Temporal Differential Fields for 4D Motion Modeling via Image-to-Video Synthesis

Hanxiao Zhang; Jie Yang; Minghui Zhang; Nassir Navab; Xin You

REVIEW 4 major objections 5 minor 1 cited by

Temporal Differential Fields for 4D Motion Modeling via Image-to-Video Synthesis

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 Mo-Diff generates 4D cardiac and lung motion videos from a single first frame, using learned temporal differential fields to keep adjacent frames consistent.

desk verdict Solid incremental I2V for single-frame 4D motion, but the warping mechanism is underspecified and the evaluation needs more statistical rigor. read the letter →

arxiv 2505.17333 v2 pith:IA7DNFSN submitted 2025-05-22 cs.CV

Xin You , Minghui Zhang , Hanxiao Zhang , Jie Yang , Nassir Navab This is my paper

classification cs.CV

keywords image-to-videosynthesistemporaldifferentialfields4DmedicalimagingdiffusionmodelsmotionmodelingrespiratorycardiacMRIconsistency

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

Mo-Diff is a two-stage diffusion pipeline that generates a 4D motion video of regular respiration-driven anatomy from a single first frame plus a frame count, instead of requiring both start and end frames from two high-dose scans. The intended payoff is clinical: cardiac MRI and lung CT motion sequences could be simulated for image-guided planning without the extra radiation exposure of a second scan. The paper argues that its temporal differential fields, generated by the Temporal Differential Diffusion Model, keep adjacent frames consistent enough that the synthesized sequences rival two-frame baselines on perceptual and temporal metrics. A sympathetic reading of the results is that the single-frame premise is viable for regular cardiac and pulmonary motion.

What carries the argument

The load-bearing object is the temporal differential field, defined as the difference between adjacent frames, F_i = I_i ⊖ I_{i-1}, with the first field set to zero. The Temporal Differential Diffusion Model (TDDM) generates these fields conditioned on the first frame and N; the prompt attention layer fuses first-frame features into the field generator at feature level rather than by naive concatenation, and the field augmented layer warps the generated fields onto the first-frame features and interleaves them with future-frame features before self-attention, so the video generator receives an explicit motion prior. The latent video diffusion backbone keeps the pipeline GPU-feasible, and N encodes the respiratory period so the model knows the speed of motion.

What would settle it

Run Mo-Diff on a 4D dataset that includes patients with arrhythmia or irregular breathing patterns, where the true inter-frame motion is not periodic. If the synthesized sequences degrade to near-static or drifting frames while a model conditioned on per-frame ECG phase tracks the ground truth, the periodic single-frame premise is falsified.

Watch

Extended reading notes

Core claim

The central claim is that regular breathing-induced motion can be modeled as an image-to-video generation problem: conditioning a latent video diffusion model on the first volume and the frame number N is sufficient to synthesize the whole 4D sequence along the intrinsic motion trajectory. To keep the animation temporally consistent, the paper introduces temporal differential fields F_i = I_i ⊖ I_{i-1} (with F_1 = 0) that encode inter-frame change, generated by a dedicated diffusion model and injected into the video generator through a prompt attention layer and a field augmented layer. The frame count N is read as the breathing period, obtained from electrocardiogram signals, and acts as an explicit velocity constraint. On the ACDC cardiac and 4D Lung datasets the method reports perceptual and temporal-consistency scores (LPIPS, FVD) competitive with or better than models that use both start and end frames.

Load-bearing premise

The load-bearing premise is that respiratory and cardiac motion is regular and periodic enough that the first frame plus an ECG-derived frame count N fully determines the future sequence; the paper itself concedes that highly unstable breathing would require additional clinical guidance.

Editorial extensions

If this is right

If the central claim holds, a clinician collecting only the first frame of a respiratory cycle can obtain the full 4D motion sequence, eliminating the need for a second high-dose CT or MRI scan for motion modeling.
Frame number N derived from ECG becomes a usable control for motion velocity: changing N changes the simulated breathing period, which could support patient-specific simulation without retraining.
The temporal differential field acts as an explicit inter-frame motion prior, which is a concrete mechanism other generative medical video models could adopt to reduce flicker and drift.
On perceptual and temporal metrics (LPIPS, FVD), single-frame Mo-Diff is claimed to rival two-frame interpolation and registration baselines, meaning the practical gap between single-frame and two-frame protocols may be narrower than previously assumed.

Reading between the lines

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

A natural extension is to condition TDDM on continuous ECG signals or respiratory phase instead of a single frame count, which would address the paper's own concession that highly unstable breathing is not yet handled.
The differential-field formulation could be transferred to other periodic anatomical motion (e.g., diaphragm, blood vessels, fetal movement) where inter-frame differences are sparse and regular.
Because the differential fields are explicit and interpretable, they could serve as an anomaly detector: a patient whose predicted fields deviate strongly from the periodic prior may indicate irregular motion or pathology.
A testable extension is to feed the predicted temporal differential fields into a flow-based interpolator as a warm start, potentially combining the generation quality of diffusion with the geometric guarantees of registration.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is a reasonable incremental contribution: it applies image-to-video synthesis to single-frame 4D cardiac and lung motion generation, with a temporal differential diffusion model and two interaction layers. The clinical motivation is clear, and the ablations show each component helps. The LPIPS and FVD improvements over two-frame baselines on ACDC and 4D Lung are credible. But the evaluation has gaps, and the ``warping'' mechanism is ambiguous enough that I'd want clarifications before trusting the central claim.

What is new: to my knowledge, this is the first attempt to simulate regular 4D medical motion from a single prompting frame plus a frame count, avoiding a second high-dose scan. The TDDM generates adjacent-frame intensity differences, and the PAL and FAL integrate these into the I2V model. The perceptual and temporal-consistency gains over methods like SVIN and UVI-Net are meaningful for clinical practice. The ablations are honest and show each module contributes.

Soft spots: FVD is reported without error bars or significance tests; with only two datasets, that's a real omission. There's no baseline that removes TDDM entirely and just runs the base LVDM, so we can't isolate what the fields add. The paper doesn't state whether Stage 2 training uses ground-truth or generated fields. More concerning, Eqs. (3) and (4) are conceptually muddy: the fields are intensity differences, yet they are used in a ``warping transform'' without specifying W. If W is a learned warp, treating intensity difference as displacement is physically unjustified; if W is just feature modulation, calling it ``warping'' is misleading. The stress-test's worry about unvalidated intermediate representations is legitimate, but not fatal here, because the final output quality is what matters and the model could learn useful cues from intensity differences even if they are not true deformation fields. Still, the paper needs to clarify W and ideally add a loss that checks whether the fields correspond to plausible motion. Finally, the regularity assumption with N derived from ECG is explicitly acknowledged, which is honest, but it limits the clinical claim to regular breathing.

Who this is for: researchers in medical image synthesis and I2V. The paper deserves a serious referee; the core idea is sound and the experiments are mostly there, but the missing statistical checks and the ambiguity around W should be addressed in revision. I'd send it to peer review and ask for those additions.

Referee Report

4 major / 5 minor

Summary. This paper proposes Mo-Diff, a two-stage image-to-video (I2V) diffusion framework for 4D medical motion modeling from a single prompting frame. Stage 1, the Temporal Differential Diffusion Model (TDDM), generates temporal differential fields from the first frame and a frame count N; Stage 2 uses these fields, via a prompt attention layer and a field augmented layer, to guide an underlying Latent Video Diffusion Model to synthesize the remaining frames. The method is evaluated on the ACDC cardiac MRI and 4D-Lung CT datasets, with baseline comparisons to two-frame interpolation methods and one single-frame diffusion baseline. Results show lower PSNR than two-frame methods but better LPIPS and FVD on both datasets, and an ablation study reports the contribution of the frame-number condition, PAL, and FAL. The paper claims that Mo-Diff simulates regular cardiac and respiratory motion along the intrinsic motion trajectory using only a first frame and a known frame count.

Significance. The application-oriented contribution is potentially significant: if validated, a clinician could generate a complete regular motion cycle from a single acquired frame plus a known period, reducing the need for a second high-dose scan. The paper is grounded in public datasets and compares against published baselines, which is a strength, and the ablation study isolates the proposed components. The central idea of explicitly generating temporal differential fields as a conditioning signal is reasonable and worth pursuing. However, the manuscript does not yet substantiate that these fields represent true anatomical motion, and the empirical evidence lacks error bars for FVD, lacks significance tests, and omits a named baseline comparison to the underlying LVDM architecture. The code link is provided, which aids reproducibility, but the missing implementation details currently limit verification.

major comments (4)

[§2.3, Eq. (4)] The warping transform W in Eq. (4) is never defined. Eq. (3) defines the temporal differential field F_i as an intensity subtraction I_i ⊖ I_{i−1}, not as a displacement or deformation field; 'warping' frame features with an intensity-difference field is not a standard operation unless W is precisely specified. Because Stage 2 applies W to every future frame, the central mechanism is not reproducible as written, and the paper's claim that the fields track the 'intrinsic motion trajectory' is unsupported without a definition of W and a validation that the generated fields correspond to anatomically meaningful deformations, for example by comparison against registration-derived displacement fields or by enforcing a cycle-consistency constraint.
[§2.3 and §3.1 (Implementation Details)] The two-stage training protocol is not described. It is not stated whether Stage 1 (TDDM) and Stage 2 (I2V) are trained sequentially or jointly, nor whether Stage 2 is trained with ground-truth differential fields computed from Eq. (3) or with fields generated by TDDM. This distinction is load-bearing: if Stage 2 uses ground-truth fields at training time but generated fields at test time, the test-time input distribution shifts, and the reported performance may not reflect the actual deployment condition. Please specify the exact training procedure, the data flow between the two stages, and the loss terms used in each stage.
[§3.2, Tables 1 and 2] FVD values are reported as point estimates without confidence intervals in both Table 1 and Table 2, and Table 2 has no error bars and no significance tests for any metric. The claimed FVD improvement over UVI-Net is a single number and cannot be assessed for statistical robustness. Please provide error bars from multiple training runs or bootstrapping, and report significance tests for the main comparisons and for the ablation rows, especially because several adjacent PSNR and LPIPS values in Table 2 are close.
[§3.2, Table 1] The only single-frame baseline in Table 1 is Condi-Diffusion; the base LVDM architecture on which Mo-Diff is built is not included as a named baseline. The closest row is the ablation row with Frame Number ✔, PAL ✗, FAL ✗ in Table 2, but it is not labeled as the base model and carries no error bars. Because the paper's central claim is that TDDM fields improve temporal consistency over a plain I2V model, this comparison should be explicit and statistically characterized.

minor comments (5)

[§3.2, text near Table 1] The sentence stating that Mo-Diff outperforms UVI-Net with a 5.9 FVD improvement matches the 4D-Lung column (121.7 vs 115.8) but not the ACDC column (94.2 vs 86.1); please specify which dataset is being referenced.
[§2.2] The statement that frame number N 'corresponds to the specific breathing period, acquired by electrocardiogram signals' conflates cardiac and respiratory gating; for the 4D-Lung dataset the respiratory period is not normally obtained from ECG. Please clarify the gating source for each modality.
[§2.3, Fig. 2] The operations labeled 'Field Warping' and 'Interleaving' in Figure 2 are not accompanied by a precise tensor-level description; please provide the exact dimension transformations or pseudocode so that the architecture is reproducible.
[§3.1, metrics] FVD is described as measuring temporal consistency and continuity, but FVD is a distributional metric over video features and does not directly isolate temporal consistency; please moderate this wording or add a dedicated temporal-consistency metric.
[General] There are several typographical issues, including 'T able 1' in the manuscript and 'samping' in Section 2.2; please proofread the final version.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: temporal differential fields are learned predictions from first frame and frame count, and final videos are evaluated against external ground-truth data and published baselines rather than being defined from the conditioning inputs.

full rationale

The derivation chain is not circular by construction. In Stage 1, TDDM predicts differential fields from Gaussian noise conditioned only on the prompting frame I1 and frame number N ('F̂ = TDDM(G1, t; I1, N)', Section 2.2); the training target fields are computed from ground-truth videos via Eq. 3, and at test time no target frame is provided. In Stage 2, the I2V network synthesizes latent embeddings guided by these fields under the standard denoising objective (Eq. 2); the unspecified warping transform W in Eq. 4 is a design component whose contribution is established by the ablation in Table 2 against external metrics, not assumed by definition. The reported PSNR/LPIPS/FVD values are computed against ground-truth ACDC and 4D-Lung videos and compared with published baselines (SVIN, Voxelmorph, UVI-Net, LDMVFI, DDM, LDDM, Condi-Diffusion), so the central claim is externally falsifiable. The only self-citations appearing in the main text are Ref. [30] (cardiac MRI context) and Ref. [31] (loosely cited to support the radiation-exposure statement); neither is load-bearing, and [31] being a vertebrae-segmentation paper used for a radiation-safety claim is a citation-practice weakness, not a circular reduction. The skeptic's concern — that intensity-difference fields (Eq. 3) are not validated as true deformations and W is unspecified — is a validity/correctness critique about whether the intermediate representation tracks real motion, not an instance of a prediction being equal to its inputs by construction. Likewise, the requirement of ECG-derived N and the admitted limitation on unstable breathing (Conclusion) concern underdetermination of conditioning, not circularity. Accordingly, no fitted constant is renamed as a prediction and no equation equates output to input; the circularity burden is low.

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

The central claim rests on the regularity assumption of respiratory motion, on the sufficiency of adjacent-frame difference fields as conditioning, and on the clinical premise that dynamic backgrounds resist registration. No ad hoc fitted constants are introduced; learned weights and hyperparameters are standard practice.

assumptions (3)

domain assumption Respiration-induced motion is regular and periodic, so a first frame plus frame count N determines the motion sequence.
Invoked in Sections 1 and 2.2 to justify single-frame forecasting; the conclusion admits the method needs ECG guidance for unstable breathing.
ad hoc to paper Adjacent-frame subtraction fields (Eq. 3) capture the motion information needed for temporally consistent synthesis.
The paper provides no analysis that these fields are sufficient or that errors do not accumulate over N frames.
domain assumption Dynamic background deviations between first and last frames cannot be removed by registration, motivating the single-frame setup.
Section 1 states this as the clinical motivation; it is not quantitatively demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Temporal Differential Fields for 4D Motion Modeling via Image-to-Video Synthesis." pith.science (2026). https://pith.science/paper/IA7DNFSN

@misc{pith2026250517333,
  author       = {Pith},
  title        = {Pith review of: Temporal Differential Fields for 4D Motion Modeling via Image-to-Video Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IA7DNFSN}},
  note         = {Machine review of arXiv:2505.17333}
}

read the original abstract

Temporal modeling on regular respiration-induced motions is crucial to image-guided clinical applications. Existing methods cannot simulate temporal motions unless high-dose imaging scans including starting and ending frames exist simultaneously. However, in the preoperative data acquisition stage, the slight movement of patients may result in dynamic backgrounds between the first and last frames in a respiratory period. This additional deviation can hardly be removed by image registration, thus affecting the temporal modeling. To address that limitation, we pioneeringly simulate the regular motion process via the image-to-video (I2V) synthesis framework, which animates with the first frame to forecast future frames of a given length. Besides, to promote the temporal consistency of animated videos, we devise the Temporal Differential Diffusion Model to generate temporal differential fields, which measure the relative differential representations between adjacent frames. The prompt attention layer is devised for fine-grained differential fields, and the field augmented layer is adopted to better interact these fields with the I2V framework, promoting more accurate temporal variation of synthesized videos. Extensive results on ACDC cardiac and 4D Lung datasets reveal that our approach simulates 4D videos along the intrinsic motion trajectory, rivaling other competitive methods on perceptual similarity and temporal consistency. Codes will be available soon.

Figures

Figures reproduced from arXiv: 2505.17333 by the authors.

**Figure 1.** The whole Mo-Diff pipeline. Stage 1: TDDM yields temporal differential fields conditioning on the prompting frame and frame number. Stage 2: the conditional diffusion model will synthesize 4D latent embeddings with the guidance of temporal fields. These latent embeddings are then transformed into 4D videos with regular motions. consistency, showcasing potentials of simulating regular cardiac and pulmonary motions. … view at source ↗

**Figure 2.** The detailed structure of prompt attention layer and field augmented layer. backgrounds, and highlight the foreground regions in the respiratory process. Thus, Fˆ can enormously boost the reconstruction performance of future frames by the I2V network. Besides, temporal fields representing regular motions will significantly improve the temporal consistency of synthesized videos. It is worth mentioning that frame numb… view at source ↗

**Figure 3.** [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Temporal error maps between different models. All frame predictions are visualized except starting and ending frames (Red box: intermediate frames) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

FB-Diff: Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging
eess.IV 2025-07 conditional novelty 5.0 of 10

FB-Diff uses learned Fourier motion bases to guide a diffusion model, improving perceptual quality and temporal consistency in 4D medical frame interpolation.

Reference graph

Works this paper leans on

35 extracted references · 33 canonical work pages · cited by 1 Pith paper

[1]

Spatial-temporal analysis of cause-specific car- diovascular hospital admission in Beijing, China

Endawoke Amsalu et al. “Spatial-temporal analysis of cause-specific car- diovascular hospital admission in Beijing, China”. In:International Journal of Environmental Health Research31.6 (2021), pp. 595–606

work page 2021
[2]

ConditionalDiffusionModel for Versatile Temporal Inpainting in 4D Cerebral CT Perfusion Imaging

JuyoungBae,ElizabethTong,andHaoChen.“ConditionalDiffusionModel for Versatile Temporal Inpainting in 4D Cerebral CT Perfusion Imaging”. In: MICCAI. Springer. 2024, pp. 67–77

work page 2024
[3]

Voxelmorph:alearningframeworkfordeformable medical image registration

GuhaBalakrishnanetal.“Voxelmorph:alearningframeworkfordeformable medical image registration”. In:IEEE TMI 38.8 (2019), pp. 1788–1800

work page 2019
[4]

DeeplearningtechniquesforautomaticMRIcardiac multi-structures segmentation and diagnosis: is the problem solved?

OlivierBernardetal.“DeeplearningtechniquesforautomaticMRIcardiac multi-structures segmentation and diagnosis: is the problem solved?” In: IEEE transactions on medical imaging37.11 (2018), pp. 2514–2525

work page 2018
[5]

Align your latents: High-resolution video syn- thesis with latent diffusion models

Andreas Blattmann et al. “Align your latents: High-resolution video syn- thesis with latent diffusion models”. In:CVPR. 2023, pp. 22563–22575

work page 2023
[6]

Ultrasound Image-to-Video Synthesis via Latent Dy- namic Diffusion Models

Tingxiu Chen et al. “Ultrasound Image-to-Video Synthesis via Latent Dy- namic Diffusion Models”. In:MICCAI. Springer. 2024, pp. 764–774

work page 2024
[7]

Ldmvfi: Video frame inter- polation with latent diffusion models

Duolikun Danier, Fan Zhang, and David Bull. “Ldmvfi: Video frame inter- polation with latent diffusion models”. In:Proceedings of the AAAI Con- ference on Artificial Intelligence. Vol. 38. 2. 2024, pp. 1472–1480

work page 2024
[8]

Jan Ehrhardt, Cristian Lorenz, et al.4D modeling and estimation of res- piratory motion for radiation therapy. Vol. 10. Springer, 2013

work page 2013

Show all 35 references

[9]

A spatiotemporal volumetric interpolation network for 4d dynamic medical image

Yuyu Guo et al. “A spatiotemporal volumetric interpolation network for 4d dynamic medical image”. In:CVPR. 2020, pp. 4726–4735

2020
[10]

Denoising diffusion proba- bilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffusion proba- bilistic models”. In:NeurIPS 33 (2020), pp. 6840–6851

2020
[11]

Video diffusion models

Jonathan Ho et al. “Video diffusion models”. In: Advances in Neural In- formation Processing Systems35 (2022), pp. 8633–8646. 10 Xin You et al

2022
[12]

Image quality metrics: PSNR vs. SSIM

Alain Hore and Djemel Ziou. “Image quality metrics: PSNR vs. SSIM”. In: 2010 20th international conference on pattern recognition. IEEE. 2010, pp. 2366–2369

2010
[13]

Data from 4D lung imaging of NSCLC patients

Geoffrey D Hugo et al. “Data from 4D lung imaging of NSCLC patients”. In: (No Title)(2016)

2016
[14]

Videointerpolationwithdiffusionmodels

SiddhantJainetal.“Videointerpolationwithdiffusionmodels”.In: CVPR. 2024, pp. 7341–7351

2024
[15]

Myocardial tagging with MR imaging: overview of normal and pathologic findings

Mi-Young Jeung et al. “Myocardial tagging with MR imaging: overview of normal and pathologic findings”. In:Radiographics 32.5 (2012), pp. 1381– 1398

2012
[16]

Diffusion deformable model for 4D temporal medical image generation

Boah Kim and Jong Chul Ye. “Diffusion deformable model for 4D temporal medical image generation”. In:MICCAI. Springer. 2022, pp. 539–548

2022
[17]

Data-Efficient Unsupervised Interpolation Without AnyIntermediateFramefor4DMedicalImages

JungEun Kim et al. “Data-Efficient Unsupervised Interpolation Without AnyIntermediateFramefor4DMedicalImages”.In: CVPR.2024,pp.11353– 11364

2024
[18]

A robotic system for transthoracic puncture of pul- monary nodules based on gated respiratory compensation

Dongyuan Li et al. “A robotic system for transthoracic puncture of pul- monary nodules based on gated respiratory compensation”. In:Computer Methods and Programs in Biomedicine244 (2024), p. 107995

2024
[19]

Sora: A review on background, technology, limitations, andopportunitiesoflargevisionmodels

Yixin Liu et al. “Sora: A review on background, technology, limitations, andopportunitiesoflargevisionmodels”.In: arXiv preprint arXiv:2402.17177 (2024)

2024 arXiv
[20]

Generating realistic brain mris via a conditional diffusion probabilistic model

Wei Peng et al. “Generating realistic brain mris via a conditional diffusion probabilistic model”. In:MICCAI. Springer. 2023, pp. 14–24

2023
[21]

Computed tomography and patient risk: Facts, perceptions and uncertainties

Stephen P Power et al. “Computed tomography and patient risk: Facts, perceptions and uncertainties”. In:World journal of radiology8.12 (2016), p. 902

2016
[22]

Learning spatio-temporal representation with pseudo- 3d residual networks

Zhaofan Qiu et al. “Learning spatio-temporal representation with pseudo- 3d residual networks”. In:ICCV. 2017, pp. 5533–5541

2017
[23]

High-resolution image synthesis with latent diffu- sion models

Robin Rombach et al. “High-resolution image synthesis with latent diffu- sion models”. In:CVPR. 2022, pp. 10684–10695

2022
[24]

Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling

Xiaoyu Shi et al. “Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling”. In:SIGGRAPH. 2024, pp. 1– 11

2024
[25]

Towards accurate generative models of video: A new metric & challenges

Thomas Unterthiner et al. “Towards accurate generative models of video: A new metric & challenges”. In:arXiv preprint arXiv:1812.01717(2018)

2018 arXiv
[26]

Dosimetric comparison of stereotactic body radiotherapy using 4D CT and multiphase CT images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage

Lu Wang et al. “Dosimetric comparison of stereotactic body radiotherapy using 4D CT and multiphase CT images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage”. In:Radiotherapy and Oncology91.3 (2009), pp. 314–324

2009
[27]

Mpvf: 4d medical image inpainting by multi-pyramid voxelflows

Tzu-Ti Wei et al. “Mpvf: 4d medical image inpainting by multi-pyramid voxelflows”.In: IEEE Journal of Biomedical and Health Informatics(2023)

2023
[28]

Cross-conditioned diffusion model for medical image to image translation

Zhaohu Xing et al. “Cross-conditioned diffusion model for medical image to image translation”. In:MICCAI. Springer. 2024, pp. 201–211

2024
[29]

Learning with explicit shape priors for medical image segmentation

Xin You et al. “Learning with explicit shape priors for medical image segmentation”. In:IEEE Transactions on Medical Imaging(2024). Mo-Diff 11

2024
[30]

Semantic difference guidance for the uncertain boundary segmentation of CT left atrial appendage

Xin You et al. “Semantic difference guidance for the uncertain boundary segmentation of CT left atrial appendage”. In:MICCAI. Springer. 2023, pp. 121–131

2023
[31]

SLoRD: Structural Low-Rank Descriptors for Shape Con- sistency in Vertebrae Segmentation

Xin You et al. “SLoRD: Structural Low-Rank Descriptors for Shape Con- sistency in Vertebrae Segmentation”. In:arXiv preprint arXiv:2407.08555 (2024)

2024 arXiv
[32]

4D myocardium reconstruction with decoupled mo- tion and shape model

Xiaohan Yuan et al. “4D myocardium reconstruction with decoupled mo- tion and shape model”. In:ICCV. 2023, pp. 21252–21262

2023
[33]

Pass: test-time prompting to adapt styles and se- mantic shapes in medical image segmentation

Chuyan Zhang et al. “Pass: test-time prompting to adapt styles and se- mantic shapes in medical image segmentation”. In:IEEE TMI (2024)

2024
[34]

Adding conditional con- trol to text-to-image diffusion models

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. “Adding conditional con- trol to text-to-image diffusion models”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 3836–3847

2023
[35]

The unreasonable effectiveness of deep features as a perceptual metric

Richard Zhang et al. “The unreasonable effectiveness of deep features as a perceptual metric”. In:CVPR. 2018, pp. 586–595

2018

Pith tools

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

[1] [1]

Spatial-temporal analysis of cause-specific car- diovascular hospital admission in Beijing, China

Endawoke Amsalu et al. “Spatial-temporal analysis of cause-specific car- diovascular hospital admission in Beijing, China”. In:International Journal of Environmental Health Research31.6 (2021), pp. 595–606

work page 2021

[2] [2]

ConditionalDiffusionModel for Versatile Temporal Inpainting in 4D Cerebral CT Perfusion Imaging

JuyoungBae,ElizabethTong,andHaoChen.“ConditionalDiffusionModel for Versatile Temporal Inpainting in 4D Cerebral CT Perfusion Imaging”. In: MICCAI. Springer. 2024, pp. 67–77

work page 2024

[3] [3]

Voxelmorph:alearningframeworkfordeformable medical image registration

GuhaBalakrishnanetal.“Voxelmorph:alearningframeworkfordeformable medical image registration”. In:IEEE TMI 38.8 (2019), pp. 1788–1800

work page 2019

[4] [4]

DeeplearningtechniquesforautomaticMRIcardiac multi-structures segmentation and diagnosis: is the problem solved?

OlivierBernardetal.“DeeplearningtechniquesforautomaticMRIcardiac multi-structures segmentation and diagnosis: is the problem solved?” In: IEEE transactions on medical imaging37.11 (2018), pp. 2514–2525

work page 2018

[5] [5]

Align your latents: High-resolution video syn- thesis with latent diffusion models

Andreas Blattmann et al. “Align your latents: High-resolution video syn- thesis with latent diffusion models”. In:CVPR. 2023, pp. 22563–22575

work page 2023

[6] [6]

Ultrasound Image-to-Video Synthesis via Latent Dy- namic Diffusion Models

Tingxiu Chen et al. “Ultrasound Image-to-Video Synthesis via Latent Dy- namic Diffusion Models”. In:MICCAI. Springer. 2024, pp. 764–774

work page 2024

[7] [7]

Ldmvfi: Video frame inter- polation with latent diffusion models

Duolikun Danier, Fan Zhang, and David Bull. “Ldmvfi: Video frame inter- polation with latent diffusion models”. In:Proceedings of the AAAI Con- ference on Artificial Intelligence. Vol. 38. 2. 2024, pp. 1472–1480

work page 2024

[8] [8]

Jan Ehrhardt, Cristian Lorenz, et al.4D modeling and estimation of res- piratory motion for radiation therapy. Vol. 10. Springer, 2013

work page 2013