Harnessing EHRs for Diffusion-based Anomaly Detection on Chest X-rays

Charmgil Hong; Harim Kim; Heeyoul Choi; Minkyu Ahn; Yuhan Wang; Yuyin Zhou

REVIEW 4 major objections 5 minor 31 references

Harnessing EHRs for Diffusion-based Anomaly Detection on Chest X-rays

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Diff3M integrates structured electronic health records into a diffusion model via image-EHR cross-attention and checkerboard masking, reporting state-of-the-art unsupervised anomaly detection on MIMIC-CXR and CheXpert.

desk verdict EHR-conditioned diffusion UAD is a real first, but the SOTA claim is built on a post-hoc scoring choice; deserves a careful referee, not a desk reject. read the letter →

arxiv 2505.17311 v1 pith:TFUHMVWZ submitted 2025-05-22 cs.CV cs.LG

Harim Kim , Yuhan Wang , Minkyu Ahn , Heeyoul Choi , Yuyin Zhou , Charmgil Hong This is my paper

classification cs.CVcs.LG

keywords DiffusionmodelsUnsupervisedanomalydetectionChestX-rayElectronichealthrecordsCross-attentionCheckerboardmaskingMulti-modalmedicalimagingMIMIC-CXR

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 tries to establish that structured electronic health records, not just pixels, should condition diffusion models for unsupervised anomaly detection on chest X-rays. It introduces Diff3M, which injects EHR-derived embeddings into the noise-prediction and masked-reconstruction networks through an image-EHR cross-attention module, and pairs it with a pixel-level checkerboard masking scheme. On MIMIC-CXR and CheXpert the method reports state-of-the-art image-level AUROC and AUPRC, including 0.617 and 0.821 on MIMIC versus 0.597 and 0.803 for the strongest prior diffusion baseline DiAD. If correct, the paper establishes that clinical tabular context helps a model tell normal anatomical variation from pathological findings without any labels.

What carries the argument

The mechanism that carries the argument is the Image-EHR Cross Attention (IECA) module. It tokenizes the EHR row into $f$ feature tokens $F \in \mathbb{R}^{f \times d}$, computes token weights $w_r = \mathrm{softmax}(F e^\top / \sqrt{d})$ against the image embedding $e$, and forms the conditional embedding $c_r = w_r F$ that conditions both the noise-prediction (NP) and masked-pixel-generation (MPG) networks. The second mechanism, Pixel-level Checkerboard Masking (PCM), applies two complementary checkerboard masks $M_1, M_2$ to the noised input, reconstructs each masked version, and combines them as $\tilde{x}_t = \mathrm{MPG}(x_t^{m_1}) \odot M_2 + \mathrm{MPG}(x_t^{m_2}) \odot M_1$, forcing the networks to use the EHR-conditioned context to fill in deliberately obscured pixels.

What would settle it

Run the MIMIC comparison several times with different seeds and again with EHR values randomly permuted; if the 0.617-vs-0.610 AUROC gap between the full-EHR and demographics-only variants does not persist beyond run-to-run noise, or if shuffled EHR values give the same lift, then the reported EHR conditioning is not the cause.

Watch

Extended reading notes

Core claim

The paper's central claim is that structured EHR data can be turned into a conditioning signal that makes diffusion-based unsupervised anomaly detection more accurate on chest X-rays. Diff3M computes a conditional embedding from the EHR record and the X-ray image, feeds it into both the noise-prediction and masked-pixel-generation networks, and reconstructs masked images into normal-like versions; the anomaly score is the pixel difference between the original and regenerated image. On MIMIC-CXR the full model reports image-level AUROC 0.617 and AUPRC 0.821, against DiAD's 0.597 and 0.803, and on CheXpert it reports AUROC 0.664 and AUPRC 0.931, matching DiAD's AUROC while raising AUPRC. The paper attributes the improvement to EHR features (especially BMI) capturing body shape, which reduces the anatomical variability that pure image-conditioned diffusion models struggle with.

Load-bearing premise

The load-bearing premise is that the chosen EHR fields (BMI, blood pressure, height, weight, plus demographics) carry genuine information about each patient's normal anatomy, so the cross-attention embedding helps reconstruct anomalies into normal-like images rather than simply adding parameters that happen to improve the score.

Editorial extensions

If this is right

On MIMIC-CXR, Diff3M reports AUROC 0.617 and AUPRC 0.821, exceeding DiAD (0.597/0.803) and the feature-based baselines PatchCore, RD4AD, and MambaAD.
On CheXpert, Diff3M with demographics-only EHR matches DiAD's AUROC (0.664) while lifting AUPRC from 0.917 to 0.931.
Ablations attribute part of the gain to the checkerboard masking strategy and part to the IECA conditioning, with the combination performing best on both datasets.
Attention-weight analysis identifies BMI as the dominant EHR feature, especially among top-decile values, indicating that the injected clinical signal is largely about body shape.
The choice of anomaly-scoring function matters: MSE suits MIMIC while max-absolute difference suits CheXpert, with disease-class analysis explaining the discrepancy.

Reading between the lines

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

If the EHR gain is real, then richer clinical inputs—lab values, medications, prior radiology reports—should push the same mechanism further, since they carry additional anatomical and physiologic context; this is a testable prediction the paper does not make.
The BMI-dominance result suggests a simpler variant: conditioning only on BMI and demographics might reproduce most of the MIMIC gain, which would clarify whether the extra EHR fields or the cross-attention parameters do the work.
Because the demographics-only variant ties DiAD on CheXpert, the EHR benefit may be dataset-dependent; reporting multiple seeds and confidence intervals would show whether the 0.007 AUROC gap on MIMIC is a stable effect.
The checkerboard masking scheme is modality-agnostic and could be evaluated on non-medical anomaly detection benchmarks, where the same masking-forces-reconstruction principle might transfer independently of EHR conditioning.

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

First, what's new: this is the first diffusion-based unsupervised anomaly detection for chest X-rays that conditions on structured EHRs, and that's a genuine, sensible idea. The IECA module is a straightforward soft attention over tabular tokens, and the checkerboard mask is a modest variant of existing masking, but the combination is not in the cited literature. The paper frames the problem well: normal anatomical variation is a confounder, and clinical variables like BMI may help.

The empirical case for SOTA, however, doesn't hold as stated. You're right to flag the per-dataset scoring selection. Table 2 shows that on CheXpert, MSE gives AUROC 0.547 and maxabs gives 0.664; on MIMIC, MSE gives 0.617 and maxabs gives 0.568. The authors report the better scoring per dataset in Table 1 without saying DiAD's scoring function. If DiAD's scoring is fixed, Diff3M's advantage may vanish or reverse: with maxabs on MIMIC, it falls below DiAD; with MSE on CheXpert, it falls well below. That's a post-hoc selection on the test set, not a protocol error a reviewer can fix by re-running. It's a load-bearing flaw in the SOTA claim. Also, the actual gains are tiny—0.007 AUROC on MIMIC over the no-EHR variant—and there are no error bars or significance tests. The CheXpert result ties with DiAD, which the authors acknowledge but then still claim SOTA overall.

The rest is okay. The DDPM/DDIM machinery is standard and correctly described. The ablation shows each module helps under the chosen scoring, but that's expected. The attention weight analysis (BMI dominating) is suggestive but not proof that EHRs add independent signal; there is no controlled experiment isolating the EHR contribution. That's a minor issue.

This deserves peer review rather than desk rejection. The novelty is real, the framing is useful, and the problems are fixable: disclose scoring selection, evaluate with a fixed protocol, report variance, and temper the CheXpert claim. As is, I'd treat it as a modest technical contribution. I would not cite it in the next year until the evaluation is cleaned up. A reading group might find the scoring-function discussion a good cautionary tale.

Recommendation: send to a serious referee, expect heavy revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes Diff3M, an unsupervised anomaly-detection method for chest X-rays that conditions a DDPM/DDIM diffusion model on structured electronic health records (EHRs). The two main technical contributions are an Image-EHR Cross Attention (IECA) module that derives an EHR-conditioned embedding from the input image and tabular features, and a Pixel-level Checkerboard Masking (PCM) strategy that masks the noised input and reconstructs it into a normal-like image during the reverse process. Training is performed on normal samples only, and anomalies are scored by the difference between the input and the reconstructed image. The method is evaluated on CheXpert and MIMIC-CXR/IV, reporting image-level AUROC/AUPRC against PatchCore, RD4AD, MambaAD, and DiAD. The authors claim state-of-the-art performance on both datasets, with the MIMIC results attributed mainly to the added EHR features.

Significance. If the empirical claims held, this would be a useful demonstration that structured clinical data can reduce anatomical-variability confounds in diffusion-based anomaly detection, and the released code would support reproducibility. The underlying diffusion machinery is standard, and the ablation structure is sensible. However, the central evidence is currently fragile: the headline comparison relies on per-dataset selection of the anomaly scoring function after seeing test results, the EHR contribution rests on a 0.007 AUROC gap with no error bars, and the CheXpert AUROC ties DiAD. The paper therefore needs substantial revision before its stated conclusions can be accepted.

major comments (4)

[§4.2 (Tables 1–2)] The reported state-of-the-art results are obtained by selecting different anomaly scoring functions per dataset after inspecting test-set performance: maxabs for CheXpert and MSE for MIMIC. With the alternative scoring, Diff3M falls below DiAD on both datasets (CheXpert MSE AUROC 0.547 vs DiAD 0.664; MIMIC maxabs AUROC 0.568 vs DiAD 0.597). Because the scoring function is part of the detection algorithm and the selection is not pre-registered or validated on a held-out split, the comparison in Table 1 is not a single fixed procedure and the SOTA claim is not yet established.
[Table 2] The key comparison supporting the EHR contribution is Ours_demo vs Ours_demo,ehr on MIMIC: AUROC 0.610 vs 0.617 and AUPRC 0.818 vs 0.821, a gap of 0.007 and 0.003 respectively. No error bars, confidence intervals, or significance tests are reported anywhere in the paper, so this gap is within plausible run-to-run variability; the conclusion that structured EHR features improve anomaly detection is therefore not supported by the current evidence.
[Table 1 and Abstract] In Table 1, Diff3M ties DiAD on CheXpert AUROC (0.664 for both), while the Abstract claims the method 'achieves state-of-the-art performance, outperforming existing UAD methods.' The only CheXpert improvement is AUPRC (0.931 vs 0.917). The claim should be qualified to acknowledge the AUROC tie and to specify which metric is improved; as written, it overstates the result.
[Eq. (1) and Algorithm 1] Equation (1) and Algorithm 1 (line 4) present a DDIM update that does not match the standard DDIM sampling formula in the cited literature. The bracket in Eq. (1) simplifies to (sqrt(1/α_t) − sqrt(1/α_{t+1}))(x_t − ε), which is not the DDIM step used in Song et al. or Wolleb et al.; as written this makes the inference procedure irreproducible and possibly inconsistent with the implementation. Please correct the equations or clarify the exact update used.

minor comments (5)

[Table 1] The header 'Oursdemo,ehr' is typeset incorrectly; it should be 'Ours_demo,ehr'.
[§4.2] The CheXpert rows of Table 2 omit the ehr condition; clarify that the additional EHR features are available only for MIMIC and state what this implies for the CheXpert comparison.
[§4.1] The implementation details do not specify how the baseline methods (PatchCore, RD4AD, MambaAD, DiAD) were configured, in particular which anomaly scoring function DiAD uses; without this information the comparison in Table 1 is not fully reproducible.
[Eq. (2)] The definition w_r = σ softmax((Fe^T)/√d) is ambiguous because both σ and softmax are used together; please clarify the exact activation and how the final conditional embedding c_r is formed.
[Eq. (3)] The notation M_{1,2} ← (M_{1,2}×(1−s)+s) is unclear about elementwise operations and whether the masks remain binary after blending; please define the operation precisely.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: Diff3M's EHR conditioning and masking are not defined in terms of its reported AUROC/AUPRC targets; results are benchmarked against external baselines.

full rationale

The derivation chain is self-contained. The IECA conditional embedding c_r is computed from tabular EHR features via a feature tokenizer and a similarity-weighted sum with the image embedding (Eq. 2); the PCM masks are static coordinate-based checkerboards scaled by time step (Eq. 3); the training loss compares predicted noise and reconstructed masked pixels against diffusion-process targets (Eq. 4); and the anomaly score is the pixel difference between the input and the DDIM-reconstructed output. None of these quantities is defined in terms of the reported detection metrics, and no network parameter is fitted to the test-set AUROC/AUPRC values. The SOTA claim is supported by comparisons to external baselines (PatchCore, RD4AD, MambaAD, DiAD). The only near-self-citation, RadFusion (ref. [31], which shares a co-author), is used as motivational background for multimodal EHR+image learning, not as a load-bearing premise for Diff3M's design or evaluation. The per-dataset choice between MSE and maxabs scoring in Sec. 4.2 is a post-hoc evaluation-selection concern and a reproducibility risk, but it is not a definitional or fitted-input circularity: the scoring function is chosen after seeing test performance rather than being derived from the method's own outputs. Overall, no step reduces by construction to its own inputs, so the circularity score is minimal.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-set hyperparameters (lambda, mask schedule, DDIM steps) and on domain assumptions about EHR informativeness and label reliability. The most consequential free choice is the anomaly scoring function, which is selected per dataset after observing test performance.

free parameters (5)

Loss weighting lambda = 0.5
Fixed in Eq. 4, balancing noise prediction and masked reconstruction; no ablation is shown for other values.
Masking scale schedule s=t/T = t/T
Introduced in Eq. 3; the schedule is chosen by design and not ablated.
DDIM inference steps = 600 (CheXpert), 400 (MIMIC)
Given in Implementation Details; not justified by experiments.
Anomaly scoring function = maxabs for CheXpert, MSE for MIMIC
Table 2 shows per-dataset best scorer; Table 1 uses these better scores without stating so, which is a post-hoc selection on the test set.
IECA token count f and embedding dimension d = Not reported
The feature tokenizer from Gorishniy et al. is used, but exact token and dimension values are absent, making replication harder.

assumptions (4)

standard math DDPM/DDIM diffusion processes can represent the distribution of normal chest X-rays and reconstruct anomalies as normal-like images.
Adopted from Ho et al. and Song et al. in Section 2; the reverse process is trained with the standard noise and reconstruction losses in Eq. 4.
domain assumption CheXpert 'No Finding' cases define the normal class, and all other labels define anomalies.
Section 4 defines the train/test protocol; label noise and unlabeled pathologies in 'No Finding' scans could violate the normality assumption.
domain assumption Paired EHR features (age, sex, AP/PA view, and for MIMIC BMI, blood pressure, height, weight) are predictive of normal anatomical variation relevant to image reconstruction.
Central premise of IECA in Section 3.1; without this, the conditioning embedding cannot separate anatomical variation from pathology.
domain assumption MIMIC-IV EHR records matched within three months of the X-ray correspond to the patient state at imaging time.
Section 4 describes the record selection window; temporal mismatch would inject noise into the conditioning signal.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Harnessing EHRs for Diffusion-based Anomaly Detection on Chest X-rays." pith.science (2026). https://pith.science/paper/TFUHMVWZ

@misc{pith2026250517311,
  author       = {Pith},
  title        = {Pith review of: Harnessing EHRs for Diffusion-based Anomaly Detection on Chest X-rays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TFUHMVWZ}},
  note         = {Machine review of arXiv:2505.17311}
}

read the original abstract

Unsupervised anomaly detection (UAD) in medical imaging is crucial for identifying pathological abnormalities without requiring extensive labeled data. However, existing diffusion-based UAD models rely solely on imaging features, limiting their ability to distinguish between normal anatomical variations and pathological anomalies. To address this, we propose Diff3M, a multi-modal diffusion-based framework that integrates chest X-rays and structured Electronic Health Records (EHRs) for enhanced anomaly detection. Specifically, we introduce a novel image-EHR cross-attention module to incorporate structured clinical context into the image generation process, improving the model's ability to differentiate normal from abnormal features. Additionally, we develop a static masking strategy to enhance the reconstruction of normal-like images from anomalies. Extensive evaluations on CheXpert and MIMIC-CXR/IV demonstrate that Diff3M achieves state-of-the-art performance, outperforming existing UAD methods in medical imaging. Our code is available at this http URL https://github.com/nth221/Diff3M

Figures

Figures reproduced from arXiv: 2505.17311 by the authors.

**Figure 1.** Training process of Diff3M. Image-EHR Cross Attention (IECA) incorporates EHR embeddings as conditional inputs for the noise prediction (NP) network at each reverse process step. A novel Pixel-level Checkerboard Masking (PCM) strategy is employed to improve generating masked inputs into normal-like images from the Masked Pixel Generation (MPG) Network. embeddings to capture temporal information. Second, we propose … view at source ↗

**Figure 2.** Details of the proposed modules. (a) The Image-EHR Cross Attention module incorporate structured clinical context into the image generation process. (b) The Pixel-level Checkerboard Masking module facilitates the regeneration of anomalies into normal-like images using two types of static masks. pathological abnormalities that deviate from predefined patterns. Instead, we employ static masking to deliberately obscur… view at source ↗

**Figure 3.** Qualitative comparison. Red circles indicate anomalous artifacts. 4.1 Performance Comparison We compare Diff3M with state-of-the-art (SOTA) feature-based models, e.g., PatchCore [23], RD4AD [5], and MambaAD [10], along with the diffusion-based model DiAD [11]. For qualitative evaluation, we compare the anomaly maps, computed as the absolute pixel-level difference between the input and output. Quantitative and Qualit… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages

[1]

In: Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Re- vised Selected Papers, Part III 14

Akcay, S., Atapour-Abarghouei, A., Breckon, T.P.: Ganomaly: Semi-supervised anomaly detection via adversarial training. In: Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Re- vised Selected Papers, Part III 14. pp. 622–637. Springer (2019)

work page 2019
[2]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Bao,J.,Sun,H.,Deng,H.,He,Y.,Zhang,Z.,Li,X.:Bmad:Benchmarksformedical anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4042–4053 (2024)

2024
[3]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bercea, C.I., Wiestler, B., Rueckert, D., Schnabel, J.A.: Diffusion models with implicit guidance for medical anomaly detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 211–220. Springer (2024)

work page 2024
[4]

In: International conference on pattern recognition

Defard,T.,Setkov,A.,Loesch,A.,Audigier,R.:Padim:apatchdistributionmodel- ing framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021)

2021
[5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class embed- ding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022)

2022
[6]

Meta-Radiology p

Fan, Y., Liao, H., Huang, S., Luo, Y., Fu, H., Qi, H.: A survey of emerging appli- cations of diffusion probabilistic models in mri. Meta-Radiology p. 100082 (2024)

work page 2024
[7]

Neurocomputing414, 346– 355 (2020)

Feng, X., Huang, Q., Li, X.: Ultrasound image de-speckling by a hybrid deep network with transferred filtering and structural prior. Neurocomputing414, 346– 355 (2020)

work page 2020
[8]

Advances in Neural Information Processing Systems34, 18932–18943 (2021)

Gorishniy, Y., Rubachev, I., Khrulkov, V., Babenko, A.: Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems34, 18932–18943 (2021)

2021

Show all 31 references

[9]

In: International conference on medical image computing and computer- assisted intervention

Han, Z., Wang, Y., Zhou, L., Wang, P., Yan, B., Zhou, J., Wang, Y., Shen, D.: Contrastive diffusion model with auxiliary guidance for coarse-to-fine pet recon- struction. In: International conference on medical image computing and computer- assisted intervention. pp. 239–249. ...

2023
[10]

arXiv preprint arXiv:2404.06564 (2024)

He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564 (2024)

2024 arXiv
[11]

In: Proceedings of the AAAI Conference on Artificial Intelligence

He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., Xie, L.: A diffusion-based framework for multi-class anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 8472–8480 (2024)

2024
[12]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[13]

In: International Workshop on Machine Learning in Medical Imaging

Iqbal, H., Khalid, U., Chen, C., Hua, J.: Unsupervised anomaly detection in med- ical images using masked diffusion model. In: International Workshop on Machine Learning in Medical Imaging. pp. 372–381. Springer (2023)

2023
[14]

dataset with uncertainty labels and expert comparison

Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al.: Chexpert: A large chest radiograph 10 Kim et al. dataset with uncertainty labels and expert comparison. In: Proceedings of the AAAI conference on...

2019
[15]

physionet (2024)

Johnson, A., Bulgarelli, L., Pollard, T., Gow, B., Moody, B., Horng, S., Celi, L., Mark, R.: Mimic-iv (version 3.1). physionet (2024)

2024
[16]

Johnson, A., Pollard, T., Mark, R., Berkowitz, S., Horng, S.: Mimic-cxr database (version 2.0. 0). physionet2, 5 (2019)

2019
[17]

PhysioNet

Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L.A., Mark, R.: Mimic-iv. PhysioNet. Available online at: https://physionet. org/content/mimiciv/1.0/(accessed August 23, 2021) pp. 49–55 (2020)

2020
[18]

IEEE Access10, 78446–78454 (2022)

Lee, S., Lee, S., Song, B.C.: Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. IEEE Access10, 78446–78454 (2022)

2022
[19]

arXiv preprint arXiv:2504.06897 (2025)

Mao, J., Wang, Y., Tang, Y., Xu, D., Wang, K., Yang, Y., Zhou, Z., Zhou, Y.: Med- segfactory: Text-guided generation of medical image-mask pairs. arXiv preprint arXiv:2504.06897 (2025)

2025 arXiv
[20]

arXiv preprint arXiv:2305.15956 (2023)

Mousakhan, A., Brox, T., Tayyub, J.: Anomaly detection with conditioned denois- ing diffusion models. arXiv preprint arXiv:2305.15956 (2023)

2023 arXiv
[21]

Diagnostics15(3), 282 (2025)

Obuchowicz, R., Lasek, J., Wodziński, M., Piórkowski, A., Strzelecki, M., Nurzyn- ska, K.: Artificial intelligence-empowered radiology—current status and critical review. Diagnostics15(3), 282 (2025)

2025
[22]

In: Proceedings of the 29th annual conference on Com- puter graphics and interactive techniques

Perlin, K.: Improving noise. In: Proceedings of the 29th annual conference on Com- puter graphics and interactive techniques. pp. 681–682 (2002)

2002
[23]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022)

2022
[24]

Schlegl,T.,Seeböck,P.,Waldstein,S.M.,Langs,G.,Schmidt-Erfurth,U.:f-anogan: Fastunsupervisedanomalydetectionwithgenerativeadversarialnetworks.Medical image analysis54, 30–44 (2019)

2019
[25]

In: International conference on information processing in medical imag- ing

Schlegl, T., Seeböck, P., Waldstein, S.M., Schmidt-Erfurth, U., Langs, G.: Unsu- pervised anomaly detection with generative adversarial networks to guide marker discovery. In: International conference on information processing in medical imag- ing. pp. 146–157. Springer (2017)

2017
[26]

arXiv preprint arXiv:2010.02502 (2020)

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

2020 arXiv
[27]

In: International Conference on Medical image computing and computer-assisted intervention

Wolleb, J., Bieder, F., Sandkühler, R., Cattin, P.C.: Diffusion models for medical anomaly detection. In: International Conference on Medical image computing and computer-assisted intervention. pp. 35–45. Springer (2022)

2022
[28]

In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

Wyatt, J., Leach, A., Schmon, S.M., Willcocks, C.G.: Anoddpm: Anomaly detec- tion with denoising diffusion probabilistic models using simplex noise. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 650–656 (2022)

2022
[29]

In: Proceedings of the AAAI conference on artificial intelligence

Yao, W., Yin, K., Cheung, W.K., Liu, J., Qin, J.: Drfuse: Learning disentangled representation for clinical multi-modal fusion with missing modality and modal inconsistency. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 16416–16424 (2024)

2024
[30]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Zhang,X.,Li,N.,Li,J.,Dai,T.,Jiang,Y.,Xia,S.T.:Unsupervisedsurfaceanomaly detection with diffusion probabilistic model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6782–6791 (2023)

2023
[31]

arXiv preprint arXiv:2111.11665 (2021)

Zhou, Y., Huang, S.C., Fries, J.A., Youssef, A., Amrhein, T.J., Chang, M., Baner- jee, I., Rubin, D., Xing, L., Shah, N., et al.: Radfusion: Benchmarking performance and fairness for multimodal pulmonary embolism detection from ct and ehr. arXiv preprint arXiv:2111.11665 (2021)

2021 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

In: Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Re- vised Selected Papers, Part III 14

Akcay, S., Atapour-Abarghouei, A., Breckon, T.P.: Ganomaly: Semi-supervised anomaly detection via adversarial training. In: Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Re- vised Selected Papers, Part III 14. pp. 622–637. Springer (2019)

work page 2019

[2] [2]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Bao,J.,Sun,H.,Deng,H.,He,Y.,Zhang,Z.,Li,X.:Bmad:Benchmarksformedical anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4042–4053 (2024)

2024

[3] [3]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Bercea, C.I., Wiestler, B., Rueckert, D., Schnabel, J.A.: Diffusion models with implicit guidance for medical anomaly detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 211–220. Springer (2024)

work page 2024

[4] [4]

In: International conference on pattern recognition

Defard,T.,Setkov,A.,Loesch,A.,Audigier,R.:Padim:apatchdistributionmodel- ing framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021)

2021

[5] [5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class embed- ding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022)

2022

[6] [6]

Meta-Radiology p

Fan, Y., Liao, H., Huang, S., Luo, Y., Fu, H., Qi, H.: A survey of emerging appli- cations of diffusion probabilistic models in mri. Meta-Radiology p. 100082 (2024)

work page 2024

[7] [7]

Neurocomputing414, 346– 355 (2020)

Feng, X., Huang, Q., Li, X.: Ultrasound image de-speckling by a hybrid deep network with transferred filtering and structural prior. Neurocomputing414, 346– 355 (2020)

work page 2020

[8] [8]

Advances in Neural Information Processing Systems34, 18932–18943 (2021)

Gorishniy, Y., Rubachev, I., Khrulkov, V., Babenko, A.: Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems34, 18932–18943 (2021)

2021