Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Conditional diffusion models for guided anomaly detection in brain images using fluid-driven anomaly randomization

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 Conditioning a latent diffusion model on fluid-randomized pseudo-pathology images yields better anomaly detection and pseudo-healthy reconstruction in 3D T1 brain MRI than unsupervised baselines and, on most datasets, than supervised…

desk verdict Useful weakly supervised 3D anomaly detection architecture that deserves referee time, but the ATLAS train/test separation for lesion maps must be clarified before the real-stroke claim can be trusted. read the letter →

arxiv 2506.10233 v1 pith:F3FPZP3X submitted 2025-06-11 eess.IV cs.CV

Ana Lawry Aguila , Peirong Liu , Oula Puonti , Juan Eugenio Iglesias This is my paper
classification eess.IVcs.CV
keywords anomalydetectiondiffusionmodelsbrainMRIlatentpseudo-pathologygenerationfluid-drivenrandomizationweaklysupervisedlearningimagereconstruction
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 sets out to show that a latent diffusion model, when conditioned on synthetic pseudo-pathology images generated by fluid-driven anomaly randomization, can detect real brain lesions and reconstruct healthy tissue in 3D T1 MRI better than reconstruction-based unsupervised methods. The motivation is practical: supervised lesion detection needs large annotated disease cohorts, while healthy-only training often fails to remove large anomalies or preserve abnormal tissue. The proposed training scheme uses healthy scans plus randomized versions of real stroke lesion maps from an auxiliary atlas, so no paired diseased/healthy images from the target population are required. Across five synthetic cohorts and the ATLAS real-stroke test set, the paper reports consistent improvements over VAEs and conditional and unconditional latent diffusion, and on most cohorts higher Dice and lower false-positive rates than a supervised inpainting baseline.

What carries the argument

The load-bearing machinery is a two-stage latent diffusion model with an added pathology-conditioning branch. An auxiliary encoder $\mathrm{Enc}_\varphi$ maps the latent code $z_p$ of the pseudo-pathology image to a 1280-dimensional vector $c_p$, which is injected through adaptive group normalization at every scale of the denoising UNet; the training objective becomes $\mathbb{E}[\|\epsilon_t - \epsilon_\theta(z_t, c_p, t)\|^2]$. Paired with this is the fluid-driven anomaly randomization of Eqs. (3)-(6), which turns healthy images into pseudo-diseased counterparts by advecting and diffusing real lesion maps under random velocity and diffusion potentials. At inference, the anomaly map $m(x_0, x_0^T)$ is a median-filtered, brain-masked product of the pixel-wise absolute difference and the LPIPS similarity, giving the localization signal.

What would settle it

A direct test: train the model on healthy cohorts with the same pipeline but delete the 20% sign-flip in Eq. (6), then run it on real hypointense lesions; if the anomaly maps stop localizing them or false positives climb sharply, the generator's intensity heuristic is a load-bearing part of the transfer, not a detail.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that feeding the denoising network a vector embedding of the pseudo-pathology image at every diffusion timestep, through adaptive group normalization layers, makes the model learn to replace anomalous tissue with healthy anatomy while leaving healthy regions intact. The pseudo-pathologies themselves are produced by an advection-diffusion PDE that randomizes real stroke segmentation maps $P_0$ with Perlin-noise potentials $\Psi$ and $\Phi$, yielding $x_p = x_h + \Delta x_h P_{T_{\max}}(x)$, where $\Delta x_h$ is sampled from a white-matter-intensity-based normal distribution. The paper claims this conditional formulation, trained only with the synthetic images, yields anomaly maps that localize real ATLAS strokes better than VAE, LDM, LDM(Tavg) and cLDM, and on most datasets better than the supervised SynthSR, with the largest gains in false-positive rate.

Load-bearing premise

The whole method hinges on the assumption that the randomized synthetic lesions look and behave enough like real pathology that training to remove them transfers to real anomalies; if real lesions have shapes, intensities, or locations the randomized generator never produces, the conditioning signal can misdirect rather than guide reconstruction.

Editorial extensions

If this is right

  • Anomaly detection can be trained without target-cohort lesion labels: healthy scans plus an auxiliary lesion atlas suffice.
  • The conditioning branch should keep improving reconstruction as more diverse pseudo-pathologies are generated, since training corrects synthetic anomalies while preserving healthy tissue.
  • On real stroke images from ATLAS, the reported Dice, average precision, and false-positive-rate gains over the second-best model mean fewer false alarms when localizing lesions.
  • The generated pseudo-healthy images can support downstream tasks such as brain parcellation and tissue segmentation, which the paper names as future work.

Reading between the lines

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

  • A testable extension is to swap the ATLAS stroke seed maps for other lesion atlases, such as white-matter hyperintensity masks, and measure whether the conditioning branch still transfers to real lesions of those types; the paper's mechanism does not require stroke-specific shapes, only realistic seed maps.
  • Because the intensity perturbation of Eq. (6) is a hand-set heuristic, a plausible next step is to learn the intensity distribution from data; if detection on non-stroke lesions improves, that would confirm that the generator's realism, not the diffusion conditioning, is the limiting factor.
  • The reported low false-positive rate on ATLAS suggests the pseudo-healthy reconstructions may be accurate enough for downstream parcellation, but whether healthy anatomy is preserved better than in UNA is not directly measured; a reconstruction-error comparison on healthy regions would separate conditioning gains from generator gains.
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

4 major / 5 minor

Summary. This paper proposes a weakly supervised conditional latent diffusion model for anomaly detection and pseudo-healthy reconstruction in 3D T1-weighted brain MRI. Pseudo-pathologies are synthesized by fluid-driven anomaly randomization applied to lesion segmentations from the ATLAS dataset (Eq. (3)-(6)), and an auxiliary encoder extracts a conditioning vector cp that modulates the diffusion UNet via adaptive group normalization. The model is evaluated on synthetic pathology images generated for HCP, ADNI, ADHD200, OASIS3, and AIBL, and on real stroke lesions from an ATLAS test cohort. The authors report that their method consistently outperforms VAE, LDM, and cLDM baselines, and surpasses supervised methods on most datasets while achieving lower false-positive rates.

Significance. The contribution is potentially significant: if the results hold, it provides a practical weakly supervised diffusion framework that leverages an auxiliary lesion database to improve localization of both synthetic and real brain pathology, and it is among the first 3D diffusion models to incorporate fluid-driven synthetic anomalies. The real ATLAS evaluation is an important external benchmark. The main strengths are the anatomically motivated pseudo-pathology generation and the public ATLAS validation. However, confidence in the central claim is currently limited by unresolved questions about dataset split separation, test-time conditioning, and statistical significance.

major comments (4)
  1. [Section 3 (Datasets)] The manuscript does not state that the Ntest=56 ATLAS subjects used for real-stroke evaluation are excluded from the Ntrain=590 ATLAS lesion maps from which training pseudo-pathologies are sampled. As written, 'initial profiles sampled from the lesion segmentation maps of ATLAS subjects' could include the test subjects' maps. If so, the model has seen the exact lesion geometry of the test cases during training and the ATLAS results in Table 1 (Dice 0.3110, AP 0.2643, FPR 0.0091) would reflect memorization rather than generalization. Please state explicitly that the split is disjoint, and if it is not, rerun the ATLAS evaluation with a training generator that excludes the 56 test lesion maps.
  2. [Section 2.2 and Section 3 (Comparison metrics)] Training conditions the denoising network on cp, the embedding of the synthetic pseudo-pathology image zp, but the evaluation paragraph says the model 'takes only pathology samples as input, using them for zT0 and zp.' This is ambiguous: if the real pathological image is encoded as zp at inference, the conditioning distribution differs from training and may leak information about the anomaly into the reconstruction. If instead a pseudo-pathology image is generated for test samples, the mechanism should be described and its potential dependence on test-time generated anomalies clarified. Please specify exactly what input is used to compute cp at inference and whether any ablation without test-time conditioning was performed.
  3. [Table 1] The results are reported as single point estimates without error bars, confidence intervals, or significance tests. Differences between methods are often small (e.g., ADNI Dice 0.2493 for Ours versus 0.2553 for SynthSR; ATLAS AUCpix 0.9413 for Ours versus 0.9539 for UNA). To support the claim of 'consistently outperforms,' please add variability across random seeds or a paired statistical test across test subjects.
  4. [Section 3 (Datasets and Implementation details)] The synthetic test sets are generated with the same fluid-driven anomaly randomization pipeline used to create training pseudo-pathologies (Eq. (3)-(6)), so performance on these cohorts partly measures memorization of the generator distribution rather than generalization to novel pathology. The ATLAS real-lesion test mitigates this concern, but only if the split issue in the first comment is resolved. Please report results on independently generated synthetic anomalies or otherwise demonstrate that the gains transfer outside the training generator's distribution.
minor comments (5)
  1. [Section 2.2] The word 'auxillary' should be 'auxiliary'.
  2. [Equation (6)] The intensity perturbation heuristic N(-mu_w/2, mu_w/2) with a 20% sign flip is a hand-set choice that directly controls pseudo-pathology realism; please provide an ablation or at least a sensitivity analysis for these parameters.
  3. [Section 3] The statement that the authors 'pre-compute 8 synthetic pathology images for each sample of the training and validation cohorts and 2 images for the test cohorts' should clarify whether the test cohorts' synthetic images are used only for evaluation and whether the corresponding healthy test subjects are excluded from training.
  4. [Section 2.2] The symbol zT0 used in the anomaly detection paragraph is not defined; it likely refers to the reconstructed latent after T_int denoising steps, but it should be defined explicitly.
  5. [References] Recent unsupervised diffusion anomaly detection baselines with implicit guidance are cited in the introduction but not compared in Table 1; adding them would strengthen the baseline comparison.

Circularity Check

2 steps flagged · score 3.0 of 10

No equation-level circularity; moderate self-citation (UNA generator) plus an unstated ATLAS train/test split leaves the external validation unverifiable.

  1. other [Section 3, Datasets (ATLAS paragraph); Table 1 (ATLAS real-stroke results)]
    "We use manual lesion segmentation maps from subjects from the ATLAS dataset [29] (Ntrain=590, Ntest=56) to generate pseudo-pathology images. Healthy scans are augmented to encode synthetic anomalies where the synthetic abnormal profiles are generated using fluid-driven anomaly randomization as described in Section 2.2, with initial profiles sampled from the lesion segmentation maps of ATLAS subjects."

    The ATLAS real-pathology evaluation in Table 1 uses Ntest=56 subjects, while training pseudo-pathologies are produced from ATLAS lesion-segmentation maps with Ntrain=590 subjects. The paper never states that the 56 test subjects are excluded from the 590-lesion-map training pool. If they are not excluded, the model has been trained on the exact lesion geometries and locations used for the real-stroke test, so the reported Dice/AP/FPR improvements (e.g., Dice 0.3110 vs 0.2711) would reflect memorization rather than generalization to unseen pathology; the one external benchmark would be internally generated. The omission is a load-bearing missing support: without an explicit disjoint split, the ATLAS result cannot be verified as independent external validation.

  2. ansatz smuggled in via citation [Section 2.2 (Eqs. 3-6) and Section 3 (Implementation details)]
    "Inspired by UNA [30], we use fluid-driven anomaly randomization to generate pseudo-pathological images that guide our model training. ... Following UNA [30], the initial pathology probability P0 is generated on-the-fly using realistic pathology segmentation maps. ... We use the anomaly randomization parameters given by [30] for generating pseudo-pathology."

    The realism and anatomical coherence of the pseudo-pathologies is the premise that makes weakly supervised training transfer to real lesions. That premise is imported from the authors' own prior work [30] (Liu, Lawry Aguila, Iglesias) via 'Following UNA [30]' and 'parameters given by [30]', rather than independently established here. Moreover, the synthetic test images are produced by the same fluid-driven randomization as the training pseudo-pathologies, so the synthetic-dataset results measure performance on the training distribution of anomalies, not on independent pathology. The new conditioning architecture is independent content, but the anomaly generator itself is a self-cited ansatz that the paper does not re-validate against real lesion morphology.

full rationale

This is an empirical methods paper, not an analytic derivation: the training objective (Eq. 7) is a standard DDPM denoising loss and the anomaly map (Eq. 8) is a standard MAE+LPIPS residual; no parameter fitted to a subset of the data is renamed as a prediction, and no equation reduces to an input by construction. The main circularity-adjacent issues are (i) the pseudo-pathology generator is inherited from the authors' own UNA paper [30], so the realism premise is a self-cited ansatz, and the synthetic test sets come from the same generator, making the synthetic evaluation in-distribution; and (ii) the paper does not state that the 56 ATLAS test subjects are disjoint from the 590 ATLAS lesion maps used to synthesize training pseudo-pathologies, which is necessary to verify the external real-stroke benchmark. The conditioning encoder and the real ATLAS benchmark (assuming a clean split) provide independent content, so the score is moderate rather than high.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced; the auxiliary encoder Enc_phi is a learned architecture, not a posited entity. The main inherited conceptual pieces are fluid-driven anomaly randomization and latent diffusion.

free parameters (3)
  • Pseudo-pathology intensity perturbation distribution = Delta x_h ~ N(-mu_w/2, mu_w/2), 20% sign flip
    Equation 6 sets synthetic lesion intensities using the mean white matter intensity mu_w; the scale is a hand-set heuristic based on prior knowledge of T1 intensities, not derived or fit in this paper.
  • Anomaly randomization PDE parameters = Inherited from UNA [30]
    Section 2.2 states the randomization parameters are taken from [30]; T_max, potentials Psi and Phi, and Perlin noise scales control the shape and smoothness of synthetic lesions and therefore the training distribution.
  • Anomaly map post-processing constants = 2 voxel shift, median filter kernel 5, 6 erosion iterations, LPIPS AlexNet weighting
    Section 4 and Eq. 8 describe alignment tolerance, weighting, smoothing, and brain-mask erosion used to build every anomaly map; these hand-chosen values affect all reported metrics.
assumptions (4)
  • standard math Standard DDPM and LDM training theory: the epsilon-matching objective in Eq. 2 is a valid surrogate for the variational bound in Eq. 1.
    The method adopts Ho et al. [19] and Rombach et al. [38] diffusion theory without re-deriving it; this is background math, not the contribution.
  • domain assumption Reconstruction assumption: a model trained on healthy and pseudo-healthy images cannot faithfully reconstruct pathological appearance, so the residual between original and pseudo-healthy image localizes anomalies.
    Introduction and Eq. 8 assume the difference between x_0 and x_T0 indicates disease; if the model learns to reconstruct anomalies, the anomaly map would be empty.
  • domain assumption Fluid-driven anomaly randomization produces pseudo-pathologies that are realistic and anatomically coherent proxies for real lesions.
    Section 2.2 generates training and synthetic test anomalies by PDE-based randomization of ATLAS lesion maps; this assumption carries the generalization from synthetic to real pathology.
  • domain assumption T1 white matter intensity prior in Eq. 6 encodes realistic lesion appearance.
    The intensity perturbation is centered at -mu_w/2 based on prior knowledge of white and gray matter intensities; the paper does not validate this against measured real lesion intensities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional diffusion models for guided anomaly detection in brain images using fluid-driven anomaly randomization." pith.science (2026). https://pith.science/paper/F3FPZP3X

@misc{pith2026250610233,
  author       = {Pith},
  title        = {Pith review of: Conditional diffusion models for guided anomaly detection in brain images using fluid-driven anomaly randomization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3FPZP3X}},
  note         = {Machine review of arXiv:2506.10233}
}
read the original abstract

Supervised machine learning has enabled accurate pathology detection in brain MRI, but requires training data from diseased subjects that may not be readily available in some scenarios, for example, in the case of rare diseases. Reconstruction-based unsupervised anomaly detection, in particular using diffusion models, has gained popularity in the medical field as it allows for training on healthy images alone, eliminating the need for large disease-specific cohorts. These methods assume that a model trained on normal data cannot accurately represent or reconstruct anomalies. However, this assumption often fails with models failing to reconstruct healthy tissue or accurately reconstruct abnormal regions i.e., failing to remove anomalies. In this work, we introduce a novel conditional diffusion model framework for anomaly detection and healthy image reconstruction in brain MRI. Our weakly supervised approach integrates synthetically generated pseudo-pathology images into the modeling process to better guide the reconstruction of healthy images. To generate these pseudo-pathologies, we apply fluid-driven anomaly randomization to augment real pathology segmentation maps from an auxiliary dataset, ensuring that the synthetic anomalies are both realistic and anatomically coherent. We evaluate our model's ability to detect pathology, using both synthetic anomaly datasets and real pathology from the ATLAS dataset. In our extensive experiments, our model: (i) consistently outperforms variational autoencoders, and conditional and unconditional latent diffusion; and (ii) surpasses on most datasets, the performance of supervised inpainting methods with access to paired diseased/healthy images.

Figures

Figures reproduced from arXiv: 2506.10233 by the authors.

Figure 1
Figure 1. The left side of the figure depicts the pathology encoding process, where [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Example reconstructions and anomaly maps for an ATLAS test sample. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    In: UNSURE (2024) 2, 6

    Baugh, M., Reynaud, H., Marimont, S.N., et al.: Image-conditioned diffusion mod- els for medical anomaly detection. In: UNSURE (2024) 2, 6

    work page 2024

  2. [2]

    Medical Image Analysis (2021) 2, 6, 8

    Baur, C., Denner, S., et al.: Autoencoders for unsupervised anomaly segmentation in brain mr images: A comparative study. Medical Image Analysis (2021) 2, 6, 8

    work page 2021

  3. [3]

    In: IMLH (2023) 2, 6

    Bercea, C., Neumayr, M., Rueckert, D., Schnabel, J.A.: Mask, stitch, and re- sample: Enhancing robustness and generalizability in anomaly detection through automatic diffusion models. In: IMLH (2023) 2, 6

    work page 2023

  4. [4]

    arXiv (2023) 2

    Bercea, C.I., Rueckert, D., Schnabel, J.A.: What do we learn? debunking the myth of unsupervised outlier detection. arXiv (2023) 2

    work page 2023

  5. [5]

    arXiv (2024) 2, 5, 6

    Bercea, C.I., Wiestler, B., Rueckert, D., Schnabel, J.A.: Diffusion models with implicit guidance for medical anomaly detection. arXiv (2024) 2, 5, 6

    work page 2024

  6. [6]

    MedIA (2023) 6

    Billot, B., Greve, D., Puonti, O., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. MedIA (2023) 6

    work page 2023

  7. [7]

    Frontiers in Systems Neuroscience (2012) 5

    Brown, M.R.G., Sidhu, G.P.S., Greiner, R., et al.: ADHD-200 global competition: diagnosing ADHD using personal characteristic data can outperform resting state fMRI measurements. Frontiers in Systems Neuroscience (2012) 5

    work page 2012

  8. [8]

    arXiv (2018) 2

    Chen, X., Konukoglu, E.: Unsupervised detection of lesions in brain mri using constrained adversarial auto-encoders. arXiv (2018) 2

    work page 2018

Show all 46 references
  1. [9]

    arXiv (2020) 2

    Chen, X., You, S., Tezcan, K.C., Konukoglu, E.: Unsupervised lesion detection via image restoration with a normative prior. arXiv (2020) 2

    2020

  2. [10]

    arXiv (2018) 2

    Denouden, T., Salay, R., Czarnecki, K., et al.: Improving reconstruction autoen- coder out-of-distribution detection with mahalanobis distance. arXiv (2018) 2

    2018

  3. [11]

    arXiv (2021) 5

    Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. arXiv (2021) 5

    2021

  4. [12]

    arXiv (2024) 2

    Durrer, A., Wolleb, J., Bieder, F., et al.: Denoising diffusion models for 3d healthy brain tissue inpainting. arXiv (2024) 2

    2024

  5. [13]

    NeuroImage (2012) 5

    Essen, D.V., Uğurbil, K., Auerbach, E., et al.: The human connectome project: A data acquisition perspective. NeuroImage (2012) 5

    2012

  6. [14]

    NeuroImage (2), 774–781 (2012) 5

    Fischl, B.: Freesurfer. NeuroImage (2), 774–781 (2012) 5

    2012

  7. [15]

    Journal of Alzheimer’s Disease Reports (2021) 5

    Fowler, C., Rainey-Smith, S.R., Bird, S.M., et al.: Fifteen years of the australian imaging, biomarkers and lifestyle (AIBL) study: Progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to alzheimer’s disease. Journal of Alzheimer’s D...

    2021

  8. [16]

    In: ICCV (2019) 2

    Gong, D., Liu, L., Le, V., et al.: Memory-augmented deep autoencoder for unsu- pervised anomaly detection. In: ICCV (2019) 2

    2019

  9. [17]

    In: CVPR Workshops (2023) 2

    Graham, M., Pinaya, W., Tudosiu, P.D., et al.: Denoising diffusion models for out-of-distribution detection. In: CVPR Workshops (2023) 2

    2023

  10. [18]

    arXiv (2023) 2, 5, 6, 8

    Graham, M., Pinaya, W., Wright, P., et al.: Unsupervised 3d out-of-distribution detection with latent diffusion models. arXiv (2023) 2, 5, 6, 8

    2023

  11. [19]

    In: Advances in Neural Information Processing Systems

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems. vol. 33, pp. 6840–6851 (2020) 2, 3

    2020

  12. [20]

    IEEE TMI (2022) 5

    Hoffmann, M., Billot, B., Greve, D., et al.: SynthMorph: Learning contrast- invariant registration without acquired images. IEEE TMI (2022) 5

    2022

  13. [21]

    Scientific Reports (2023) 5

    Iglesias, J.E.: A ready-to-use machine learning tool for symmetric multi-modality registration of brain MRI. Scientific Reports (2023) 5

    2023

  14. [22]

    Science Advances (2023) 6, 8 Title Suppressed Due to Excessive Length 11

    Iglesias, J.E., Billot, B., Balbastre, Y., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances (2023) 6, 8 Title Suppressed Due to Excessive Length 11

    2023

  15. [23]

    In: MICCAI (2023) 2

    Jiang, A., Huang, C., Cao, Q., et al.: Multi-scale cross-restoration framework for electrocardiogram anomaly detection. In: MICCAI (2023) 2

    2023

  16. [24]

    CoRR (2014) 6

    Kingma, D., Ba, J.: Adam: A method for stochastic optimization. CoRR (2014) 6

    2014

  17. [25]

    In: NeurIPS

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. In: NeurIPS. vol. 25 (2012) 5

    2012

  18. [26]

    arXiv (2021) 2

    Kumar, S., Sotiras, A.: Normvae: Normative modeling on neuroimaging data using variational autoencoders. arXiv (2021) 2

    2021

  19. [27]

    Alzheimer’s & Dementia (2018) 5

    LaMontagne, P., Keefe, S., Lauren, W., et al.: OASIS-3: Longitudinal neuroimag- ing, clinical, and cognitive dataset for normal aging and alzheimer’s disease. Alzheimer’s & Dementia (2018) 5

    2018

  20. [28]

    In: ISBI (2024) 2

    Laso, P., Cerri, S., Sorby-Adams, A., et al.: Quantifying white matter hyperinten- sity and brain volumes in heterogeneous clinical and low-field portable MRI. In: ISBI (2024) 2

    2024

  21. [29]

    Liew, S.L., Anglin, J., Banks, N., et al.: A large, open source dataset of stroke anatomical brain images and manual lesion segmentations. Sci. data (2018) 5

    2018

  22. [30]

    In: CVPR (2025) 2, 3, 4, 6, 8

    Liu, P., Lawry Aguila, A., Iglesias, J.E.: Unraveling normal anatomy via fluid- driven anomaly randomization. In: CVPR (2025) 2, 3, 4, 6, 8

    2025

  23. [31]

    In: ECCV (2024) 2

    Liu, P., Puonti, O., Hu, X., et al.: Brain-id: Learning contrast-agnostic anatomical representations for brain imaging. In: ECCV (2024) 2

    2024

  24. [32]

    In: MICCAI (2024) 2

    Liu, P., Puonti, O., Sorby-Adams, A., et al.: PEPSI: Pathology-enhanced pulse- sequence-invariant representations for brain MRI. In: MICCAI (2024) 2

    2024

  25. [33]

    arXiv (2022) 2, 5

    Pinaya, W., Graham, M., Gray, R., et al.: Fast unsupervised brain anomaly detec- tion and segmentation with diffusion models. arXiv (2022) 2, 5

    2022

  26. [34]

    In: Deep Generative Models

    Pinaya, W., Tudosiu, P.D., Dafflon, J., et al.: Brain imaging generation with latent diffusion models. In: Deep Generative Models. pp. 117–126 (2022) 5

    2022

  27. [35]

    arXiv (2021) 2

    Pinaya, W., Tudosiu, P.D., Gray, R., et al.: Unsupervised brain anomaly detection and segmentation with transformers. arXiv (2021) 2

    2021

  28. [36]

    arXiv (2022) 5

    Preechakul, K., Chatthee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion au- toencoders: Toward a meaningful and decodable representation. arXiv (2022) 5

    2022

  29. [37]

    arXiv (2017) 6

    Ramachandran, P., et al.: Searching for activation functions. arXiv (2017) 6

    2017

  30. [38]

    arXiv (2021) 3, 5, 6

    Rombach, R., Blattmann, A., Lorenz, D., et al.: High-resolution image synthesis with latent diffusion models. arXiv (2021) 3, 5, 6

    2021

  31. [39]

    In: MICCAI

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241 (2015) 5

    2015

  32. [40]

    arXiv (2017) 2

    Schlegl, T., Seeböck, P., Waldstein, S., et al.: Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. arXiv (2017) 2

    2017

  33. [41]

    MedIA (2019) 2

    Schlegl, T., Seeböck, P., Waldstein, S., et al.: f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. MedIA (2019) 2

    2019

  34. [42]

    Alzheimer’s & Dementia (2017) 5

    Weiner, M., Veitch, D., Aisen, P., et al.: The Alzheimer’s disease neuroimaging initiative 3. Alzheimer’s & Dementia (2017) 5

    2017

  35. [43]

    arXiv (2020) 2

    Zhou, K., Xiao, Y., Yang, J., et al.: Encoding structure-texture relation with p-net for anomaly detection in retinal images. arXiv (2020) 2

    2020

  36. [44]

    arXiv (2023) 2

    Zhou, Y.: Rethinking reconstruction autoencoder-based out-of-distribution detec- tion. arXiv (2023) 2

    2023

  37. [45]

    arXiv (2018) 2

    Zimmerer, D., Kohl, S., Petersen, J., et al.: Context-encoding variational autoen- coder for unsupervised anomaly detection. arXiv (2018) 2

    2018

  38. [46]

    In: ICLR (2018) 2

    Zong, B., Song, Q., Min, M.R., et al.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: ICLR (2018) 2

    2018

Pith tools

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