Test Time Training for 4D Medical Image Interpolation

Qikang Zhang; Yingjie Lei; Zhonghao Xie; Zihao Zheng; Ziyang Chen

REVIEW 4 major objections 6 minor 22 references

Test Time Training for 4D Medical Image Interpolation

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Test-time training lifts 4D medical interpolation accuracy

desk verdict A plausible first application of TTT to 4D medical interpolation, but the reported gains are within noise and no distribution shift is tested, so the empirical claims do not hold. read the letter →

arxiv 2502.02341 v1 pith:KPGQC3JE submitted 2025-02-04 eess.IV cs.AIcs.CV

Qikang Zhang , Yingjie Lei , Zihao Zheng , Ziyang Chen , Zhonghao Xie This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords test-timetraining4Dmedicalimageinterpolationself-supervisedlearningdomainadaptationrotationpredictionmaskedautoencoderanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes TTT4MII, a test-time training framework for 4D medical image interpolation that adapts a pre-trained model to each test instance without needing ground-truth labels. The key idea is to update a shared feature extractor using a self-supervised auxiliary task—rotation prediction or 3D masked autoencoding—before the main interpolation head makes its prediction. The authors demonstrate on the Cardiac and 4D-Lung datasets that this adaptation yields higher PSNR (33.73 dB and 34.02 dB) than prior interpolation methods, including the UVI-Net baseline they build on. The paper also compares three test-time adaptation schemes (naïve, online, and mini-batch) and finds they perform nearly identically. If the approach holds, it offers a label-free route to adapting interpolation models to new scanners or protocols at deployment time.

What carries the argument

The core mechanism is a shared feature extractor $f$ (a 3D AlexNet) that feeds both the main interpolation head $h$ (UVI-Net) and a self-supervised auxiliary head $g$. At test time, only the auxiliary loss $L_g$ is optimized—either a four-class rotation-prediction cross-entropy loss or a masked-autoencoder reconstruction loss—to update $f$'s parameters, leaving $h$ frozen. After adaptation, $h \circ f$ predicts the intermediate frame. The paper implements three TTT schemes: Naïve (adapt on all test batches before predicting), Online (reset to $\theta_0$ and adapt per batch), and Mini-batch (sequentially update across batches), formalized by different update rules for $\theta$.

What would settle it

Run TTT4MII and its UVI-Net baseline on a genuinely out-of-distribution test set, such as cardiac MRI after training on CT or scans from a different center, and compare per-scan PSNR; if the adapted model does not consistently beat the frozen baseline, the distribution-shift claim collapses. Alternatively, a paired significance test across the existing test scans would show whether the 0.14 dB and 0.02 dB PSNR gains are statistically meaningful or within noise.

Watch

Extended reading notes

Core claim

The paper claims that self-supervised test-time training improves 4D medical image interpolation accuracy on unseen test data. Concretely, TTT4MII freezes the interpolation head, updates only a shared feature extractor via an auxiliary self-supervised loss on the test input, then performs interpolation. On the Cardiac dataset it reports a PSNR of 33.73 dB, NCC of 0.571, and NMSE of 2.230, exceeding the UVI-Net baseline (33.59 dB, 0.565, 2.384). On 4D-Lung it reports PSNR 34.02 dB, SSIM 0.981, NCC 0.320, and NMSE 0.551, again marginally above UVI-Net (34.00 dB, 0.980, 0.320, 0.552). The paper further claims that the 3D masked autoencoder auxiliary task outperforms rotation prediction, and that the three TTT schemes deliver comparable accuracy.

Load-bearing premise

The method assumes that self-supervised adaptation on the unlabeled test scan improves interpolation accuracy, and that the standard test splits used here genuinely represent a distribution shift.

Editorial extensions

If this is right

If the central claim holds, interpolation models can be updated at inference time on unlabeled patient scans, potentially reducing quality drops when a deployed model meets a new scanner or imaging protocol.
Because the framework is architecture-agnostic, it could be retrofitted to any feature-extractor-plus-head interpolation model and transferred to other image-to-image tasks such as segmentation and registration.
Since the three TTT schemes give nearly identical accuracy, a deployment can choose the scheme by latency, with Online TTT allowing per-batch adaptation without waiting for the whole test set.
The better accuracy of 3D-MAE over rotation prediction suggests that reconstruction-based self-supervision may be a stronger general-purpose auxiliary task for medical image interpolation.

Reading between the lines

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

The reported gains over UVI-Net are small (0.14 dB on Cardiac, 0.02 dB on 4D-Lung) and the paper does not report per-run variance for baselines, so the practical significance of the improvement is not yet established; a paired, per-scan comparison with confidence intervals would be needed.
The paper frames its motivation as handling distribution shifts, but the experiments use the standard test splits of Cardiac and 4D-Lung, leaving the actual shift scenario untested; a cross-scanner or cross-center evaluation would directly test the claim.
A risk is that adapting the shared encoder with a self-supervised loss could harm task-specific features; the paper's design of freezing the interpolation head mitigates but does not eliminate this, and the narrow metric gaps suggest sensitivity to initialization.

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

Here's my read. This paper is the first to apply test-time training (TTT) to 4D medical image interpolation, using rotation prediction and a 3D masked autoencoder as auxiliary tasks. That is a legitimate extension of the TTT literature, and the paper credits that literature properly. The ablation covering three adaptation schemes (Naive, Online, Mini-batch) and two self-supervised tasks is a reasonable piece of work, and the clinical motivation—adapting to unseen scanners without labels—is real.

The problem is that the empirical core does not hold up. On Cardiac, the PSNR gain over UVI-Net is 0.14 dB (33.73 vs 33.59), while the paper's own standard deviation is ±0.247. On 4D-Lung, the gain is 0.02 dB (34.02 vs 34.00, std ±0.341). Both deltas are inside the noise. No variance is reported for any baseline, so significance cannot be assessed. More telling, the three TTT schemes give nearly identical numbers on every metric—Cardiac PSNR is 33.70 for all three rotation rows and 33.73/33.72/33.73 for MAE. The paper itself calls the differences 'surprisingly subtle,' but that is the correct conclusion: the TTT step has no measurable effect here.

The second weakness is conceptual. The experiments never introduce a distribution shift. Cardiac and 4D-Lung use their standard train/test splits from the same acquisition protocol, so the setup tests interpolation performance, not adaptation to a new distribution. Even if the deltas were significant, they would not support the paper's stated motivation.

What is good: the method is simple and plausible, the writing is clear, and the citation pattern is honest—the components are attributed to Sun, Gandelsman, and He. The abstract promises code but no link is provided, which is a minor issue.

I would send this to a reviewer rather than desk-reject it, because the task is meaningful and the idea could work under a properly designed shift experiment. But the current empirical claims are unsupported. The referee should require a real distribution-shift test, variance for all methods, and an explanation of why the three schemes behave identically. Without those, this is a workshop-level application note. For a reader working on TTT or medical interpolation, it is a useful cautionary example, but not a citable result.

Referee Report

4 major / 6 minor

Summary. The paper proposes TTT4MII, a test-time training framework for 4D medical image interpolation. It shares a feature extractor between the interpolation head and a self-supervised auxiliary head, and at test time adapts the feature extractor using either rotation prediction or a 3D masked autoencoder objective. Three adaptation schemes are introduced: Naive, Online, and Mini-batch TTT. Experiments on the Cardiac and 4D-Lung datasets compare against prior interpolation methods and report PSNR, NCC, SSIM, and NMSE. The paper claims significant improvements over the strongest baseline UVI-Net.

Significance. If the claimed gains were real, the paper would make a useful contribution by adapting interpolation models to test distributions without labels, a plausible need across scanners and protocols. It would also be one of the first TTT applications to 4D medical image interpolation and provides a clear comparison of three TTT schemes. However, the empirical support is too weak: the improvements over the strongest baseline are within the reported noise, no distribution shift is evaluated, and the three schemes give nearly identical results, casting doubt on whether the adaptation step has any effect.

major comments (4)

[Section V-A, Tables I and II] The central claim of significant improvement over the strongest baseline is not supported by the reported numbers. On Cardiac, the proposed method achieves 33.73 dB PSNR versus UVI-Net's 33.59 dB (Delta = 0.14 dB), while Table IV reports a standard deviation of +/-0.247 dB for the proposed method on the same dataset. On 4D-Lung, the gain is 34.02 dB versus 34.00 dB (Delta = 0.02 dB), with +/-0.341 dB reported. No standard deviation is reported for any baseline, so a significance test cannot be performed. The words 'significant' and 'significantly' in the abstract and conclusion are therefore unsupported.
[Section IV-A and Section V-A] The method is motivated by distribution shifts, but the experiments never introduce one. The Cardiac and 4D-Lung datasets are split into standard train/test sets from the same acquisition protocol: 90/10 scans for Cardiac and 18/2 patients for 4D-Lung. There is no experiment with a different scanner, imaging protocol, or preprocessing that would constitute the shift described in the introduction. Without a distribution shift, the test-time adaptation would be expected to have little effect, and the small metric differences in Tables I and II are consistent with that expectation. The claim that TTT 'adapts to a new test distribution' is thus not empirically demonstrated.
[Section III-D and Table IV] The three TTT schemes are algorithmically distinct (Section III-D): Naive TTT performs multi-epoch global adaptation, Online TTT adapts independently per batch from the initial weights, and Mini-batch TTT streams sequentially. Yet Table IV shows essentially identical metrics for all three schemes across every condition. For example, on Cardiac with rotation prediction, PSNR is 33.70 +/- 0.256 for all three schemes and NMSE is 2.263 +/- 0.282 for all three; on 4D-Lung with 3D-MAE, PSNR is 34.02 for all three schemes. This pattern strongly suggests that the test-time update does not change the interpolation output. The authors acknowledge the 'surprisingly subtle' differences in Section V-B but provide only post-hoc explanations; they should report evidence that adaptation actually alters the model, for example loss curves over TTT epochs, weight-change magnitudes, or an ablation that disables TTT entirely.
[Section IV-C and Section III-C] The implementation details are not sufficient to verify the adaptation protocol or reproduce the results. The paper does not specify which parameters of the UVI-Net architecture are frozen during testing, how the 3D AlexNet feature extractor is connected to the UVI-Net decoder, what learning rate and optimizer are used during the 50 test-time epochs, or how many gradient steps are taken per batch for Online and Mini-batch TTT. The 3D-MAE loss is described only in words and not written as an equation. The code is announced only as 'TTT4DMII' with no URL. These omissions matter because the identical-results pattern in Table IV cannot be checked without knowing whether adaptation actually ran.

minor comments (6)

[Section III-C, Eq. (1)] The symbol theta is used for both model parameters and the ground-truth rotation label (theta_{i,c}); use a different symbol, such as y, for the label.
[Section IV-B] The text says 'we use five common metrics' but lists only PSNR, NCC, SSIM, and NMSE; correct the count or add the missing metric.
[Table III] The 'Time Cost' entries are given as '0.93766s per' without a unit (per sample? per batch?); specify the unit in the table header or entries.
[Section II-A] It is not explained how registration methods such as TransMorph [3] and VoxelMorph [10] are adapted to the interpolation task; add a sentence describing the baseline adaptation.
[Abstract] The phrase 'different distribution' should be plural ('different distributions'), and the paper uses 'distribution shifts' inconsistently elsewhere.
[Section IV-C] The code availability statement 'The code is available at TTT4DMII' is not an actionable identifier; provide a URL or repository name.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the TTT adaptation objective is defined on the test inputs only, and the interpolation targets are never used in adaptation, while all comparisons are against external baselines.

full rationale

The paper proposes a test-time training framework for 4D medical image interpolation. The derivation chain is self-contained: the self-supervised auxiliary loss Lg(g(f(x)), x) (rotation prediction or 3D-MAE reconstruction) is minimized on the input frames, and the interpolation prediction h(f(x)) is evaluated on the intermediate frame, which is never used in the adaptation. There is no equation in which the target interpolation frame appears in the training objective, so the prediction does not reduce to the fit by construction. Comparisons are made against external baselines (UVI-Net, VoxelMorph, etc.) rather than against a self-cited or self-defined quantity. The paper contains no load-bearing self-citations: all method components (TTT, rotation prediction, MAE) are cited from independent prior work, and the base model UVI-Net is by different authors. Although the reported improvements are small and within the paper's own standard deviations, and although no explicit distribution shift is tested, these are concerns about statistical significance and experimental design, not about circularity. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper relies on established machine learning concepts (TTT, rotation prediction, MAE) without inventing new entities. The main untested assumptions are the transferability of self-supervised signals to the interpolation target and the presence of a distribution shift in the evaluation setup.

free parameters (2)

Loss weight between main interpolation loss and self-supervised auxiliary loss during joint training = not reported
The paper states that training optimizes both the main network and the self-supervised network, but the relative weighting of these losses is never specified. This choice affects the learned representation and is needed to reproduce the method.
Mask ratio for 3D masked autoencoder = 80%
The masking ratio is set to 80%, a standard choice for MAE, but it is a manually chosen hyperparameter that influences the difficulty of the auxiliary task and therefore the adaptation behavior.

assumptions (3)

domain assumption Self-supervised auxiliary tasks on test data improve the shared feature extractor for the interpolation task.
This is the core premise of test-time training. The paper adopts it from prior work without task-specific proof or validation for 4D medical interpolation.
domain assumption The standard test splits of Cardiac and 4D-Lung represent a distribution shift relative to the training splits.
The paper motivates TTT by distribution shift, but the experiments use ordinary same-dataset train/test splits. No synthetic or real shift is introduced, so this assumption is untested.
domain assumption The shared feature extractor with UVI-Net as the base model provides a sufficient architecture for the interpolation task.
The paper selects UVI-Net as the base and does not analyze whether the auxiliary self-supervised head could interfere with the main task, or whether the architecture choice limits the benefit of TTT.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Test Time Training for 4D Medical Image Interpolation." pith.science (2026). https://pith.science/paper/KPGQC3JE

@misc{pith2026250202341,
  author       = {Pith},
  title        = {Pith review of: Test Time Training for 4D Medical Image Interpolation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPGQC3JE}},
  note         = {Machine review of arXiv:2502.02341}
}

read the original abstract

4D medical image interpolation is essential for improving temporal resolution and diagnostic precision in clinical applications. Previous works ignore the problem of distribution shifts, resulting in poor generalization under different distribution. A natural solution would be to adapt the model to a new test distribution, but this cannot be done if the test input comes without a ground truth label. In this paper, we propose a novel test time training framework which uses self-supervision to adapt the model to a new distribution without requiring any labels. Indeed, before performing frame interpolation on each test video, the model is trained on the same instance using a self-supervised task, such as rotation prediction or image reconstruction. We conduct experiments on two publicly available 4D medical image interpolation datasets, Cardiac and 4D-Lung. The experimental results show that the proposed method achieves significant performance across various evaluation metrics on both datasets. It achieves higher peak signal-to-noise ratio values, 33.73dB on Cardiac and 34.02dB on 4D-Lung. Our method not only advances 4D medical image interpolation but also provides a template for domain adaptation in other fields such as image segmentation and image registration.

Figures

Figures reproduced from arXiv: 2502.02341 by the authors.

**Figure 1.** Up: Training before deployment. A standard training workflow, we optimize the self-supervised objective. Down: Training during deployment. Observations are collected from the test set, and we optimize only the auxiliary self-supervised objective. inspired by masked autoencoders [9], we implement a 3D MAE model designed for slice-based image inputs. In this setting, we remove parts of the input image, and the model l… view at source ↗

**Figure 2.** Visualization of three TTT schemes: (a) Na¨ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Visualization of interpolation results on the Cardiac and 4D-Lung dataset. The top row shows the predicted frames (left) and the ground truth (right) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

[1]

A spatiotemporal volumetric interpolation network for 4d dynamic medical image,

Y . Guo, L. Bi, E. Ahn, D. D. Feng, Q. Wang, and J. Kim, “A spatiotemporal volumetric interpolation network for 4d dynamic medical image,” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 4725–4734, 2020

work page 2020
[2]

Mpvf: 4d medical image inpainting by multi-pyramid voxel flows,

T.-T. Wei, C.-T. Kuo, Y .-C. Tseng, and J.-J. Chen, “Mpvf: 4d medical image inpainting by multi-pyramid voxel flows,” IEEE Journal of Biomedical and Health Informatics, vol. 27, pp. 5872–5882, 2023

work page 2023
[3]

Transmorph: Transformer for unsupervised med- ical image registration,

J. Chen, E. C. Frey, Y . He, W. P. Segars, Y . Li, and Y . Du, “Transmorph: Transformer for unsupervised med- ical image registration,” Medical Image Analysis, vol. 82, p. 102615, 2022

work page 2022
[4]

Fourier-net+: Leveraging band-limited repre- sentation for efficient 3d medical image registration,

X. Jia, A. Thorley, A. Gomez, W. Lu, D. Kotecha, and J. Duan, “Fourier-net+: Leveraging band-limited repre- sentation for efficient 3d medical image registration,” 2023

work page 2023
[5]

R2net: Efficient and flexible diffeomorphic image registration using lipschitz continu- ous residual networks,

A. Joshi and Y . Hong, “R2net: Efficient and flexible diffeomorphic image registration using lipschitz continu- ous residual networks,” Medical Image Analysis , vol. 89, p. 102917, 2023

work page 2023
[6]

Implicit neural representations for deformable image registration,

J. M. Wolterink, J. C. Zwienenberg, and C. Brune, “Implicit neural representations for deformable image registration,” in Proceedings of The 5th International Conference on Medical Imaging with Deep Learning (E. Konukoglu, B. Menze, A. Venkataraman, C. Baum- gartner, Q. Dou, and S. Albarqouni, eds.), vol. 172 of Proceedings of Machine Learning Research , ...

work page 2022
[7]

Test-time training with self-supervision for generalization under distribution shifts,

Y . Sun, X. Wang, Z. Liu, J. Miller, A. Efros, and M. Hardt, “Test-time training with self-supervision for generalization under distribution shifts,” in International conference on machine learning , pp. 9229–9248, PMLR, 2020

work page 2020
[8]

Unsuper- vised domain adaptation through self-supervision,

Y . Sun, E. Tzeng, T. Darrell, and A. A. Efros, “Unsuper- vised domain adaptation through self-supervision,” arXiv preprint arXiv:1909.11825, 2019

arXiv 1909

Show all 22 references

[9]

Masked autoencoders are scalable vision learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 16000–16009, 2022

2022
[10]

V oxelmorph: A learning framework for deformable medical image registration,

G. Balakrishnan, A. Zhao, M. R. Sabuncu, J. V . Guttag, and A. V . Dalca, “V oxelmorph: A learning framework for deformable medical image registration,” IEEE Transac- tions on Medical Imaging , vol. 38, pp. 1788–1800, 2018

2018
[11]

Diffusion deformable model for 4d temporal medical image generation,

B. Kim and J.-C. Ye, “Diffusion deformable model for 4d temporal medical image generation,” ArXiv, vol. abs/2206.13295, 2022

2022 arXiv
[12]

Data- efficient unsupervised interpolation without any interme- diate frame for 4d medical images,

J. Kim, H. Yoon, G. Park, K. Kim, and E. Yang, “Data- efficient unsupervised interpolation without any interme- diate frame for 4d medical images,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11353–11364, 2024

2024
[13]

A comprehensive survey on test-time adaptation under distribution shifts,

J. Liang, R. He, and T. Tan, “A comprehensive survey on test-time adaptation under distribution shifts,” Inter- national Journal of Computer Vision , pp. 1–34, 2024

2024
[14]

Test- time training with masked autoencoders,

Y . Gandelsman, Y . Sun, X. Chen, and A. Efros, “Test- time training with masked autoencoders,” Advances in Neural Information Processing Systems , vol. 35, pp. 29374–29385, 2022

2022
[15]

Test time training for industrial anomaly segmentation,

A. Costanzino, P. Z. Ramirez, M. D. Moro, A. Aiezzo, G. Lisanti, S. Salti, and L. D. Stefano, “Test time training for industrial anomaly segmentation,” 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 3910–3920, 2024

2024
[16]

Test-time training for matching-based video object segmentation,

J. Bertrand, G. Kordopatis-Zilos, Y . Kalantidis, and G. Tolias, “Test-time training for matching-based video object segmentation,” in Thirty-seventh Conference on Neural Information Processing Systems , 2023

2023
[17]

Unsupervised learning of visual representations by solving jigsaw puzzles,

M. Noroozi and P. Favaro, “Unsupervised learning of visual representations by solving jigsaw puzzles,” in European conference on computer vision , pp. 69–84, Springer, 2016

2016
[18]

Self-supervised learning for medical image analysis using image context restoration,

L. Chen, P. Bentley, K. Mori, K. Misawa, M. Fujiwara, and D. Rueckert, “Self-supervised learning for medical image analysis using image context restoration,” Medical image analysis , vol. 58, p. 101539, 2019

2019
[19]

Multimodal self-supervised learning for medical image analysis,

A. Taleb, C. Lippert, T. Klein, and M. Nabi, “Multimodal self-supervised learning for medical image analysis,” in International conference on information processing in medical imaging , pp. 661–673, Springer, 2021

2021
[20]

Unsupervised visual representation learning by context prediction,

C. Doersch, A. K. Gupta, and A. A. Efros, “Unsupervised visual representation learning by context prediction,” 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1422–1430, 2015

2015
[21]

Unsupervised representation learning by predicting image rotations,

S. Gidaris, P. Singh, and N. Komodakis, “Unsupervised representation learning by predicting image rotations,” arXiv preprint arXiv:1803.07728 , 2018

2018 arXiv
[22]

A simple framework for contrastive learning of visual representations,

T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning, pp. 1597–1607, PMLR, 2020

2020

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

A spatiotemporal volumetric interpolation network for 4d dynamic medical image,

Y . Guo, L. Bi, E. Ahn, D. D. Feng, Q. Wang, and J. Kim, “A spatiotemporal volumetric interpolation network for 4d dynamic medical image,” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 4725–4734, 2020

work page 2020

[2] [2]

Mpvf: 4d medical image inpainting by multi-pyramid voxel flows,

T.-T. Wei, C.-T. Kuo, Y .-C. Tseng, and J.-J. Chen, “Mpvf: 4d medical image inpainting by multi-pyramid voxel flows,” IEEE Journal of Biomedical and Health Informatics, vol. 27, pp. 5872–5882, 2023

work page 2023

[3] [3]

Transmorph: Transformer for unsupervised med- ical image registration,

J. Chen, E. C. Frey, Y . He, W. P. Segars, Y . Li, and Y . Du, “Transmorph: Transformer for unsupervised med- ical image registration,” Medical Image Analysis, vol. 82, p. 102615, 2022

work page 2022

[4] [4]

Fourier-net+: Leveraging band-limited repre- sentation for efficient 3d medical image registration,

X. Jia, A. Thorley, A. Gomez, W. Lu, D. Kotecha, and J. Duan, “Fourier-net+: Leveraging band-limited repre- sentation for efficient 3d medical image registration,” 2023

work page 2023

[5] [5]

R2net: Efficient and flexible diffeomorphic image registration using lipschitz continu- ous residual networks,

A. Joshi and Y . Hong, “R2net: Efficient and flexible diffeomorphic image registration using lipschitz continu- ous residual networks,” Medical Image Analysis , vol. 89, p. 102917, 2023

work page 2023

[6] [6]

Implicit neural representations for deformable image registration,

J. M. Wolterink, J. C. Zwienenberg, and C. Brune, “Implicit neural representations for deformable image registration,” in Proceedings of The 5th International Conference on Medical Imaging with Deep Learning (E. Konukoglu, B. Menze, A. Venkataraman, C. Baum- gartner, Q. Dou, and S. Albarqouni, eds.), vol. 172 of Proceedings of Machine Learning Research , ...

work page 2022

[7] [7]

Test-time training with self-supervision for generalization under distribution shifts,

Y . Sun, X. Wang, Z. Liu, J. Miller, A. Efros, and M. Hardt, “Test-time training with self-supervision for generalization under distribution shifts,” in International conference on machine learning , pp. 9229–9248, PMLR, 2020

work page 2020

[8] [8]

Unsuper- vised domain adaptation through self-supervision,

Y . Sun, E. Tzeng, T. Darrell, and A. A. Efros, “Unsuper- vised domain adaptation through self-supervision,” arXiv preprint arXiv:1909.11825, 2019

arXiv 1909