Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

A causal diffusion model with voxel-level anatomical masks generates brain MRIs whose counterfactuals replicate subtle disease-related cortical changes for the first time.

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

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

A mask-guided causal diffusion model generates 3D brain MRI counterfactuals whose cortical volume measurements match known alcohol-use-disorder effects, but those effects are inserted via fitted masks rather than discovered.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful architecture and careful quality experiments, but the headline AUD replication is an in-loop consistency check, not an independent finding. the 3 major comments →

arxiv 2509.09054 v1 pith:VFZ7C5RD submitted 2025-09-10 cs.CV

Integrating Anatomical Priors into a Causal Diffusion Model

Binxu Li , Wei Peng , Mingjie Li , Ehsan Adeli , Kilian M. Pohl This is my paper
classification cs.CV
keywords brain MRI synthesiscounterfactual generationdiffusion modelsprobabilistic causal graphanatomical priorsalcohol use disordercortical volume3D medical imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper introduces PCGM, a diffusion-based generator for 3D brain MRIs that builds explicit anatomical constraints into the generation process. It claims that, unlike prior counterfactual generators, PCGM can alter an MRI in a way that matches the subtle, locally restricted effects of a disease—here, alcohol use disorder—rather than just changing overall appearance. The authors demonstrate that standard cortical volume measurements taken from PCGM's counterfactual MRIs replicate the group differences between patients and controls previously reported in the neuroscience literature. If correct, this would make synthetic MRIs a viable tool for studying subtle morphological effects in psychiatric and neurological disorders, where data are scarce and expensive.

Core claim

PCGM combines three modules: a probabilistic graph module (PGM) that encodes known causal relations between metadata (age, sex, diagnosis) and regional brain volumes; a counterfactual mask generator (CMG) that converts the PGM's predicted volume changes into spatial binary masks by moving only the boundary between each cortical region and the cerebrospinal fluid; and a mask-guided diffusion module (MGD) that injects those masks into a 3D latent diffusion model via a 3D ControlNet. The central claim is that this voxel-level anatomical conditioning is what allows the model to preserve subtle, medically relevant local variations. Empirically, the paper shows that PCGM's unconditional MRIs score

What carries the argument

The Counterfactual Mask Generator (CMG) is the load-bearing component: it turns a scalar volume change for a cortical ROI into a binary mask that alters only the ROI–CSF boundary, preserving the gray–white boundary. That mask is then encoded by a 3D extension of ControlNet and used as voxel-level conditioning in a counterfactual denoising UNet, whose latent is decoded by a dedicated 3D diffusion decoder. The mask is the mechanism that forces the diffusion process to make the disease-related change in the right place.

Load-bearing premise

The central result rests on the anatomical premise that a small change in a cortical ROI's volume moves only the boundary between that ROI and the cerebrospinal fluid, never the boundary with white matter or neighboring ROIs; if real AUD-related tissue loss occurs at the gray–white boundary or is distributed, the masks would misplace the effect and the reported replication would be an artifact.

What would settle it

Take real longitudinal MRIs from individuals who developed AUD, segment them at both time points, and measure whether the observed volume change indeed occurs predominantly at the ROI–CSF boundary. If substantial changes occur at the ROI–white-matter boundary, the CMG's masking rule is falsified. A simpler test: generate counterfactuals with the mask applied to the gray–white boundary instead of the CSF boundary and check whether the measured group differences vanish.

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

If this is right

  • If PCGM is correct, standard morphometric pipelines can be applied to synthetic MRIs to test hypotheses about disease effects, without acquiring new patient scans.
  • Counterfactual MRIs generated from a single baseline scan can model disease progression (e.g., aging) with subject-specific trajectories.
  • The separation of causal volume prediction (PGM) from image-level rendering (MGD) means the model can generate counterfactuals for any metadata intervention supported by the graph.
  • The quality of unconditional synthesis is high enough that clinical experts cannot reliably distinguish real from synthetic MRIs, addressing a known pitfall of earlier generators.

Where Pith is reading between the lines

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

  • If the anatomical premise is transferred to other diseases or brain regions, the mask rule would need revalidation; one could test this by comparing counterfactuals against real longitudinal scans of patients who develop the condition.
  • The PGM is learned from cross-sectional data; whether the counterfactual volumes are truly 'causal' depends on the identifiability of the graph, which the paper does not establish.
  • A stronger validation would be to generate counterfactuals with the mask applied to the wrong tissue boundary (e.g., gray–white) and show that the replicated findings disappear, confirming that the mask placement, not just the label conditioning, drives the effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes PCGM, a three-module generative framework for 3D brain MRI counterfactuals. The Probabilistic Graph Module (PGM) learns a normalizing-flow SCM mapping metadata (age, sex, diagnosis) to cortical ROI volumes; the Counterfactual Mask Generator (CMG) converts intervened volume scores into binary masks by modifying only the CSF-side boundary of each ROI; and the Mask Guided Diffusion (MGD) module, based on a 3D ControlNet and a 3D diffusion decoder, generates counterfactual MRIs conditioned on masks and metadata. The authors evaluate three tasks: unconditional synthesis quality against six baselines, age-conditioned counterfactuals against longitudinal ventricle growth, and AUD-conditioned counterfactuals against the Sullivan et al. 2018 cortical-volume findings. The paper claims, for the first time, that measurements from generated counterfactual MRIs replicate subtle disease-related cortical effects.

Significance. If the central claim were supported, this would be a notable advance in counterfactual MRI generation: it demonstrates a concrete mechanism for injecting anatomical priors into a diffusion model and shows that the resulting images can preserve subtle regional morphology. The paper includes multiple positive features: a complete 3D architecture with a novel 3D diffusion decoder, comparison against six baselines on unconditional quality, a FreeSurfer-based anatomical plausibility analysis, a user study, and a longitudinal validation of aging counterfactuals. These components are valuable and largely plausible. However, the headline disease-replication claim is currently not convincing, because the measured effect is largely introduced by construction: the PGM is trained on the same AUD cohort, the CMG writes the PGM volumes into masks, and the outcome is measured with the same segmenter used to create those masks. The significance of the paper therefore depends on whether the authors can break this measurement loop or substantially reframe the claim.

major comments (3)
  1. [§4.4, Eq. (1)–(2), Table 3]
  2. [§3.2]
  3. [§4.4]
minor comments (5)
  1. [§4.4]
  2. [Table 1]
  3. [Table 3]
  4. [Conclusion]
  5. [Throughout]

Circularity Check

3 steps flagged

The headline AUD 'replication' is a fitted-value recovery: PGM is trained on the same cohort, the CMG writes the PGM volume change into the conditioning mask, and SynthSeg+ both defines the condition and measures the outcome.

specific steps
  1. fitted input called prediction [Section 4.4 (Task 3), PGM training]
    "For each test fold of the out-of-sample AUD data set, we then used the remaining data to train the PGM module, which creates a causal graph relating scalar variables (i.e., age, sex, diagnosis) to the volume of cortical regions (Frontal, Parietal, Insula, Cingulate and Temporal)."

    The PGM is trained on the same AUD/control cohort whose published findings ([18]) are the replication target. Its counterfactual ROI volumes are fitted conditional estimates of diagnosis-to-volume effects, not independent predictions. The later 'replication' in Table 3 therefore checks that fitted group differences survive synthesis and re-measurement; it does not test an external or emergent prediction. The abstract's 'for the first time ... replicate' overstates this in-the-loop consistency result.

  2. self definitional [Section 3.2 (Counterfactual Mask Generator)]
    "To increase the mask by d := v_PGM − v_orig voxels, we simulate ’increasing’ the voxel-level probabilities Pk(l = k) by a scalar α > 1 for that ROI until d voxels are added."

    The CMG is defined so that the modified ROI mask has volume v_PGM, the counterfactual volume produced by the PGM. Section 3.3 then conditions the diffusion model on this mask ('we further incorporate the counterfactual mask to directly guide the diffusion process on a voxel-level'). Measuring the ROI volume in the generated image and finding the control-vs-AUD difference is therefore recovering the very volume change that was written into the conditioning mask. The effect is injected by construction rather than emergent from the generative model.

  3. other [Section 3 (pipeline) and Section 4.4 (evaluation)]
    "Given a t1w brain MRI, the approach first extracts a volume score, mask, and probability map of each region of interest (ROI) via SynthSeg+ [57]."

    SynthSeg+ is the only described source of volume scores, masks, and probability maps. The CMG modifies these SynthSeg+ masks and probabilities, the MGD is conditioned on them, and the evaluation extracts 'cortical volume measurements' from the generated MRIs using the same pipeline. The measurement loop is closed: the segmenter that defines the conditioning mask is the same segmenter that measures the outcome. Thus Table 3 demonstrates consistency of a SynthSeg+-defined condition with a SynthSeg+-based measurement, not independent confirmation of a disease effect.

full rationale

The paper is self-contained against external benchmarks for image quality (Task 1) and ventricular aging (Task 2); those experiments compare against baselines and are not circular. The circularity is concentrated in the central disease-replication claim of the abstract and Section 4.4. The PGM is trained on the same AUD/control dataset whose published findings [18] are the replication target, so the counterfactual ROI volumes are fitted conditional estimates. The CMG then turns those volumes into binary masks by construction, and the MGD is voxel-wise conditioned on those masks; re-measuring the generated images and finding the group difference is therefore recovering the quantity that was injected through the mask. The measurement loop is closed because SynthSeg+ supplies both the mask probabilities used by the CMG and the volume scores used in evaluation. The 'for the first time' language overstates what is demonstrated: a consistency check that the diffusion decoder is faithful to a SynthSeg+-based mask condition, not an independent neurobiological confirmation. Score 8 rather than 10 because the task is non-trivially accomplished—the MGD does render the mask changes into plausible MRI intensities and a segmentation pipeline recovers them—but the scientific claim that the findings are 'replicated' reduces by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper's core mechanism rests on fitted PGM parameters and on anatomical/statistical assumptions about causal structure, segmentation accuracy, and boundary behavior. No new physical entities are postulated. The main free parameter is the PGM, whose fitted output is the disease effect that the headline experiment re-measures.

free parameters (2)
  • PGM normalizing-flow parameters for ROI volumes = learned via 5-fold training on the in-house AUD/control dataset
    Counterfactual ROI volume scores in Section 4.4 are outputs of these fitted functions; they determine the masks that drive the generated MRIs. The replicated AUD effect is therefore a fitted value.
  • MGD model weights (Causal Encoder, denoising UNet, 3D Diffusion Decoder, 3D ControlNet) = trained on 3,954 control MRIs from ADNI and NCANDA
    These parameters determine the image quality and anatomical plausibility reported in Tasks 1-3. They are standard learned parameters, not ad hoc constants, and do not by themselves introduce circularity.
axioms (5)
  • domain assumption Exogenous noises u_k are independent and mechanisms f_k are invertible normalizing flows.
    Section 3.1. Needed for the abduction-action-prediction counterfactual computation; if the noise model is misspecified, the counterfactual volumes are wrong.
  • domain assumption Diagnosis causally affects cortical ROI volumes with the specified graph and no unobserved confounders.
    Section 3.1 and 4.4. The PGM treats diagnosis as a parent of ROI volumes; this is assumed and not tested.
  • domain assumption A small volume change of a cortical ROI alters only the ROI-CSF boundary, not the white matter boundary.
    Section 3.2, citing [61]. The CMG mask modification relies on this anatomical rule; false placement would invalidate the counterfactual anatomy.
  • domain assumption SynthSeg+ segmentations and probability maps are accurate enough to define ROI boundaries and rank boundary voxels.
    Section 3.2. CMG uses SynthSeg+ probabilities to decide which CSF voxels become ROI; measurement also uses SynthSeg+ and FreeSurfer.
  • domain assumption DDIM inversion of the causal encoding preserves subject identity while permitting counterfactual edits.
    Section 3.4. The counterfactual generation starts from an intermediate latent of the original MRI; if inversion does not preserve anatomy, outputs are not subject-specific.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Anatomical Priors into a Causal Diffusion Model." pith.science (2026). https://pith.science/paper/VFZ7C5RD

@misc{pith2026250909054,
  author       = {Pith},
  title        = {Pith review of: Integrating Anatomical Priors into a Causal Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFZ7C5RD}},
  note         = {Machine review of arXiv:2509.09054}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D brain MRI studies often examine subtle morphometric differences between cohorts that are hard to detect visually. Given the high cost of MRI acquisition, these studies could greatly benefit from image syntheses, particularly counterfactual image generation, as seen in other domains, such as computer vision. However, counterfactual models struggle to produce anatomically plausible MRIs due to the lack of explicit inductive biases to preserve fine-grained anatomical details. This shortcoming arises from the training of the models aiming to optimize for the overall appearance of the images (e.g., via cross-entropy) rather than preserving subtle, yet medically relevant, local variations across subjects. To preserve subtle variations, we propose to explicitly integrate anatomical constraints on a voxel-level as prior into a generative diffusion framework. Called Probabilistic Causal Graph Model (PCGM), the approach captures anatomical constraints via a probabilistic graph module and translates those constraints into spatial binary masks of regions where subtle variations occur. The masks (encoded by a 3D extension of ControlNet) constrain a novel counterfactual denoising UNet, whose encodings are then transferred into high-quality brain MRIs via our 3D diffusion decoder. Extensive experiments on multiple datasets demonstrate that PCGM generates structural brain MRIs of higher quality than several baseline approaches. Furthermore, we show for the first time that brain measurements extracted from counterfactuals (generated by PCGM) replicate the subtle effects of a disease on cortical brain regions previously reported in the neuroscience literature. This achievement is an important milestone in the use of synthetic MRIs in studies investigating subtle morphological differences.

Figures

Figures reproduced from arXiv: 2509.09054 by Binxu Li, Ehsan Adeli, Kilian M. Pohl, Mingjie Li, Wei Peng.

Figure 1
Figure 1. Figure 1: The framework of our Probabilistic Causal Graph Model (PCGM). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Real and synthetic MRIs created by generative approaches. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Age trajectory modeling. Each trajectory in the plot represents the volume of the ventricle of one subject measured at different ages. The subjects were participants of NCANDA (ages 14 - 26 years), ADNI (ages 60 - 90 years), or diagnosed with HIV (ages 33 - 74 years). Among the three methods, the trajectories extracted from the counterfactual MRIs produced by our proposed CDM align the best (i.e., lowest N… view at source ↗
Figure 4
Figure 4. Figure 4: Anatomical Plausibility. Displayed are the natural logarithm of the absolute Cohen’s d (|d|) for 32 regions and three methods. |d|< 0.2 is viewed as a small effect, 0.2 <|d|< 0.5 indicates a medium effect, and |d| > 0.5 corresponds to a large effect. Our method records the smallest d-score for the majority of regions, indicating the highest overall anatomical plausibility among the 3 methods. Note, the plo… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Modality-Aware and Anatomical Vector-Quantized Autoencoding for Multimodal Brain MRI

    cs.CV 2026-04 unverdicted novelty 7.0

    NeuroQuant is a modality-aware 3D VQ-VAE that uses dual-stream encoding, a shared anatomical codebook, and FiLM to achieve superior multi-modal brain MRI reconstruction.

Reference graph

Works this paper leans on

79 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    MRI of the neonatal brain: a review of methodological challenges and neuroscientific advances,

    J. Dubois, M. Alison, S. J. Counsell, L. Hertz-Pannier, P. S. H¨ uppi, and M. J. Benders, “MRI of the neonatal brain: a review of methodological challenges and neuroscientific advances,”Journal of MRI, vol. 53, no. 5, pp. 1318–1343, 2021

    2021

  2. [2]

    Association of brain atrophy with disease progression independent of relapse activity in patients with relapsing multiple sclerosis,

    A. Cagolet al., “Association of brain atrophy with disease progression independent of relapse activity in patients with relapsing multiple sclerosis,”JAMA neurology, vol. 79, no. 7, pp. 682–692, 2022

    2022

  3. [3]

    Artificial intelligence assistive software tool for automated detection and quantification of amyloid-related imaging abnormalities,

    D. M. Simaet al., “Artificial intelligence assistive software tool for automated detection and quantification of amyloid-related imaging abnormalities,”JAMA Network Open, vol. 7, no. 2, 2024, Art. no. e2355800

    2024

  4. [4]

    Missing MRI Pulse Sequence Synthesis Using Multi-Modal Generative Ad- versarial Network,

    A. Sharma and G. Hamarneh, “Missing MRI Pulse Sequence Synthesis Using Multi-Modal Generative Ad- versarial Network,”IEEE Transactions on Medical Imaging, vol. 39, no. 4, pp. 1170–1183, 2020

    2020

  5. [6]

    Ensuring useful adoption of generative artificial intelligence in healthcare,

    J. A. Jindal, M. P. Lungren, and N. H. Shah, “Ensuring useful adoption of generative artificial intelligence in healthcare,”Journal of the American Medical Informatics Association, vol. 31, no. 6, pp. 1441–1444, 2024

    2024

  6. [7]

    Self-improving generative foundation model for synthetic medical image generation and clinical applications,

    J. Wanget al., “Self-improving generative foundation model for synthetic medical image generation and clinical applications,”Nature Medicine, vol. 31, no. 2, pp. 609–617, 2025

    2025

  7. [8]

    Toward general text-guided multimodal brain MRI synthesis for diagnosis and medical image analysis,

    Y. Wanget al., “Toward general text-guided multimodal brain MRI synthesis for diagnosis and medical image analysis,”Cell Reports Medicine, vol. 6, no. 6, 2025, Art. no. 102182

    2025

  8. [9]

    A vision–language foundation model for the generation of realistic chest X-ray images,

    C. Bluethgenet al., “A vision–language foundation model for the generation of realistic chest X-ray images,” Nature Biomedical Engineering, vol. 9, no. 4, pp. 494–506, 2024

    2024

  9. [10]

    Realistic morphology-preserving generative modelling of the brain,

    P.-D. Tudosiuet al., “Realistic morphology-preserving generative modelling of the brain,”Nature Machine Intelligence, vol. 6, no. 7, pp. 811–819, 2024

    2024

  10. [11]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 10 684–10 695

    2022

  11. [12]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esseret al., “Scaling rectified flow transformers for high-resolution image synthesis,” inProceedings of the 41st International Conference on Machine Learning, 2024, pp. 12 606–12 633

    2024

  12. [13]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 4195–4205

    2023

  13. [14]

    Evaluating the Quality of Brain MRI Generators,

    J. Wu, W. Peng, B. Li, Y. Zhang, and K. M. Pohl, “Evaluating the Quality of Brain MRI Generators,” inMedical Image Computing and Computer Assisted Intervention, Lecture Notes in Computer Science, vol. 15010, 2024, pp. 297–307

    2024

  14. [15]

    Medical image generation using generative adversarial networks: A review,

    N. K. Singh and K. Raza, “Medical image generation using generative adversarial networks: A review,” in Health informatics: A computational perspective in healthcare, 2021, pp. 77–96

    2021

  15. [16]

    Structural brain imaging in Alzheimer’s disease and mild cognitive impairment: biomarker analysis and shared morphometry database,

    C. Ledig, A. Schuh, R. Guerrero, R. A. Heckemann, and D. Rueckert, “Structural brain imaging in Alzheimer’s disease and mild cognitive impairment: biomarker analysis and shared morphometry database,”Scientific reports, vol. 8,, no. 1, 2018, Art. no. 11258

    2018

  16. [17]

    Multi-label, multi-domain learning identifies compounding effects of HIV and cognitive impairment,

    J. Zhanget al., “Multi-label, multi-domain learning identifies compounding effects of HIV and cognitive impairment,”Medical Image Analysis, vol. 75, 2022, Art. no. 102246. Integrating Anatomical Priors into a Causal Diffusion Model 13

    2022

  17. [18]

    The role of aging, drug dependence, and hepatitis C comorbidity in alcoholism cortical compromise,

    E. V. Sullivanet al., “The role of aging, drug dependence, and hepatitis C comorbidity in alcoholism cortical compromise,”JAMA Psychiatry, vol. 75, no. 5, pp. 474–483, 2018

    2018

  18. [19]

    Evaluation of 3D counterfactual brain MRI generation,

    P. Sun, W. Peng, L. Li, Y. Wang, and K. M. Pohl, “Evaluation of 3D counterfactual brain MRI generation,”Deep Generative Models, Lecture Notes in Computer Science, Accepted. [Online]. Available: https://arxiv.org/abs/2508.02880

    Pith/arXiv arXiv

  19. [20]

    Counterfactual MRI Generation with Denoising Diffusion Models for Interpretable Alzheimer’s Disease Effect Detection,

    N. J. Dhinagar, S. I. Thomopoulos, E. Laltoo, and P. M. Thompson, “Counterfactual MRI Generation with Denoising Diffusion Models for Interpretable Alzheimer’s Disease Effect Detection,” inIEEE Engineering in Medicine and Biology Society (EMBC), 2024, pp. 1–6

    2024

  20. [21]

    Brain latent progression: Individual-based spatiotemporal disease progression on 3D Brain MRIs via latent diffusion,

    L. Puglisi, D. C. Alexander, and D. Rav ` ı, “Brain latent progression: Individual-based spatiotemporal disease progression on 3D Brain MRIs via latent diffusion,”Medical Image Analysis, vol. 106, 2025, Art. no. 103734

    2025

  21. [22]

    High fidelity image counterfactuals with probabilistic causal models,

    F. De Sousa Ribeiro, T. Xia, M. Monteiro, N. Pawlowski, and B. Glocker, “High fidelity image counterfactuals with probabilistic causal models,” inInternational Conference on Machine Learning, vol. 202, 2023, pp. 7390– 7425

    2023

  22. [23]

    Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis,

    Y. Yeganehet al., “Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 7685–7695

    2025

  23. [24]

    Latent Causal Modeling for 3D Brain MRI Counterfactuals,

    W. Penget al., “Latent Causal Modeling for 3D Brain MRI Counterfactuals,”Deep Generative Models, Lecture Notes in Computer Science, Accepted. [Online]. Available: https://arxiv.org/abs/2409.05585

    arXiv

  24. [25]

    Ventricular enlargement as a possible measure of Alzheimer’s disease progression vali- dated using the Alzheimer’s disease neuroimaging initiative database,

    S. M. Nestoret al., “Ventricular enlargement as a possible measure of Alzheimer’s disease progression vali- dated using the Alzheimer’s disease neuroimaging initiative database,”Brain, vol. 131, no. 9, pp. 2443–2454, 2008

    2008

  25. [26]

    Adding Conditional Control to Text-to-Image Diffusion Models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding Conditional Control to Text-to-Image Diffusion Models,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 3836–3847

    2023

  26. [28]

    Alzheimer’s disease Neuroimaging Initiative (ADNI) clinical characterization,

    R. C. Petersenet al., “Alzheimer’s disease Neuroimaging Initiative (ADNI) clinical characterization,”Neu- rology, vol. 74, no. 3, pp. 201–209, 2010

    2010

  27. [29]

    The National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA): a multisite study of adolescent development and substance use,

    S. A. Brownet al., “The National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA): a multisite study of adolescent development and substance use,”Journal of studies on alcohol and drugs, vol. 76, no. 6, pp. 895–908, 2015

    2015

  28. [30]

    Novel Machine Learning Identifies Brain Patterns Distinguishing Diagnostic Membership of Human Immunodeficiency Virus, Alcoholism, and Their Comorbidity of Individuals,

    E. Adeli, N. M. Zahr, A. Pfefferbaum, E. V. Sullivan, and K. M. Pohl, “Novel Machine Learning Identifies Brain Patterns Distinguishing Diagnostic Membership of Human Immunodeficiency Virus, Alcoholism, and Their Comorbidity of Individuals,”Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, vol. 4, no. 6, pp. 589–599, 2019

    2019

  29. [31]

    Image registration methods: A survey,

    B. Zitova and J. Flusser, “Image registration methods: A survey,”Image and vision computing, vol. 21, no. 11, pp. 977–1000, 2003

    2003

  30. [32]

    Accurate registration of serial 3D MR brain images and its application to visualizing change in neurodegenerative disorders,

    P. A. Freeborough, R. P. Woods, and N. C. Fox, “Accurate registration of serial 3D MR brain images and its application to visualizing change in neurodegenerative disorders,”Journal of computer assisted tomography, vol. 20, no. 6, pp. 1012–1022, 1996

    1996

  31. [33]

    A survey of medical image registration,

    J. A. Maintz and M. A. Viergever, “A survey of medical image registration,”Medical image analysis, vol. 2, no. 1, pp. 1–36, 1998

    1998

  32. [34]

    Medical image registration,

    D. L. Hill, P. G. Batchelor, M. Holden, and D. J. Hawkes, “Medical image registration,”Physics in medicine & biology, vol. 46, no. 3, 2001, Art. no. R1

    2001

  33. [35]

    Score-based diffusion models for accelerated MRI,

    H. Chung and J. C. Ye, “Score-based diffusion models for accelerated MRI,”Medical Image Analysis, vol. 80, 2022, Art. no. 102479

    2022

  34. [36]

    Modelling the Distribution of 3D Brain MRI Using a 2D Slice V AE,

    A. Volokitinet al., “Modelling the Distribution of 3D Brain MRI Using a 2D Slice V AE,” inMedical Image Computing and Computer Assisted Intervention, Lecture Notes in Computer Science, vol. 12267, 2020, pp. 657–666

    2020

  35. [37]

    Image synthesis in multi-contrast MRI with conditional generative adversarial networks,

    S. U. Dar, M. Yurt, L. Karacan, A. Erdem, E. Erdem, and T. Cukur, “Image synthesis in multi-contrast MRI with conditional generative adversarial networks,”IEEE Transactions on Medical Imaging, vol. 38, no. 10, pp. 2375–2388, 2019

    2019

  36. [38]

    Medical image synthesis for data augmentation and anonymization using generative adversarial networks,

    H.-C. Shinet al., “Medical image synthesis for data augmentation and anonymization using generative adversarial networks,” inSimulation and Synthesis in Medical Imaging, Lecture Notes in Computer Science, vol. 11037, 2018, pp. 1–11

    2018

  37. [39]

    3D cGAN based cross-modality MR image synthesis for brain tumor segmentation,

    B. Yu, L. Zhou, L. Wang, J. Fripp, and P. Bourgeat, “3D cGAN based cross-modality MR image synthesis for brain tumor segmentation,” in2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018, pp. 626–630

    2018

  38. [40]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 12, pp. 4217–4228, 2021

    2021

  39. [41]

    Cycle consistent embedding of 3D brains with auto-encoding generative adversarial networks,

    S. Xing, H. Sinha, and S. J. Hwang, “Cycle consistent embedding of 3D brains with auto-encoding generative adversarial networks,” inMedical Imaging with Deep Learning, 2021, pp. 118–126

    2021

  40. [42]

    Learning implicit brain MRI manifolds with deep learning,

    C. Bermudez, A. J. Plassard, L. T. Davis, A. T. Newton, S. M. Resnick, and B. A. Landman, “Learning implicit brain MRI manifolds with deep learning,” inMedical Imaging 2018: Image Processing, vol. 10574. SPIE, 2018, pp. 408–414. 14 B. Li et al

    2018

  41. [43]

    GAN-based synthetic brain MR image generation,

    C. Hanet al., “GAN-based synthetic brain MR image generation,” in2018 IEEE 15th International Sympo- sium on Biomedical Imaging (ISBI 2018), 2018, pp. 734–738

    2018

  42. [44]

    Generation of 3D Brain MRI Using Auto-Encoding Generative Adversarial Networks,

    G. Kwon, C. Han, and D.-s. Kim, “Generation of 3D Brain MRI Using Auto-Encoding Generative Adversarial Networks,” inMedical Image Computing and Computer Assisted Intervention, Lecture Notes in Computer Science, vol. 9, 2019, pp. 118–126

    2019

  43. [45]

    Degenerative adversarial neuroimage nets for brain scan simulations: Application in ageing and dementia,

    D. Raviet al., “Degenerative adversarial neuroimage nets for brain scan simulations: Application in ageing and dementia,”Medical image analysis, vol. 75, 2022, Art. no. 102257

    2022

  44. [46]

    Metadata-conditioned generative models to synthesize anatomically-plausible 3D brain MRIs,

    W. Penget al., “Metadata-conditioned generative models to synthesize anatomically-plausible 3D brain MRIs,”Medical Image Analysis, vol. 98, 2024, Art. no. 103325

    2024

  45. [47]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851

    2020

  46. [48]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” inAdvances in Neural Infor- mation Processing Systems, vol. 34, 2021, pp. 8780–8794

    2021

  47. [49]

    Anatomically constrained CT image translation for heterogeneous blood vessel seg- mentation,

    G. La Barberaet al., “Anatomically constrained CT image translation for heterogeneous blood vessel seg- mentation,” inBritish Machine Vision Virtual Conference, 2022, Art. no. 776

    2022

  48. [50]

    Diffusion Models for Medical Anomaly Detection,

    J. Wolleb, F. Bieder, R. Sandk¨ uhler, and P. C. Cattin, “Diffusion Models for Medical Anomaly Detection,” inMedical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science, vol. 13438, 2022, pp. 35–45

    2022

  49. [51]

    Three-dimensional medical image synthesis with denoising diffusion probabilistic models,

    Z. Dorjsembe, S. Odonchimed, and F. Xiao, “Three-dimensional medical image synthesis with denoising diffusion probabilistic models,” inMedical Imaging with Deep Learning, 2022. [Online]. Available: https://openreview.net/forum?id=Oz7lKWVh45H

    2022

  50. [52]

    Generating realistic brain MRIs via a conditional diffusion probabilistic model,

    W. Peng, E. Adeli, T. Bosschieter, S. H. Park, Q. Zhao, and K. M. Pohl, “Generating realistic brain MRIs via a conditional diffusion probabilistic model,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention, Lecture Notes in Computer Science, vol. 14227, 2023, pp. 14–24

    2023

  51. [53]

    Sadm: Sequence-aware diffusion model for longitudinal medical image generation,

    J. S. Yoon, C. Zhang, H.-I. Suk, J. Guo, and X. Li, “Sadm: Sequence-aware diffusion model for longitudinal medical image generation,” inInternational Conference on Information Processing in Medical Imaging, 2023, pp. 388–400

    2023

  52. [54]

    MedGen3D: A deep generative framework for paired 3D image and mask generation,

    K. Hanet al., “MedGen3D: A deep generative framework for paired 3D image and mask generation,” in Medical Image Computing and Computer Assisted Intervention, Lecture Notes in Computer Science, vol. 14220, 2023, pp. 759–769

    2023

  53. [55]

    Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3D deep generative models,

    G. Pomboet al., “Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3D deep generative models,”Medical Image Analysis, vol. 84, 2023, Art. no. 102723

    2023

  54. [56]

    Training calibration-based counterfactual explain- ers for deep learning models in medical image analysis,

    J. J. Thiagarajan, K. Thopalli, D. Rajan, and P. Turaga, “Training calibration-based counterfactual explain- ers for deep learning models in medical image analysis,”Scientific reports, vol. 12, no. 1, 2022, Art. no. 597

    2022

  55. [57]

    Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets,

    B. Billot, Y. Colin, Magdamo Cheng, S. Das, and J. E. Iglesias, “Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets,”Proceedings of the National Academy of Sciences (PNAS), vol. 120, no. 9, 2023, Art. no. e2216399120

    2023

  56. [58]

    Deep structural causal models for tractable counter- factual inference,

    N. Pawlowski, D. Coelho de Castro, and B. Glocker, “Deep structural causal models for tractable counter- factual inference,”Advances in Neural Information Processing Systems, vol. 33, pp. 857–869, 2020

    2020

  57. [59]

    Variational inference with normalizing flows,

    D. Rezende and S. Mohamed, “Variational inference with normalizing flows,” inInternational Conference on Machine Learning, vol. 37, 2015, pp. 1530–1538

    2015

  58. [60]

    High fidelity image counterfactuals with probabilistic causal models,

    F. De Sousa Ribeiro, T. Xia, M. Monteiro, N. Pawlowski, and B. Glocker, “High fidelity image counterfactuals with probabilistic causal models,” inInternational Conference on Machine Learning, 2023, pp. 7390–7425

    2023

  59. [61]

    Reduced cerebral grey matter observed in alcoholics using magnetic resonance imaging,

    T. L. Jerniganet al., “Reduced cerebral grey matter observed in alcoholics using magnetic resonance imaging,” Alcoholism: Clinical and Experimental Research, vol. 15, no. 3, pp. 418–427, 1991

    1991

  60. [62]

    Od-vae: An omni-dimensional video compressor for improving latent video diffusion model,

    L. Chenet al., “Od-vae: An omni-dimensional video compressor for improving latent video diffusion model,” arXiv preprint arXiv:2409.01199, 2024. [Online]. Available: https://arxiv.org/abs/2409.01199

    Pith/arXiv arXiv 2024

  61. [63]

    Frozen in time: A joint video and image encoder for end- to-end retrieval,

    M. Bain, A. Nagrani, G. Varol, and A. Zisserman, “Frozen in time: A joint video and image encoder for end- to-end retrieval,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 1728–1738

    2021

  62. [64]

    MedSyn: Text-guided anatomy-aware synthesis of high-fidelity 3D CT images,

    Y. Xuet al., “MedSyn: Text-guided anatomy-aware synthesis of high-fidelity 3D CT images,”IEEE Trans- actions on Medical Imaging, vol. 43, no. 10, pp. 3648–3660, 2024

    2024

  63. [65]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 3836–3847

    2023

  64. [66]

    Diffusion autoencoders: Toward a meaningful and decodable representation,

    K. Preechakul, N. Chatthee, S. Wizadwongsa, and S. Suwajanakorn, “Diffusion autoencoders: Toward a meaningful and decodable representation,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 10 609–10 619

    2022

  65. [67]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” in2nd International Conference on Learning Representations, Y. Bengio and Y. LeCun, Eds., 2014. [Online]. Available: http: //arxiv.org/abs/1312.6114

    Pith/arXiv arXiv 2014

  66. [68]

    Null-text inversion for editing real images using guided diffusion models,

    R. Mokady, A. Hertz, K. Aberman, Y. Pritch, and D. Cohen-Or, “Null-text inversion for editing real images using guided diffusion models,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 6038–6047. Integrating Anatomical Priors into a Causal Diffusion Model 15

    2023

  67. [69]

    Neural Discrete Representation Learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural Discrete Representation Learning,” inAdvances in Neural Information Processing Systems, 2017, pp. 6309–6318

    2017

  68. [70]

    Generation of Synthetic Rat Brain MRI Scans with a 3D Enhanced Alpha Generative Adversarial Network,

    A. Ferreira, R. Magalh˜ aes, S. M´ eriaux, and V. Alves, “Generation of Synthetic Rat Brain MRI Scans with a 3D Enhanced Alpha Generative Adversarial Network,”Applied Sciences, vol. 12, no. 10, 2022, Art. no., 4844

    2022

  69. [71]

    Hierarchical amortized GAN for 3D high resolution medical image synthesis,

    L. Sun, J. Chen, Y. Xu, M. Gong, K. Yu, and K. Batmanghelich, “Hierarchical amortized GAN for 3D high resolution medical image synthesis,”IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 8, pp. 3966–3975, 2022

    2022

  70. [72]

    Brain imaging generation with latent diffusion models,

    W. H. Pinayaet al., “Brain imaging generation with latent diffusion models,” inDeep Generative Models, Lecture Notes in Computer Science, vol. 13609, 2022, pp. 117–126

    2022

  71. [73]

    Alzheimer’s Disease Neuroimaging Initiative (ADNI): clinical characterization,

    R. C. Petersenet al., “Alzheimer’s Disease Neuroimaging Initiative (ADNI): clinical characterization,”Neu- rology, vol. 74, no. 3, pp. 201–209, 2010

    2010

  72. [74]

    Classifier-Free Diffusion Guidance,

    J. Ho and T. Salimans, “Classifier-Free Diffusion Guidance,” inNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. [Online]. Available: https://openreview.net/forum? id=qw8AKxfYbI

    2021

  73. [75]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two time-scale update rule converge to a local Nash equilibrium,”Advances in Neural Information Processing Systems, vol. 30, pp. 6626–6637, 2017

    2017

  74. [76]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. Simoncelli, and A. Bovik, “Multiscale structural similarity for image quality assessment,” inThe Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, vol. 2, 2003, pp. 1398–1402

    2003

  75. [77]

    FreeSurfer,

    B. Fischl, “FreeSurfer,”NeuroImage, vol. 62, no. 2, pp. 774–781, 2012

    2012

  76. [78]

    The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites,

    B. J. Caseyet al., “The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites,”Developmental cognitive neuroscience, vol. 32, pp. 43–54, 2018

    2018

  77. [79]

    UK Biobank: An open access resource for identifying the causes of a wide range of complex diseases of middle and old age,

    C. Sudlowet al., “UK Biobank: An open access resource for identifying the causes of a wide range of complex diseases of middle and old age,”PLoS medicine, vol. 12, no. 3, 2015, Art. no. e1001779

    2015

  78. [80]

    The Human Connectome Project: A data acquisition perspective,

    D. C. Van Essenet al., “The Human Connectome Project: A data acquisition perspective,”Neuroimage, vol. 62, no. 4, pp. 2222–2231, 2012

    2012

  79. [81]

    Using effect size—or why the P value is not enough,

    G. M. Sullivan and R. Feinn, “Using effect size—or why the P value is not enough,”Journal of Graduate Medical Education, vol. 4, no. 3, pp. 279–282, 2012

    2012

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.