Pith. sign in

REVIEW 6 major objections 5 minor 35 references

Hierarchical Diffusion Framework for Pseudo-Healthy Brain MRI Inpainting with Enhanced 3D Consistency

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-stage axial-then-coronal diffusion pipeline can inpaint pseudo-healthy brain MRIs with 3D consistency, outperforming slice-wise and full-3D baselines on data-scarce medical benchmarks.

desk verdict Useful engineering paper on pseudo-healthy MRI inpainting with a genuine two-stage perpendicular diffusion pipeline, but the load-bearing z-detail claim needs a sharper ablation and thinner-metric analysis. read the letter →

arxiv 2507.17911 v1 pith:4M2RNDYZ submitted 2025-07-23 eess.IV cs.CV

Dou Hoon Kwark , Shirui Luo , Xiyue Zhu , Yudu Li , Zhi-Pei Liang , Volodymyr Kindratenko This is my paper
classification eess.IVcs.CV
keywords diffusionmodelbrainMRIinpaintingpseudo-healthysynthesisvolumetricconsistency3Dmedicalimagingcoarse-to-finegenerationtissue-awareattention
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

Pseudo-healthy inpainting replaces pathological regions in brain MRI with synthetic healthy tissue so that downstream tools designed for healthy brains can be used on lesioned scans. The paper argues that a hierarchical two-stage diffusion pipeline—axial inpainting followed by coronal refinement—can achieve volumetric consistency without the large training data required by full 3D generative models. On two public benchmark MRI datasets, the method is reported to outperform slice-wise and 3D baselines in reconstruction realism, volumetric consistency, and downstream tissue segmentation, with the largest gains on large lesions. If correct, this gives data-scarce medical settings a practical route to 3D-consistent synthesis using only 2D diffusion networks.

What carries the argument

The load-bearing object is the two-stage perpendicular coarse-to-fine pipeline: (1) axial 2D diffusion with depth-wise 1D convolutions and adaptive z-axis resampling, followed by cubic interpolation back to the original slice count; (2) coronal 2D diffusion trained to predict only the residual between the interpolated volume and the true anatomy. The adaptive resampling step keeps full volumetric context within a GPU-memory ceiling by increasing slice spacing rather than dropping slices, and the Tissue-Aware Attention Module—token-based cross-slice self-attention combined with gated convolution—lets the refinement stage adaptively emphasize tissue-specific features. Together these components turn '3D consistency' into a cheap 2D residual-refinement problem.

What would settle it

Measure the axial (superior–inferior) high-frequency content inside the inpainted mask before and after the coronal refinement stage on volumes containing thin structures such as cortical sulci or small vessels; if the coronal stage does not measurably increase z-axis sharpness beyond the interpolated axial output, the residual-refinement assumption is falsified.

Watch

Extended reading notes

Core claim

The central claim is that 3D consistency in pseudo-healthy brain MRI inpainting can be decomposed into two perpendicular 2D diffusion tasks rather than learned by an explicit 3D generator. An axial model first fills the masked region on a z-resampled volume to obtain globally coherent structure, the volume is interpolated back to its original slice count, and a coronal model then predicts only the residual needed to recover superior–inferior detail. A Tissue-Aware Attention Module in the coronal stage mixes slice-level features according to tissue type, which the paper shows improves white matter, gray matter, and cerebrospinal fluid Dice scores. Reported results on two benchmarks show the method surpassing the prior slice-wise and full-3D baselines in masked-region SSIM, PSNR, and MSE and in downstream segmentation.

Load-bearing premise

The load-bearing premise is that a coronal 2D diffusion model trained to predict only the residual can restore fine superior–inferior detail that interpolation blurs; if thin structures such as cortical sulci or small vessels remain blurred, the claimed 3D consistency without explicit 3D modeling does not follow.

Editorial extensions

If this is right

  • Data-scarce volumetric imaging tasks can pursue 3D-consistent synthesis with 2D diffusion models rather than training expensive 3D generators.
  • Large lesions that span many slices receive full global context despite a fixed GPU memory ceiling, so inpainting quality no longer degrades sharply with lesion size.
  • The residual formulation means the second stage only corrects remaining high-frequency errors, making the refinement computationally lightweight and easier to retrain on new data.
  • Downstream tissue segmentation in the inpainted region improves, which supports use of pseudo-healthy inpainting as a preprocessing step in automated neuroimaging pipelines.

Reading between the lines

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

  • The same axial-then-coronal residual strategy could transfer to other sparse-data 3D synthesis tasks such as cross-modal MRI translation or lesion filling in CT, since nothing in the design is T1-specific.
  • Because adaptive resampling changes slice spacing, the axial model is effectively trained under variable anisotropic resolution; one testable extension is to evaluate how performance degrades as the resampling factor grows, predicting a graceful rather than abrupt failure curve.
  • The Tissue-Aware Attention Module's benefit is measured through segmentation Dice in the masked region, suggesting a direct test: hold the pipeline fixed and replace TAM with ordinary cross-slice attention to see whether tissue-boundary sharpness, rather than overall fidelity, is what changes.
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

6 major / 5 minor

Summary. This paper proposes a two-stage 2D diffusion framework for pseudo-healthy brain MRI inpainting. The first stage is an axial diffusion model with depth-wise 1D convolutions and adaptive z-resampling, producing a coarse but globally consistent inpainting; the second stage is a coronal residual diffusion model, augmented with a Tissue-Aware Attention Module (TAM), intended to restore fine superior-inferior details lost during interpolation. The authors claim that this hierarchy replaces data-hungry 3D networks while achieving volumetric consistency, and report comparisons against Autoencoder 3D, Pix2Pix 3D, 3D DDPM, and Make-A-Volume on BraTS 2023 and HCP datasets, including downstream segmentation with FAST and FastSurfer. Code is released.

Significance. If the central claim is validated, the contribution is practically valuable: a data-efficient alternative to full 3D diffusion for volumetric inpainting, especially in medical settings with limited training data. The release of code, fixed dataset splits, and the use of downstream segmentation as an evaluation proxy are strengths. However, the evidence currently does not fully support the strong claims of 'outperforming state-of-the-art baselines in both realism and volumetric consistency': the residual-refinement mechanism is not directly tested, the ablation does not isolate the coronal stage, and several evaluation gaps remain. The reported gains on BraTS are small and no significance tests are provided.

major comments (6)
  1. [Section 3.3 / Table 1] The central 'outperforms' claim is not supported by statistical significance testing. On BraTS, the improvement over Make-A-Volume is small: SSIM 0.805 vs 0.793, PSNR 22.039 vs 21.912, with overlapping standard deviations. The authors should report paired significance tests (e.g., Wilcoxon signed-rank or paired bootstrap) with confidence intervals for SSIM, PSNR, MSE, and Dice scores across the 213/251 test subjects, or temper the claim accordingly.
  2. [Section 2.3] The load-bearing residual-refinement assumption is not directly tested. During training, the coronal refinement model is conditioned on xm = Blur(x0) * m + x0 * (1 - m), but during inference it is conditioned on the cubic-interpolated axial-stage output, which is not a Gaussian blur of the ground truth and whose blur is mask-dependent and anisotropic. The paper reports no analysis of this train/inference mismatch and no structure-size or slice-thickness stratified evaluation. As a result, the claim that the coronal stage restores fine superior-inferior detail remains unverified.
  3. [Section 3.4 / Table 3] The ablation study only toggles the TAM module; it does not ablate the coronal refinement stage, the adaptive resampling, or the axial stage. Consequently, the improvement in Dice scores shown in Table 3 cannot be attributed to the perpendicular coarse-to-fine mechanism. The authors should add ablations that remove the coronal stage and/or replace the residual formulation with a non-residual coronal model, and report image-quality metrics alongside Dice scores.
  4. [Section 2.2 / Section 3.3] The claim of 'volumetric consistency' is not measured directly. The reported metrics (SSIM, PSNR, MSE, segmentation Dice) are all computed within the masked region and do not quantify inter-slice continuity, out-of-plane coherence, or the absence of slice-wise discontinuities. The authors should either include a dedicated volumetric-consistency metric (e.g., inter-slice gradient statistics, slice-to-slice agreement, or a 3D structural metric) or revise the wording so that 'volumetric consistency' refers only to downstream anatomical plausibility rather than measured 3D coherence.
  5. [Section 3.1] The HCP evaluation uses synthetic inpainting masks obtained by transplanting BraTS tumor masks into tumor-free HCP volumes. This testbed, while convenient for obtaining ground truth, does not establish performance on real large lesions in clinical data. The paper's claim that the method 'excels when broader structural context is required' is therefore supported only by a synthetic proxy. The authors should acknowledge this limitation explicitly and, if possible, include a small real-lesion dataset or clearly restrict the claim to the synthetic benchmark.
  6. [Section 3.2] The training setups of the baseline models are not described. It is unclear whether Autoencoder 3D, Pix2Pix 3D, 3D DDPM, and Make-A-Volume were retrained on the same training splits, with comparable iterations, learning rates, and conditioning schemes. Without this information, a fair comparison cannot be established. Additionally, reference [16] (Hinton and Salakhutdinov 2006) does not clearly correspond to a '3D Autoencoder' and reference [19] is for 2D Pix2Pix, not 'Pix2Pix 3D'; the authors should clarify the exact baseline architectures and training protocols.
minor comments (5)
  1. [Section 2.3] The text states that the coronal model restores 'anterior-posterior details,' but coronal slices display the left-right and superior-inferior axes; the anterior-posterior axis is perpendicular to the coronal slice plane. This should be corrected to 'superior-inferior details.'
  2. [Table 1] There are typographical spacing issues in the table, e.g., '16 .385', '17 .333', '14 .828', which should be fixed to standard decimal formatting.
  3. [Section 3.2] The value of Zmax for the coronal stage is not specified; the text mentions batch size 16 for stage 2 but does not state whether Zmax is also 16 for that stage. This should be clarified.
  4. [Section 3.3 / Figure 2] The qualitative comparison in Figure 2 would benefit from zoomed-in crops of the masked region, as the reported stripe artifacts and fine anatomical details are difficult to discern at the current scale.
  5. [Section 2.3 / Table 3] The name 'Tissue-Aware Attention Module' may overstate what the module does, since it does not use tissue-type labels or a tissue-specific loss; the evidence for tissue awareness is indirect through downstream segmentation Dice. Consider renaming or adding a direct analysis of tissue-boundary fidelity.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found; the framework is an empirical combination of standard diffusion components, evaluated against external baselines.

full rationale

The paper's claimed derivation chain is not circular in the load-bearing sense. The axial stage is a standard DDPM trained with the noise-matching objective in Eq. (2), and the coronal refinement stage is a residual DDPM whose training context is x_m = Blur(x_0) ⊙ m + x_0 ⊙ (1 - m) (Sec. 2.1); neither defines a prediction as the value of a fitted parameter. The two-perpendicular-views idea is attributed to Lee et al. [23], an external source, and the design is not justified by a self-cited uniqueness theorem. TAM (Sec. 2.3) is trained through the diffusion denoising loss, not directly on FAST/FastSurfer Dice, so the downstream segmentation evaluation, though aligned in objective, is not a reduction of the method's output to its own training target. The only self-citation, Ref. [35], appears in the bibliography and is not invoked as evidence in the body, so it is not load-bearing. The train/test mismatch between the coronal stage's blurred-ground-truth context and the interpolated axial output at inference is a validity concern for the claimed z-detail restoration, but it is a correctness risk rather than circularity. No predicted quantity reduces by construction to an input, so the circularity score is minimal.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim is empirical and rests on several background assumptions: the validity of DDPM conditioning for inpainting, the transfer of perpendicular 2D diffusion to sequential coarse-to-fine refinement, the validity of a synthetic HCP testbed built by transplanting BraTS masks, and the use of segmentation Dice as a proxy for anatomical plausibility. The method's main hand-chosen knobs are GPU-memory-limited hyperparameters (Zmax, batch sizes, noise schedule, training length), none of which are derived from the data. The only new component singled out as an invention is the Tissue-Aware Attention Module, whose benefit is shown only in an in-paper ablation.

free parameters (4)
  • Zmax (maximum volume depth after adaptive resampling) = 24 slices
    Hardware ceiling from GPU memory and batch size; not derived from data (Section 3.2).
  • Diffusion noise schedule endpoints = linear 1e-4 to 2e-2 over T=1000
    Taken from prior work [6]; affects sampling behavior (Section 3.2).
  • Training iterations per stage = 3.25e5
    Arbitrary stopping point; no early stopping or convergence criterion reported (Section 3.2).
  • Learning rate and optimizer settings = Adam, lr=1e-4
    Standard choices; no hyperparameter search reported (Section 3.2).
assumptions (5)
  • standard math The DDPM training objective (Eq. 2) yields valid image-to-image inpainting models when conditioned on (x_t, x_m, m, t).
    Relies on established diffusion theory [17] without new derivation (Section 2.1).
  • domain assumption Perpendicular 2D diffusion models can approximate 3D volumetric coherence, following Lee et al. [23].
    The hierarchical framework assumes this result transfers to sequential coarse-to-fine inpainting (Section 2.3).
  • domain assumption FAST and FastSurfer segmentation Dice within the masked region is a valid proxy for anatomical plausibility of synthetic tissue.
    Used as the primary downstream evaluation without validating segmenter behavior on synthetic input (Section 3.3).
  • domain assumption HCP volumes with transplanted BraTS tumor masks form a valid pseudo-healthy inpainting testbed.
    The authors construct this evaluation set and assume it preserves the properties of real pathological volumes (Section 3.1).
  • domain assumption Cubic interpolation from the resampled axial inpaint to the original slice grid preserves enough structure for coronal refinement to recover lost z-detail.
    Stated in Section 2.3 as the motivation for the second stage; no explicit validation of this recoverability.
invented entities (1)
  • Tissue-Aware Attention Module (TAM)
    purpose: Adds content-adaptive cross-slice attention and gated convolution to the coronal refinement stage to distinguish WM/GM/CSF.
    Only evidence is the in-paper ablation (Table 3); no external validation or theoretical guarantee. This is an architectural component, not a physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Diffusion Framework for Pseudo-Healthy Brain MRI Inpainting with Enhanced 3D Consistency." pith.science (2026). https://pith.science/paper/4M2RNDYZ

@misc{pith2026250717911,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Diffusion Framework for Pseudo-Healthy Brain MRI Inpainting with Enhanced 3D Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4M2RNDYZ}},
  note         = {Machine review of arXiv:2507.17911}
}
read the original abstract

Pseudo-healthy image inpainting is an essential preprocessing step for analyzing pathological brain MRI scans. Most current inpainting methods favor slice-wise 2D models for their high in-plane fidelity, but their independence across slices produces discontinuities in the volume. Fully 3D models alleviate this issue, but their high model capacity demands extensive training data for reliable, high-fidelity synthesis -- often impractical in medical settings. We address these limitations with a hierarchical diffusion framework by replacing direct 3D modeling with two perpendicular coarse-to-fine 2D stages. An axial diffusion model first yields a coarse, globally consistent inpainting; a coronal diffusion model then refines anatomical details. By combining perpendicular spatial views with adaptive resampling, our method balances data efficiency and volumetric consistency. Our experiments show our approach outperforms state-of-the-art baselines in both realism and volumetric consistency, making it a promising solution for pseudo-healthy image inpainting. Code is available at https://github.com/dou0000/3dMRI-Consistent-Inpaint.

Figures

Figures reproduced from arXiv: 2507.17911 by the authors.

Figure 1
Figure 1. Inference in our hierarchical framework for 3D inpainting: at each stage, the model denoises (xt, t) conditioned on the context image and mask for T −1 steps, yielding xˆ0 (the noise sample xt and mask are omitted from the schematic for clarity). 2.1 Preliminary Denoising Diffusion Probabilistic Models (DDPMs) [17] have become powerful tools for image synthesis. They model a forward diffusion process that gradually … view at source ↗
Figure 2
Figure 2. Qualitative Comparison. Qualitative analysis on HCP dataset for generation and segmentation using FAST. The inpainting region is indicated by red boundary. 3.3 Performance Evaluation To evaluate our proposed framework, we compared it against four baseline mod￾els. We selected 3D Autoencoder [16] and Pix2Pix [19] for their widely recognized success in generative tasks, and a memory-efficient 3D DDPM [2] for compariso… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [16]

    E., Salakhutdinov, R

    Hinton, G. E., Salakhutdinov, R. R.: Reducing the dimensionality of data with neural networks. Science 313, 504–507 (2006)

    work page 2006

  2. [19]

    A.: Image-to-Image Translation with Conditional GANs

    Isola, P., Zhu, J.-Y., Zhou, T., Efros, A. A.: Image-to-Image Translation with Conditional GANs. In: CVPR, 5967–5976 (2017)

    work page 2017

  3. [1]

    Ashburner, J.: Computational anatomy with the SPM software. Magn. Reson. Imag- ing 27(8), 1163–1174 (2009)

    work page 2009

  4. [2]

    C.: Memory-efficient 3-D DDPM for medical imaging

    Bieder, F., Wolleb, J., Durrer, A., Sandkuehler, R., Cattin, P. C.: Memory-efficient 3-D DDPM for medical imaging. In: MIDL, 552–567 (2024)

    work page 2024

  5. [3]

    N., Puonti, O., Thielscher, A., Van Leemput, K., Fischl, B., Dalca, A

    Billot, B., Greve, D. N., Puonti, O., Thielscher, A., Van Leemput, K., Fischl, B., Dalca, A. V., Iglesias, J. E.: SynthSeg: MRI segmentation across contrast and res- olution without retraining. Med. Image Anal. 86, 102789 (2023)

    work page 2023

  6. [4]

    In: CVPR, 1251–1258 (2017)

    Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: CVPR, 1251–1258 (2017)

    work page 2017

  7. [5]

    R., Modo, M., Vernon, A

    Crum, W. R., Modo, M., Vernon, A. C., Barker, G. J., Williams, S. C.: Registration of challenging pre-clinical brain images. J. Neurosci. Methods 216(1), 62–77 (2013)

    work page 2013

  8. [6]

    NeurIPS 34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. NeurIPS 34, 8780–8794 (2021)

    work page 2021

Show all 35 references
  1. [7]

    A., Baltruschat, I

    Dohmen, M., Klemens, M. A., Baltruschat, I. M., et al.: Similarity and quality metrics for MR image-to-image translation. Sci. Rep. 15, 3853 (2025)

    2025

  2. [8]

    C., Wolleb, J.: DDPM for healthy-brain inpainting

    Durrer, A., Cattin, P. C., Wolleb, J.: DDPM for healthy-brain inpainting. In: CMDA Challenge, 35–45 (2023)

    2023

  3. [9]

    In: DGDM Workshop, 87–97 (2024)

    Durrer, A., Wolleb, J., Bieder, F., Friedrich, P., Melie-Garcia, L., et al.: Denois- ing diffusion models for 3-D healthy-brain inpainting. In: DGDM Workshop, 87–97 (2024)

    2024

  4. [10]

    S., Glasser, M

    Elam, J. S., Glasser, M. F., Harms, M. P., Sotiropoulos, S. N., Andersson, J. L., et al.: The Human Connectome Project: a retrospective. NeuroImage 244, 118543 (2021) 10 Kwark et al

    2021

  5. [11]

    NeuroImage 62(2), 774–781 (2012)

    Fischl, B.: FreeSurfer. NeuroImage 62(2), 774–781 (2012)

    2012

  6. [12]

    C.: WDM: 3-D wavelet diffusion models for high-resolution image synthesis (2024)

    Friedrich, P., Wolleb, J., Bieder, F., Durrer, A., Cattin, P. C.: WDM: 3-D wavelet diffusion models for high-resolution image synthesis (2024)

    2024

  7. [13]

    Guo, C., Ferreira, D., Fink, K., Westman, E., Granberg, T.: Repeatability of FreeSurfer, FSL-SIENAX and SPM volumetry in multiple sclerosis. Eur. Radiol. 29(3), 1355–1364 (2019)

    2019

  8. [14]

    MELBA 2, 611–656 (2024)

    Hassanaly, R., Brianceau, C., Solal, M., Colliot, O., Burgos, N.: Pseudo-healthy reconstruction for anomaly detection in FDG-PET. MELBA 2, 611–656 (2024)

    2024

  9. [15]

    NeuroImage 219, 117012 (2020)

    Henschel, L., Conjeti, S., Estrada, S., Diers, K., Fischl, B., Reuter, M.: FastSurfer – A deep-learning neuroimaging pipeline. NeuroImage 219, 117012 (2020)

    2020

  10. [17]

    NeurIPS 33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840–6851 (2020)

    2020

  11. [18]

    E., Billot, B., Balbastre, Y., Magdamo, C., Arnold, S

    Iglesias, J. E., Billot, B., Balbastre, Y., Magdamo, C., Arnold, S. E., et al.: SynthSR: Converting clinical scans to high-resolution T1-weighted images. Sci. Adv. 9(5), eadd3607 (2023)

    2023

  12. [20]

    F., Behrens, T

    Jenkinson, M., Beckmann, C. F., Behrens, T. E., Woolrich, M. W., Smith, S. M.: FSL. NeuroImage 62(2), 782–790 (2012)

    2012

  13. [21]

    arXiv:2211.03364 (2022)

    Khader, F., et al.: Medical diffusion – 3-D denoising diffusion probabilistic models. arXiv:2211.03364 (2022)

    2022 arXiv

  14. [22]

    arXiv:2305.08992 (2023)

    Kofler, F., Meissen, F., Steinbauer, F., Graf, R., Oswald, E., et al.: BraTS 2023: Local synthesis of healthy brain tissue via inpainting. arXiv:2305.08992 (2023)

    2023 arXiv

  15. [23]

    C.: Improving 3-D imaging with perpendicular 2-D diffusion models

    Lee, S., Chung, H., Park, M., Park, J., Ryu, W.-S., Ye, J. C.: Improving 3-D imaging with perpendicular 2-D diffusion models. In: ICCV, 10710–10720 (2023)

    2023

  16. [24]

    Liu, X., Xing, F., Yang, C., Kuo, C. C. J., El Fakhri, G., Woo, J.: Symmetric- constrained inpainting for brain-MRI registration. In: BrainLes Workshop, 80–91 (2021)

    2021

  17. [25]

    Liu, X., Xiang, C., Lan, L., Li, C., Xiao, H., Liu, Z.: Lesion-region inpainting for pseudo-healthy image synthesis. Front. Microbiol. 15 (2024)

    2024

  18. [26]

    Imaging Neurosci

    Pollak, C., Kügler, D., Bauer, T., Rüber, T., Reuter, M.: FastSurfer-LIT: Lesion inpainting for whole-brain MRI segmentation. Imaging Neurosci. 3 (2025)

    2025

  19. [27]

    M., Emsell, L., Blommaert, J., Zhylka, A., Kovacs, S., et al.: Vir- tual brain grafting for whole-brain parcellation with large lesions

    Radwan, A. M., Emsell, L., Blommaert, J., Zhylka, A., Kovacs, S., et al.: Vir- tual brain grafting for whole-brain parcellation with large lesions. NeuroImage 229, 117731 (2021)

    2021

  20. [28]

    In: CVPR, 10684–10695 (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR, 10684–10695 (2022)

    2022

  21. [29]

    S., Hahn, A., Lanzenberger, R.: Cortical-thickness estimatesfromFreeSurferandCAT12inAlzheimer’sdisease.J.Neuroimaging28(5), 515–523 (2018)

    Seiger, R., Ganger, S., Kranz, G. S., Hahn, A., Lanzenberger, R.: Cortical-thickness estimatesfromFreeSurferandCAT12inAlzheimer’sdisease.J.Neuroimaging28(5), 515–523 (2018)

    2018

  22. [30]

    IEEE Trans

    Tian, C., Xu, Y., Zuo, W., Zhang, B., Fei, L., Lin, C.-W.: Coarse-to-Fine CNN for image super-resolution. IEEE Trans. Multimedia 23, 1489–1502 (2021)

    2021

  23. [31]

    In: NeurIPS, 5998–6008 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., et al.: Attention is all you need. In: NeurIPS, 5998–6008 (2017)

    2017

  24. [32]

    S.: Free-form image inpainting with gated convolution

    Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T. S.: Free-form image inpainting with gated convolution. In: ICCV, 4471–4480 (2019)

    2019

  25. [33]

    M.: MR image segmentation using HMRF-EM

    Zhang, Y., Brady, M., Smith, S. M.: MR image segmentation using HMRF-EM. IEEE Trans. Med. Imaging 20(1), 45–57 (2001) Title Suppressed Due to Excessive Length 11

    2001

  26. [34]

    In: MICCAI, 592–601 (2023)

    Zhu, L., Xue, Z., Jin, Z., Liu, X., He, J., Liu, Z., Yu, L.: Make-a-Volume: Latent diffusion for 3-D MRI synthesis. In: MICCAI, 592–601 (2023)

    2023

  27. [35]

    H., Zhu, R., Hong, K., Tao, Y., et al.: Diff-Ensembler: Ensem- bling 2-D diffusion models for volumetric translation

    Zhu, X., Kwark, D. H., Zhu, R., Hong, K., Tao, Y., et al.: Diff-Ensembler: Ensem- bling 2-D diffusion models for volumetric translation. arXiv:2501.07430 (2025)

    2025 arXiv

Pith tools

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