Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Reverse Imaging for Wide-spectrum Generalization of Cardiac MRI Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that cardiac MRI segmentation models can generalize to unseen imaging sequences by first reverse-inferring tissue spin properties (PD, T1, T2) from a single bSSFP image, then synthesizing training images of arbitrary new se

desk verdict Physics-based augmentation with strong empirical gains; the reverse-inference framing overstates identifiability, but the method and results merit serious review. read the letter →

arxiv 2508.21254 v1 pith:CMG55LV4 submitted 2025-08-28 cs.CV

Yidong Zhao , Peter Kellman , Hui Xue , Tongyun Yang , Yi Zhang , Yuchi Han , Orlando Simonetti , Qian Tao This is my paper
classification cs.CV
keywords cardiacMRIsegmentationdomaingeneralizationreverseimagingdiffusionmodelspinpropertiesdataaugmentationMRphysics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the reason segmentation models fail on new MRI sequences is not a missing style-transfer trick, but a missing physical variable: every MRI image is determined by the same underlying tissue spin properties — proton density, T1, and T2 — plus the specific pulse sequence used to acquire it. The authors introduce Reverse Imaging, which estimates those spin properties from a standard bSSFP cine image by running a diffusion model guided by the MRI physics forward model, then uses the estimated properties to synthesize training images of arbitrary sequences, such as MOLLI T1 mapping and GRE. A nnUNet trained on ACDC bSSFP images augmented this way reaches LV Dice of 91.6% on real MOLLI images and 96.0% on device images, versus 39.9% and 91.9% for the baseline model. The method is zero-shot: it needs no target-domain images at inference, only a learned prior over spin properties from the mSASHA T1/T2 mapping dataset. If correct, this suggests that cross-sequence generalization in cardiac MRI is a solvable physics problem rather than an open domain-adaptation challenge.

What carries the argument

The load-bearing mechanism is the guided reverse diffusion step in Eq. (10): at each denoising step, the DDPM score estimate is combined with the gradient of the squared error between the observed image and the forward model applied to the denoised spin estimate. This couples the learned spin prior with the MRI physics likelihood, producing an approximate posterior sample that reconstructs the input and yields new contrasts when other sequence equations are applied. The spin prior itself is a DDPM trained on mSASHA-derived joint T1/T2 maps, and the forward models used are the standard bSSFP, MOLLI, and GRE signal equations.

What would settle it

Take a subject or phantom with both a bSSFP image and a reference quantitative T1/T2 map acquired at a known flip angle; run Reverse Imaging on the bSSFP image and compare the estimated T1/T2 maps voxel-wise with the reference. If the mismatch is large or if augmenting with images synthesized from the estimated maps gives significantly worse segmentation Dice than augmenting with images synthesized from the true maps, the method's utility depends on the accuracy of its spin inference rather than on its data-augmentation principle. A cheaper test: repeat the MOLLI experiment with the flip angle

Watch

Extended reading notes

Core claim

The central discovery is that ill-posed spin-property estimation from a single qualitative MRI image becomes feasible when a generative prior learned from quantitative T1/T2 mapping data regularizes the inversion, and that the resulting approximate spin maps are sufficient to synthesize training images whose contrasts span the sequences a segmentation model will later face. The authors formalize cross-sequence translation as sampling from p(xj|xi) by introducing a latent z=(PD,T1,T2), with likelihoods given by the Bloch-equation forward models of each sequence and a diffusion model over z learned from mSASHA as the prior. Reverse Imaging solves the posterior p(z|xi) via guided reverse diffus

Load-bearing premise

The entire pipeline assumes the ACDC bSSFP images are accurately described by the bSSFP signal equation evaluated at a fixed 45-degree flip angle; any mismatch between this model and the true acquisition parameters biases the inferred spin properties and hence every synthesized training image.

Editorial extensions

If this is right

  • A segmentation model trained once on bSSFP plus physics-synthesized MOLLI and GRE images can segment real MOLLI and device images zero-shot, without ever seeing target-domain data.
  • Target-to-source translation also works, but is less reliable for GRE because the short echo time carries almost no T2 weighting; using Reverse Imaging for augmentation is the stronger strategy.
  • The estimated spin properties, though approximate, reproduce known relative T1/T2 contrast patterns in the heart, giving the method an interpretable intermediate representation.
  • Any new imaging sequence with a known signal equation can be added to the augmentation, widening the spectrum of covered contrasts without retraining the inverse step.

Reading between the lines

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

  • The success of the method suggests that physics-driven augmentation could replace style-transfer domain adaptation in other MRI applications, such as brain or abdominal imaging, whenever a quantitative spin-prior dataset is available — the requirement is a forward model, not paired target data.
  • A direct test of the weakest assumption would be to re-run Reverse Imaging with the bSSFP flip angle set to other plausible values (e.g., 30 or 60 degrees) and check whether segmentation gains change materially; if they do, the method's robustness hinges on accurate sequence parameter knowledge.
  • The GRE result reveals an information-theoretic limit: when the source image carries little T2 weighting, T2 cannot be well inferred, so T2-dependent synthetic contrasts degrade. This predicts that providing multiple source images with complementary contrast (e.g., different flip angles) would improve the inferred spin properties.
  • The released estimated spin properties for the ACDC dataset could serve as a public testbed for comparing physics-constrained inverse solvers against reference quantitative maps where those coexist for the same subjects.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Reverse Imaging, a physics-driven method for cross-sequence generalization of cardiac MRI segmentation. The core idea is to estimate voxel-wise spin properties (PD, T1, T2) from observed images by solving an ill-posed inverse problem regularized by a diffusion prior learned from mSASHA quantitative maps. From the estimated spin maps, the authors synthesize images of unseen sequences (MOLLI, GRE) using Bloch-equation forward models. They evaluate two uses: target-to-source translation (RI-T2S) and physics-based augmentation (RI-Aug) of a nnUNet trained on ACDC bSSFP. Reported results show large Dice improvements on MOLLI and device GRE/bSSFP datasets compared with baseline and BayeSeg, e.g., LV Dice 91.6% vs 39.9% on MOLLI and 96.0% vs 91.9% on device data.

Significance. If the mechanism and results hold, the paper would make a valuable contribution: zero-shot cross-sequence generalization without target-domain data, with an interpretable latent space tied to MR physics. The empirical improvement is large and the release of code and estimated spin maps is a practical strength. However, the central scientific claim—that the method 'reversely infers' subject-specific spin properties from a single bSSFP image—is not established. The bSSFP signal equation depends only on PD and the T1/T2 ratio, so the data likelihood is flat along the T1-T2 direction; the absolute T1 and T2 values in the posterior are supplied by the mSASHA prior rather than by the source image. This does not necessarily invalidate the augmentation gains, but it changes the interpretation from subject-level reverse inference to prior-driven contrast synthesis. The manuscript also fixes the ACDC flip angle to 45° without sensitivity analysis, and lacks comparisons to standard intensity augmentation or synthetic-inversion sanity checks. These issues are addressable, so major revision is appropriate.

major comments (4)
  1. [§2.3, Eq. (10) and Eq. (4)] The likelihood term in the reverse-diffusion guidance uses fi(˜z0(zt)) with the bSSFP model fSS(z)=PD sinω/(1+cosω+(1−cosω)T1/T2). This function is invariant to any simultaneous change in T1 and T2 that preserves r=T1/T2 and PD. Consequently, the gradient in Eq. (10) carries no information that separates T1 from T2; the posterior over absolute T1 and T2 is inherited from the diffusion prior. The paper claims to 'reversely infer the underlying spin properties' and that the same subject-specific z drives synthesis. What is actually demonstrated is that the image constrains PD and the ratio, while the prior fills in the absolute scales. The synthesis may still be useful if the prior is representative, which would explain the segmentation gains, but the paper should either (a) provide explicit evidence of subject-level identifiability, e.g., a synthetic experiment where a bSSFP image is gene
  2. [§3, Experiments] The flip angle of ACDC bSSFP is unknown and fixed to ω=45° without any sensitivity analysis. Because fSS in Eq. (4) depends on ω nonlinearly, a wrong FA will bias the estimated PD and the T1/T2 ratio, and consequently all synthesized MOLLI and GRE images used for RI-Aug and RI-T2S. The authors should report segmentation Dice for a range of plausible FAs (e.g., 30°, 45°, 60°) and, ideally, per-subject or vendor-specific FA information if available. Without this, the robustness of the method to a key free parameter is unquantified.
  3. [§4.2, Table 1 and Fig. 3] The comparison does not include a control for generic data augmentation. The reported gain of RI-Aug over baseline could in principle be due to the additional synthetic training data alone, independent of the physical plausibility of the synthesized contrasts. The authors should compare against a baseline trained with standard intensity augmentation (e.g., random brightness/contrast/gamma, or simple contrast-inversion augmentation) and ideally against a non-physics generative augmentation method. This would isolate the contribution of physics-based synthesis.
  4. [§4.1 and Fig. 2] The paper states that estimated spin properties are 'not used for quantitative evaluation purposes,' but the central narrative depends on the estimated z being meaningful. A direct quantitative validation of the estimated PD/T1/T2 is missing. Since mSASHA provides joint T1/T2 maps, a natural experiment is: use mSASHA-derived z to simulate bSSFP images, run Reverse Imaging on those simulated images, and compare the recovered z with the ground-truth mSASHA maps (e.g., Bland-Altman or correlation). This would concretely test how much of z is recovered from the image versus filled in by the prior. Without such a test, the interpretability claim remains speculative.
minor comments (5)
  1. [§2.1, Eq. (1)] The text says 'the inference p(xi|xj)' but the equation and context concern p(xj|xi); please correct the direction notation.
  2. [§3, Datasets] The dataset description says 'Dev-bSSPF' but the abbreviation is inconsistent with the main text's use of 'Dev-bSSFP'; please standardize.
  3. [§3, Experimental Settings] The sentence 'The DDPM is pretrained on ACDC images for 36,000 steps and then fine-tuned with the real spin properties' is confusing: the prior is supposed to be a distribution over z=(PD,T1,T2), not over images. Please clarify whether the pretraining is on image space or z space, and how the fine-tuning is performed.
  4. [§4.2, Statistical tests] The statistical significance statement (p<0.05) lacks detail on the test used and whether multiple comparisons were corrected. Please specify the test and how the 131 MOLLI slices and 25 device subjects are treated (independent vs clustered).
  5. [Fig. 2] The colorbars for estimated spin properties are unlabeled; adding units (e.g., ms for T1/T2) and a scale for PD would improve interpretability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reverse-inference prior is external and target-domain data are unseen.

full rationale

The claimed derivation chain does not reduce to its inputs. The spin prior p(z) is trained on mSASHA-derived T1/T2 maps [7], which are external quantitative measurements, not outputs of the ACDC images or the target MOLLI/Device data. The likelihoods are standard Bloch signal equations (Eqs. 4-6), and the guided reverse diffusion (Eq. 10) follows the standard diffusion-posterior-sampling formulation [8]. RI-Aug synthesizes MOLLI and GRE training images from source ACDC images via forward models and evaluates on real target data, so the target-domain contrast is never injected into training or fitting. The bSSFP model (Eq. 4) depends on PD and the ratio T1/T2, so individual T1 and T2 are not identifiable from a single bSSFP image and are informed by the prior; the paper explicitly acknowledges this by stating Reverse Imaging 'does not provide precise spin property estimation' and that estimates are not used for quantitative evaluation (Sec. 4.1), and it approximates the ACDC flip angle as 45 degrees (Sec. 3). These are identifiability/calibration limitations, not circularity. The only self-citation is the mSASHA source [7] (co-author Kellman), but it is an empirical dataset and is not used to define the segmentation prediction; it is therefore not load-bearing. No equation reduces to the fitted target or to a self-citation chain.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

Central claim rests on the physical premise that spin properties are shared across sequences, on the fidelity of three forward signal models, and on a prior learned from 63 mSASHA subjects. The main hand-chosen numerical values are the ACDC flip angle (approximated to 45 degrees) and the guidance step size xi=400. No new physical entities are introduced.

free parameters (2)
  • Flip angle for ACDC bSSFP (omega) = 45 degrees (approximation)
    Section 3: exact FA for ACDC scans is unknown, so the forward model Eq. (4) is evaluated at 45 degrees; estimated spin properties and all synthetic contrasts inherit this choice.
  • Guidance step size xi = 400
    Section 3, used in Eq. (10) for Reverse Imaging; no sensitivity study given.
assumptions (5)
  • standard math Bayesian decomposition p(xj|xi)=integral p(xj|z)p(z|xi)dz with z=(PD,T1,T2)
    Eq. (1)-(2) in Section 2.1; assumes z is a sufficient latent variable separating sequence appearance from tissue content.
  • domain assumption Different MR sequences of the same subject share identical underlying spin properties z
    Section 1 and Fig. 1; fundamental premise, plausible but ignores motion, flow, B1 and off-resonance effects beyond the signal equations.
  • domain assumption Signal equations (4)-(6) are accurate forward models for bSSFP, MOLLI and GRE
    Section 2.1; Eq. (4) assumes balanced SSFP with fixed 45 degree flip angle and no off-resonance; Eq. (5) uses phase-insensitive MOLLI approximation; Eq. (6) assumes spoiled GRE.
  • domain assumption mSASHA T1/T2 maps provide a representative prior over cardiac spin properties
    Section 2.2; trained on 63 subjects, 186 slices; may under-represent pathology extremes.
  • domain assumption PD approximately equals A from mSASHA fitting is sufficient for contrast synthesis
    Section 2.2 after Eq. (7), citing [1,2]; amplitude A carries PD-like scaling but is not a direct proton-density measurement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reverse Imaging for Wide-spectrum Generalization of Cardiac MRI Segmentation." pith.science (2026). https://pith.science/paper/CMG55LV4

@misc{pith2026250821254,
  author       = {Pith},
  title        = {Pith review of: Reverse Imaging for Wide-spectrum Generalization of Cardiac MRI Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMG55LV4}},
  note         = {Machine review of arXiv:2508.21254}
}
read the original abstract

Pretrained segmentation models for cardiac magnetic resonance imaging (MRI) struggle to generalize across different imaging sequences due to significant variations in image contrast. These variations arise from changes in imaging protocols, yet the same fundamental spin properties, including proton density, T1, and T2 values, govern all acquired images. With this core principle, we introduce Reverse Imaging, a novel physics-driven method for cardiac MRI data augmentation and domain adaptation to fundamentally solve the generalization problem. Our method reversely infers the underlying spin properties from observed cardiac MRI images, by solving ill-posed nonlinear inverse problems regularized by the prior distribution of spin properties. We acquire this "spin prior" by learning a generative diffusion model from the multiparametric SAturation-recovery single-SHot acquisition sequence (mSASHA) dataset, which offers joint cardiac T1 and T2 maps. Our method enables approximate but meaningful spin-property estimates from MR images, which provide an interpretable "latent variable" that lead to highly flexible image synthesis of arbitrary novel sequences. We show that Reverse Imaging enables highly accurate segmentation across vastly different image contrasts and imaging protocols, realizing wide-spectrum generalization of cardiac MRI segmentation.

Figures

Figures reproduced from arXiv: 2508.21254 by the authors.

Figure 1
Figure 1. Images acquired by sequences i and j from a subject share the same underlying spin properties z. Reverse Imaging (dashed arrows) estimates zˆ from the prior p(z) that explains the observed image xi. Images of an arbitrary sequence can be generated by MR physics from zˆ (solid arrows). In this process, the only unknown is the prior p(z), which we learn from the mSASHA data by a diffusion model (red). Imaging. Reverse… view at source ↗
Figure 2
Figure 2. (a) Reverse Imaging a bSSFP image xi (yellow). The estimated zˆ accurately reconstructs xˆi ≈ xi. (b) Generating GRE (red) and MOLLI (blue) images with zˆ. The FA in degrees and TR/TE or tINV in milliseconds are given for each image. (c) Reverse Imaging a MOLLI readout for target-to-source (T2S) translation. evaluate the robustness against contrast variation, we also report the Dice for each baseline image of MOLLI … view at source ↗
Figure 3
Figure 3. Segmentation accuracy on each MOLLI readout during inversion recovery. Image Baseline BayeSeg RI-Aug. LL-#1 LL-#2 LL-#3 LL-#4 LL-#5 Dev-GRE Dev-bSSFP [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Segmentation of 5 MOLLI baseline images (LL-#1-#5) and 2 device cine im￾ages, all of which have different contrast than bSSFP cine images used for training. T2S transform (RI-T2S) significantly improved the segmentation accuracy for MOLLI, but a performance drop is obs…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

  1. [1]

    Magnetic resonance in medicine74(1), 93–105 (2015)

    Akçakaya, M., Basha, T.A., Weingärtner, S., Roujol, S., Berg, S., Nezafat, R.: Im- proved quantitative myocardial T2 mapping: Impact of the fitting model. Magnetic resonance in medicine74(1), 93–105 (2015)

    work page 2015

  2. [2]

    Magnetic resonance in medicine76(3), 888–896 (2016)

    Akçakaya, M., Weingärtner, S., Basha, T.A., Roujol, S., Bellm, S., Nezafat, R.: Joint myocardial T1 and T2 mapping using a combination of saturation recovery and t2-preparation. Magnetic resonance in medicine76(3), 888–896 (2016)

    work page 2016

  3. [3]

    Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging37(11), 2514–2525 (2018)

    work page 2018

  4. [4]

    IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

    Campello, V.M., Gkontra, P., Izquierdo, C., Martin-Isla, C., Sojoudi, A., Full, P.M., Maier-Hein, K., Zhang, Y., He, Z., Ma, J., et al.: Multi-centre, multi-vendor and multi-disease cardiac segmentation: the M&Ms challenge. IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

    work page 2021

  5. [5]

    In: Sta- tistical Atlases and Computational Models of the Heart

    Chen, C., Ouyang, C., Tarroni, G., Schlemper, J., Qiu, H., Bai, W., Rueckert, D.: Unsupervised multi-modal style transfer for cardiac MR segmentation. In: Sta- tistical Atlases and Computational Models of the Heart. Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full Quantification Challenges: 10th Interna- tional Workshop, STACOM 2019, Held in Conjun...

    work page 2019

  6. [6]

    Medical Image Analysis82, 102597 (2022)

    Chen, C., Qin, C., Ouyang, C., Li, Z., Wang, S., Qiu, H., Chen, L., Tarroni, G., Bai,W.,Rueckert,D.:EnhancingMRimagesegmentationwithrealisticadversarial data augmentation. Medical Image Analysis82, 102597 (2022)

    work page 2022

  7. [7]

    Magnetic Resonance in Medicine87(6), 2775–2791 (2022)

    Chow, K., Hayes, G., Flewitt, J.A., Feuchter, P., Lydell, C., Howarth, A., Pagano, J.J., Thompson, R.B., Kellman, P., White, J.A.: Improved accuracy and precision with three-parameter simultaneous myocardial T1 and T2 mapping using multi- parametric SASHA. Magnetic Resonance in Medicine87(6), 2775–2791 (2022)

    work page 2022

  8. [8]

    arXiv preprint arXiv:2209.14687 (2022) 10 Y

    Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sam- pling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022) 10 Y. Zhao et al

    arXiv 2022

Show all 29 references
  1. [9]

    IEEE Transactions on Medical Imaging40(12), 3604–3616 (2021)

    Cui, Z., Li, C., Du, Z., Chen, N., Wei, G., Chen, R., Yang, L., Shen, D., Wang, W.: Structure-driven unsupervised domain adaptation for cross-modality cardiac segmentation. IEEE Transactions on Medical Imaging40(12), 3604–3616 (2021)

    2021

  2. [10]

    Journal of Cardiovascular Magnetic Resonance 15(1), 41 (2013)

    Ferreira, P.F., Gatehouse, P.D., Mohiaddin, R.H., Firmin, D.N.: Cardiovascu- lar magnetic resonance artefacts. Journal of Cardiovascular Magnetic Resonance 15(1), 41 (2013)

    2013

  3. [11]

    In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

    Gao, S., Zhou, H., Gao, Y., Zhuang, X.: Joint modeling of image and label statistics for enhancing model generalizability of medical image segmentation. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 360–369. Springer (2022)

    2022

  4. [12]

    Medical Image Analysis 89, 102889 (2023)

    Gao, S., Zhou, H., Gao, Y., Zhuang, X.: BayeSeg: Bayesian modeling for medical image segmentation with interpretable generalizability. Medical Image Analysis 89, 102889 (2023)

    2023

  5. [13]

    Lippin- cott Williams & Wilkins (2012)

    Hashemi, R.H., Bradley, W.G., Lisanti, C.J.: MRI: the basics: The Basics. Lippin- cott Williams & Wilkins (2012)

    2012

  6. [14]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  7. [15]

    Nature methods 18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021)

    2021

  8. [16]

    Jour- nal of cardiovascular magnetic resonance16(1), 2 (2014)

    Kellman, P., Hansen, M.S.: T1-mapping in the heart: accuracy and precision. Jour- nal of cardiovascular magnetic resonance16(1), 2 (2014)

    2014

  9. [17]

    IEEE Journal of Biomedical and Health Informatics27(7), 3302–3313 (2023)

    Martín-Isla, C., Campello, V.M., Izquierdo, C., Kushibar, K., Sendra-Balcells, C., Gkontra, P., Sojoudi, A., Fulton, M.J., Arega, T.W., Punithakumar, K., et al.: Deep learning segmentation of the right ventricle in cardiac MRI: the M&Ms chal- lenge. IEEE Journal of Biomedical ...

    2023

  10. [18]

    Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine52(1), 141–146 (2004)

    Messroghli, D.R., Radjenovic, A., Kozerke, S., Higgins, D.M., Sivananthan, M.U., Ridgway, J.P.: Modified look-locker inversion recovery (MOLLI) for high-resolution t1 mapping of the heart. Magnetic Resonance in Medicine: An Official Journal of the International Society for Mag...

    2004

  11. [19]

    IEEE Transactions on Medical Imaging42(4), 1095–1106 (2022)

    Ouyang, C., Chen, C., Li, S., Li, Z., Qin, C., Bai, W., Rueckert, D.: Causality- inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging42(4), 1095–1106 (2022)

    2022

  12. [20]

    In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd In- ternational Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22

    Ouyang, C., Kamnitsas, K., Biffi, C., Duan, J., Rueckert, D.: Data efficient un- supervised domain adaptation for cross-modality image segmentation. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd In- ternational Conference, Shenzhen, China, Oc...

    2019

  13. [21]

    Signal Processing Magazine, IEEE25(2), 72–82 (2008)

    Pauly, J.M.: Compressed sensing MRI. Signal Processing Magazine, IEEE25(2), 72–82 (2008)

    2008

  14. [22]

    https://github.com/huggingface/diffusers (2022)

    von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Nair, D., Paul, S., Berman, W., Xu, Y., Liu, S., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)

    2022

  15. [23]

    Journal of cardiovascular magnetic resonance12(1), 71 (2010)

    Ridgway, J.P.: Cardiovascular magnetic resonance physics for clinicians: part I. Journal of cardiovascular magnetic resonance12(1), 71 (2010)

    2010

  16. [24]

    European radiology13, 2409–2418 (2003) Reverse Imaging for CMR Segmentation 11

    Scheffler, K., Lehnhardt, S.: Principles and applications of balanced SSFP tech- niques. European radiology13, 2409–2418 (2003) Reverse Imaging for CMR Segmentation 11

    2003

  17. [25]

    Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine51(4), 661–667 (2004)

    Schmitt, P., Griswold, M.A., Jakob, P.M., Kotas, M., Gulani, V., Flentje, M., Haase, A.: Inversion recovery TrueFISP: quantification of t1, t2, and spin density. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine...

    2004

  18. [26]

    arXiv preprint arXiv:2111.08005 (2021)

    Song, Y., Shen, L., Xing, L., Ermon, S.: Solving inverse problems in medical imag- ing with score-based generative models. arXiv preprint arXiv:2111.08005 (2021)

    2021 arXiv

  19. [27]

    Radiology290(1), 81–88 (2019)

    Tao, Q., Yan, W., Wang, Y., Paiman, E.H., Shamonin, D.P., Garg, P., Plein, S., Huang, L., Xia, L., Sramko, M., et al.: Deep learning–based method for fully auto- matic quantification of left ventricle function from cine MR images: a multivendor, multicenter study. Radiology290...

    2019

  20. [28]

    IEEE Transactions on Medical Imaging 43(1), 4–14 (2022)

    Xie, Q., Li, Y., He, N., Ning, M., Ma, K., Wang, G., Lian, Y., Zheng, Y.: Unsu- pervised domain adaptation for medical image segmentation by disentanglement learning and self-training. IEEE Transactions on Medical Imaging 43(1), 4–14 (2022)

    2022

  21. [29]

    Journal of Cardiovascular Magnetic Resonance25(1), 64 (2023)

    Xu, Z., Li, W., Wang, J., Wang, F., Sun, B., Xiang, S., Luo, X., Meng, Y., Wang, X., Wang, X., et al.: Reference ranges of myocardial t1 and t2 mapping in healthy chinese adults: a multicenter 3t cardiovascular magnetic resonance study. Journal of Cardiovascular Magnetic Reson...

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.