or pneumonia, could in principle be handled the same way given a suitable counterfactual generator.

Reading between the lines

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

If the anatomy-preservation assumption holds, counterfactual-first preprocessing could generalize beyond segmentation to any task where disease-related appearance variation confounds an otherwise well-trained model.
The gains likely depend on the causal graph capturing confounding variables such as scanner and sex, so deployment to a site with different scanner or population characteristics could degrade the benefit.
A direct extension would test the same pipeline on other structures and pathologies, since the paper evaluates only lungs and pleural effusion.
The volume-density overlap evidence is indirect; a direct test would measure whether CF-Seg's improvement comes from boundary correction rather than systematic volume bias.

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.

Desk Editor's Note

read the letter

Dear colleague,

The thing to know: this is a clean, honest empirical application paper, not a methodological breakthrough. It takes an existing counterfactual generator for chest X-rays, feeds the pseudo-healthy image into an off-the-shelf U-Net, and shows that on pleural-effusion cases you get better lung segmentations than segmenting the original image. The authors are careful to note the generator is not new; the contribution is the application and the evaluation.

What's genuinely good: the evaluation includes expert segmentations on 140 images and a blinded preference study with two radiologists on 600 images total. On PE images, radiologists preferred the CF-Seg masks over CheXMask about 70% of the time, and the Dice gains on the expert subset—though modest (1–2 points overall, larger on the worst cases)—are consistent across two datasets and both lungs. The volume-distribution analysis, at least on the expert set, also lines up with what you'd expect if CF-Seg were closer to expert anatomy.

The soft spots are real and mostly center on one untested assumption: that the counterfactual removes effusion while preserving lung boundaries. That's the load-bearing claim in the intro, and the paper never verifies it. They don't measure boundary displacement between original and CF images, and they don't ask experts to segment the CF images directly. The preference study is strong, but it compares masks, not anatomical truth. The small expert-Dice set is the only direct evidence, and it has no error bars or significance tests, so the 'substantial' improvement in the abstract is not statistically supported. The large-scale proxy in Sec. 4.5 is weak: assuming NF and PE populations have identical lung-volume densities is unfounded, and the test would be passed by any method that makes PE images look like NF images, regardless of anatomical accuracy. That proxy should not be used to infer the causal mechanism.

There are also two practical issues: no code or data is available yet (the paper says it 'will be'), and the causal graph for PadChest adds scanner and sex as independent parents without validation. Neither is fatal, but code is important given the dependence on a specific pretrained generator.

All that said, the paper deserves a serious referee. The clinical question is relevant, the experiments are honest in their limitations, and the core idea is worth testing further. I'd accept it for peer review, but I'd insist on an anatomy-preservation validation—paired expert annotations on original vs. CF images, or at least a boundary-displacement metric—and proper statistical reporting. As is, the conclusion is plausible but not proven.

Referee Report

4 major / 5 minor

Summary. The paper proposes CF-Seg, an inference-time framework that first generates pseudo-healthy counterfactual (CF) images of chest X-rays with pleural effusion using a deep structural causal model (the HVAE of Ribeiro et al. [5]), and then feeds these images to a standard U-Net for lung segmentation. The central claim is that segmenting the counterfactual image improves anatomical lung segmentation without retraining the segmentation model. The authors evaluate CF-Seg on MIMIC-CXR and PadChest through (i) a radiologist preference study on 300 images per dataset, (ii) expert segmentations on 140 images (50 pleural-effusion + 20 no-finding per dataset) with Dice-score comparisons, and (iii) a large-scale proxy that compares lung-volume distributions between healthy and effusion populations.

Significance. If the anatomy-preservation property of the counterfactual generator holds, this is a simple and attractive contribution: it converts a diseased image into a form that a pre-trained segmentor handles better, with no model changes, making it compatible with arbitrary segmentation networks. The paper includes a radiologist preference study and expert-annotated Dice comparisons, which are valuable clinical evidence for the approach. The reliance on an existing generator and the restriction to one pathology limit the scope, but as a proof-of-concept the result is compelling. I also credit the authors for committing to release code and segmentation masks, which will aid reproducibility and follow-up work.

major comments (4)

[Sec. 3.1-3.2, Fig. 5] The causal mechanism asserted in the central claim—that the counterfactual removes pleural effusion while preserving the underlying lung anatomy—is never directly verified. The paper does not measure lung-boundary displacement between the original and CF images, nor does it obtain expert segmentations on the CF images to confirm they agree with expert segmentations on the original images. The indirect evidence in Table 1 could in principle arise from a CF image that simply looks healthy: the U-Net was trained on CheXMask silver labels that undersegment PE lungs, so a healthier-looking image could produce a larger mask that scores higher against expert masks even if the CF boundary deviates from the true boundary of the effused lung. Please add a direct anatomy-preservation test, e.g., compare radiologist segmentations drawn on CF images against those on the original images, or compute a boundary-distance metric (such as Hausdorff distance) between segmentations on original and CF images to confirm that changes are confined to regions where effusion was present.
[Sec. 4.4, Table 1] All Dice results are reported as means without error bars, confidence intervals, or paired significance tests. With n=50 PE images per dataset (and roughly 35 in the ΔV+ subgroup), the observed gains of 0.5–2.7 Dice points, while consistent, may not be statistically significant, especially for the 'All' columns. Please report per-image variability and run a paired test (e.g., Wilcoxon signed-rank) between U-Net and CF-Seg on the same images. Additionally, the ΔV+ selection criterion and the Dice computation use the same expert masks, so the subgroup result should be accompanied by an appropriate caveat or by a clear pre-specified definition.
[Sec. 4.5, Fig. 6] The large-scale evaluation relies on the assumption that healthy (NF) and pleural-effusion (PE) populations have identical lung-volume densities. This assumption is not established and could be false: pleural effusion may compress or collapse lung tissue, shifting the true volume distribution. The KDE overlap in Fig. 6 is only a qualitative visual comparison; no quantitative divergence metric is reported. Please validate the equal-volume assumption on the expert-annotated subset (e.g., compare Expert NF vs Expert PE volumes), or reframe the claim to state only that CF-Seg makes the PE volume distribution resemble the NF distribution, without asserting that this is anatomically correct.
[Sec. 4.3, Fig. 3] The preference study compares CF-Seg against CheXMask labels, not against the same U-Net applied to the original image. Since CF-Seg is literally the U-Net run on the counterfactual, the preference results could be driven by the U-Net's output characteristics rather than by the counterfactual intervention. Please include a U-Net-on-original arm in the user study (or replace the CheXMask baseline with it) so that the effect of the counterfactual is isolated.

minor comments (5)

[Sec. 4.2] The name 'Riberio' appears in the text; the correct spelling is 'Ribeiro' (as in the reference list).
[Fig. 2] The schematic is unclear on which components are frozen or pre-trained; please annotate the DSCM and U-Net blocks more explicitly for reproducibility.
[Table 1] Define 'All' and 'ΔV+' in the caption; currently they appear without explanation in the table.
[Sec. 4.5] The sentence 'We hypothesize that irrespective of the disease stage...' should be explicitly framed as an unverified assumption, and the argument should acknowledge that this hypothesis is not tested.
[Sec. 4.3] Please clarify the exact number of radiologists per dataset and the total number of images each radiologist reviewed; the text says 'one for each dataset' and '300 test images for each dataset', but it would help to state this explicitly in the figure caption or text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: CF-Seg's segmentation gains are tested against external expert annotations and radiologist preference, not derived from the counterfactual generator's own outputs.

full rationale

The paper's central claim is empirical: applying pseudo-healthy counterfactuals before segmentation improves anatomical segmentation. This claim is validated in Sec. 4.3 by radiologist preference and in Sec. 4.4 against expert-drawn ground-truth segmentations, both of which are external to the counterfactual generator and to the U-Net's training labels. The only self-citation is the use of the HVAE/DSCM from Ribeiro et al. [5] as an off-the-shelf counterfactual generator. The paper does not fit any parameter of that generator to the expert segmentations, nor does it define its evaluation metric in terms of the generator's outputs. The assumption that counterfactuals preserve anatomy is inherited from [5] and is not independently verified here, but this is a validity assumption rather than a circular reduction: the segmentation improvement is measured against external masks and human judgment. The large-scale proxy in Sec. 4.5 is presented as an explicit hypothesis about population lung volumes, not as a restatement of the method's objective. No equation or fitted parameter reduces to the claimed result, and no load-bearing uniqueness theorem or ansatz is smuggled in via self-citation. Therefore the derivation chain is not circular; the observed improvements are independently testable against expert annotations.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the validity of the counterfactual generator and on the proxy assumption used for large-scale evaluation. These are domain assumptions imported from prior work or stated as hypotheses, not established in this paper.

free parameters (1)

PadChest causal graph variables = hand-chosen (scanner, sex, disease)
In Sec 4.2, the authors state 'for PadChest, we consider three variables scanner, sex and disease, assuming they are independent.' This is a modeling choice that affects counterfactual generation, but no fitting procedure is described.

assumptions (3)

domain assumption Counterfactual images generated by the HVAE preserve underlying anatomy while removing disease appearance.
Stated in Sec 1 ('the underlying lung structure should remain unchanged') and assumed throughout; not verified in this paper.
domain assumption Lung volume density across the population is similar between no-finding and pleural-effusion groups.
Used in Sec 4.5 as a proxy to evaluate CF-Seg on large test sets. The hypothesis is stated but not validated.
domain assumption The causal graph and exogenous noise model of the DSCM from [5] are correctly specified for chest X-rays.
The paper relies on the causal model from [5] without re-validation; for PadChest they assume scanner, sex, and disease are independent causes of the image.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CF-Seg: Counterfactuals meet Segmentation." pith.science (2026). https://pith.science/paper/JVRVBSKM

@misc{pith2026250616213,
  author       = {Pith},
  title        = {Pith review of: CF-Seg: Counterfactuals meet Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JVRVBSKM}},
  note         = {Machine review of arXiv:2506.16213}
}

read the original abstract

Segmenting anatomical structures in medical images plays an important role in the quantitative assessment of various diseases. However, accurate segmentation becomes significantly more challenging in the presence of disease. Disease patterns can alter the appearance of surrounding healthy tissues, introduce ambiguous boundaries, or even obscure critical anatomical structures. As such, segmentation models trained on real-world datasets may struggle to provide good anatomical segmentation, leading to potential misdiagnosis. In this paper, we generate counterfactual (CF) images to simulate how the same anatomy would appear in the absence of disease without altering the underlying structure. We then use these CF images to segment structures of interest, without requiring any changes to the underlying segmentation model. Our experiments on two real-world clinical chest X-ray datasets show that the use of counterfactual images improves anatomical segmentation, thereby aiding downstream clinical decision-making.

Figures

Figures reproduced from arXiv: 2506.16213 by the authors.

**Figure 2.** Overview of the proposed CF-Seg framework. Instead of directly segment [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Preference study results. Experts were shown CheXMask and our CF [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: For (Top) MIMIC and (Bottom) PadChest, density plots of lung volumes [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Qualitative comparison of lung segmentation masks generated by dif [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Comparison of right lung volume density plots measured by different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

[5]

ICML (2023)

De Sousa Ribeiro, F., Xia, T., Monteiro, M., Pawlowski, N., Glocker, B.: High fidelity image counterfactuals with probabilistic causal models. ICML (2023)

work page 2023
[1]

In: International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging

Baugh, M., Reynaud, H., Marimont, S.N., Cechnicka, S., Müller, J.P., Tarroni, G., Kainz, B.: Image-conditioned diffusion models for medical anomaly detection. In: International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. pp. 117–127. Springer (2024)

work page 2024
[2]

In: MICCAI 2023

Baugh, M., Tan, J., Müller, J.P., Dombrowski, M., Batten, J., Kainz, B.: Many tasks make light work: Learning to localise medical anomalies from multiple syn- thetic tasks. In: MICCAI 2023. pp. 162–172. Springer (2023)

work page 2023
[3]

Medical image anal- ysis 66, 101797 (2020) 10 Mehta et al

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 anal- ysis 66, 101797 (2020) 10 Mehta et al

work page 2020
[4]

In: MIDL

Cohen, J.P., Brooks, R., En, S., Zucker, E., Pareek, A., Lungren, M.P., Chaudhari, A.: Gifsplanation via latent shift: a simple autoencoder approach to counterfactual generation for chest x-rays. In: MIDL. pp. 74–104 (2021)

work page 2021
[6]

MIDL (2024)

Fathi, N., Kumar, A., Nichyporuk, B., Havaei, M., Arbel, T.: Decodex: Confounder detector guidance for improved diffusion-based counterfactual explanations. MIDL (2024)

work page 2024
[7]

Scientific Data11(1), 511 (2024)

Gaggion, N., Mosquera, C., Mansilla, L., Saidman, J.M., Aineseder, M., Milone, D.H., Ferrante, E.: Chexmask: a large-scale dataset of anatomical segmentation masks for multi-center chest x-ray images. Scientific Data11(1), 511 (2024)

2024
[8]

Machine Learning for Biomedical Imaging2, 611–656 (2024)

Hassanaly, R., Brianceau, C., Solal, M., Colliot, O., Burgos, N.: Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain fdg pet. Machine Learning for Biomedical Imaging2, 611–656 (2024)

work page 2024

Show all 35 references

[9]

IEEE Transactions on Medical Imaging40(12), 3519–3530 (2021)

Huang, H., Zheng, H., Lin, L., Cai, M., Hu, H., Zhang, Q., Chen, Q., Iwamoto, Y., Han, X., Chen, Y.W., et al.: Medical image segmentation with deep atlas prior. IEEE Transactions on Medical Imaging40(12), 3519–3530 (2021)

2021
[10]

In: ICML

Ilse, M., Tomczak, J.M., Forré, P.: Selecting data augmentation for simulating interventions. In: ICML. pp. 4555–4562. PMLR (2021)

2021
[11]

IEEE transactions on medical imaging33(2), 233–245 (2013)

Jaeger, S., Karargyris, A., Candemir, S., Folio, L., Siegelman, J., Callaghan, F., Xue, Z., Palaniappan, K., Singh, R.K., Antani, S., et al.: Automatic tuberculosis screening using chest radiographs. IEEE transactions on medical imaging33(2), 233–245 (2013)

2013
[12]

In: MICCAI 2024

Jaus, A., Seibold, C., Reiß, S., Heine, L., Schily, A., Kim, M., Bahnsen, F.H., Her- rmann, K., Stiefelhagen, R., Kleesiek, J.: Anatomy-guided pathology segmentation. In: MICCAI 2024. pp. 3–13. Springer (2024)

2024
[13]

Scientific data10(1), 1 (2023)

Johnson, A.E., Bulgarelli, L., Shen, L., Gayles, A., Shammout, A., Horng, S., Pollard, T.J., Hao, S., Moody, B., Gow, B., et al.: Mimic-iv, a freely accessible electronic health record dataset. Scientific data10(1), 1 (2023)

2023
[14]

Scientific Reports13(1), 20899 (2023)

Khomduean, P., Phuaudomcharoen, P., Boonchu, T., Taetragool, U., et al.: Seg- mentation of lung lobes and lesions in chest ct for the classification of covid-19 severity. Scientific Reports13(1), 20899 (2023)

2023
[15]

ICLR (2017)

Kocaoglu, M., Snyder, C., Dimakis, A.G., Vishwanath, S.: Causalgan: Learning causal implicit generative models with adversarial training. ICLR (2017)

2017
[16]

In: MICCAI Workshop on Clinical Image-Based Procedures

Kumar, A., Fathi, N., Mehta, R., Nichyporuk, B., Falet, J.P.R., Tsaftaris, S., Arbel, T.: Debiasing counterfactuals in the presence of spurious correlations. In: MICCAI Workshop on Clinical Image-Based Procedures. pp. 276–286. Springer (2023)

2023
[17]

IEEE Transactions on Medical Imaging39(12), 3813–3820 (2020)

Larrazabal, A.J., Martínez, C., Glocker, B., Ferrante, E.: Post-dae: anatomically plausible segmentation via post-processing with denoising autoencoders. IEEE Transactions on Medical Imaging39(12), 3813–3820 (2020)

2020
[18]

Scientific Reports12(1), 8649 (2022)

Liu, W., Luo, J., Yang, Y., Wang, W., Deng, J., Yu, L.: Automatic lung segmen- tation in chest x-ray images using improved u-net. Scientific Reports12(1), 8649 (2022)

2022
[19]

In: MICCAI

Müller, P., Meissen, F., Brandt, J., Kaissis, G., Rueckert, D.: Anatomy-driven pathology detection on chest x-rays. In: MICCAI. pp. 57–66. Springer (2023)

2023
[20]

In: Machine Learning in Medical Imaging

Navarro, F., Shit, S., Ezhov, I., Paetzold, J., Gafita, A., Peeken, J.C., Combs, S.E., Menze, B.H.: Shape-aware complementary-task learning for multi-organ segmen- tation. In: Machine Learning in Medical Imaging. pp. 620–627. Springer (2019) CF-Seg: Counterfactuals meet Segmen...

2019
[21]

IEEE Transactions on Medical Imaging37(2), 384–395 (2017)

Oktay, O., Ferrante, E., Kamnitsas, K., Heinrich, M., Bai, W., et al.: Anatomically constrained neural networks (acnns): application to cardiac image enhancement and segmentation. IEEE Transactions on Medical Imaging37(2), 384–395 (2017)

2017
[22]

NeurIPS33, 857–869 (2020)

Pawlowski, N., Coelho de Castro, D., Glocker, B.: Deep structural causal models for tractable counterfactual inference. NeurIPS33, 857–869 (2020)

2020
[23]

Cambridge university press (2009)

Pearl, J.: Causality. Cambridge university press (2009)

2009
[24]

Medical Image Analysis 84, 102723 (2023)

Pombo, G., Gray, R., Cardoso, M.J., Ourselin, S., Rees, G., Ashburner, J., Nachev, P.: Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3d deep generative models. Medical Image Analysis 84, 102723 (2023)

2023
[25]

In: MICCAI

Puglisi, L., Alexander, D.C., Ravì, D.: Enhancing spatiotemporal disease progres- sion models via latent diffusion and prior knowledge. In: MICCAI. pp. 173–183. Springer (2024)

2024
[26]

BMC medical imaging20, 1–13 (2020)

Reamaroon, N., Sjoding, M.W., Derksen, H., Sabeti, E., Gryak, J., Barbaro, R.P., Athey, B.D., Najarian, K.: Robust segmentation of lung in chest x-ray: applications in analysis of acute respiratory distress syndrome. BMC medical imaging20, 1–13 (2020)

2020
[27]

In: MICCAI

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)

2015
[28]

In: MICCAI Workshop on Data Engineering in Medical Imaging

Roschewitz, M., de Sousa Ribeiro, F., Xia, T., Khara, G., Glocker, B.: Counter- factual contrastive learning: robust representations via causal image synthesis. In: MICCAI Workshop on Data Engineering in Medical Imaging. pp. 22–32 (2024)

2024
[29]

Conference on Causal Learning and Reasoning (2022)

Sanchez, P., Tsaftaris, S.A.: Diffusion causal models for counterfactual estimation. Conference on Causal Learning and Reasoning (2022)

2022
[30]

In: ICLR 2023 (2023)

Xia, K.M., Pan, Y., Bareinboim, E.: Neural causal models for counterfactual iden- tification and estimation. In: ICLR 2023 (2023)

2023
[31]

Medical Image Analysis64, 101719 (2020)

Xia,T.,Chartsias,A.,Tsaftaris,S.A.:Pseudo-healthysynthesiswithpathologydis- entanglement and adversarial learning. Medical Image Analysis64, 101719 (2020)

2020
[32]

In: CVPR

Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., Wang, J.: Causalvae: Disentangled representation learning via neural structural causal models. In: CVPR. pp. 9593– 9602 (2021)

2021
[33]

In: MICCAI 2019

Yao, J., Cai, J., Yang, D., Xu, D., Huang, J.: Integrating 3d geometry of organ for improving medical image segmentation. In: MICCAI 2019. pp. 318–326. Springer (2019)

2019
[34]

arXiv:2304.13431 (2023)

Zhou, X., Wu, O., Ng, M.K.: Implicit counterfactual data augmentation for robust learning. arXiv:2304.13431 (2023)

2023 arXiv
[35]

IEEE Transactions on Medical Imaging41(10), 2728–2738 (2022)

Zimmerer, D., Full, P.M., Isensee, F., Jäger, P., Adler, T., Petersen, J., Köhler, G., Ross, T., Reinke, A., Kascenas, A., et al.: Mood 2020: A public benchmark for out- of-distribution detection and localization on medical images. IEEE Transactions on Medical Imaging41(10), 2...

2022

Pith tools

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

[1] [5]

ICML (2023)

De Sousa Ribeiro, F., Xia, T., Monteiro, M., Pawlowski, N., Glocker, B.: High fidelity image counterfactuals with probabilistic causal models. ICML (2023)

work page 2023

[2] [1]

In: International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging

Baugh, M., Reynaud, H., Marimont, S.N., Cechnicka, S., Müller, J.P., Tarroni, G., Kainz, B.: Image-conditioned diffusion models for medical anomaly detection. In: International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. pp. 117–127. Springer (2024)

work page 2024

[3] [2]

In: MICCAI 2023

Baugh, M., Tan, J., Müller, J.P., Dombrowski, M., Batten, J., Kainz, B.: Many tasks make light work: Learning to localise medical anomalies from multiple syn- thetic tasks. In: MICCAI 2023. pp. 162–172. Springer (2023)

work page 2023

[4] [3]

Medical image anal- ysis 66, 101797 (2020) 10 Mehta et al

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 anal- ysis 66, 101797 (2020) 10 Mehta et al

work page 2020

[5] [4]

In: MIDL

Cohen, J.P., Brooks, R., En, S., Zucker, E., Pareek, A., Lungren, M.P., Chaudhari, A.: Gifsplanation via latent shift: a simple autoencoder approach to counterfactual generation for chest x-rays. In: MIDL. pp. 74–104 (2021)

work page 2021

[6] [6]

MIDL (2024)

Fathi, N., Kumar, A., Nichyporuk, B., Havaei, M., Arbel, T.: Decodex: Confounder detector guidance for improved diffusion-based counterfactual explanations. MIDL (2024)

work page 2024

[7] [7]

Scientific Data11(1), 511 (2024)

Gaggion, N., Mosquera, C., Mansilla, L., Saidman, J.M., Aineseder, M., Milone, D.H., Ferrante, E.: Chexmask: a large-scale dataset of anatomical segmentation masks for multi-center chest x-ray images. Scientific Data11(1), 511 (2024)

2024

[8] [8]

Machine Learning for Biomedical Imaging2, 611–656 (2024)

Hassanaly, R., Brianceau, C., Solal, M., Colliot, O., Burgos, N.: Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain fdg pet. Machine Learning for Biomedical Imaging2, 611–656 (2024)

work page 2024