CUTE-MRI: Conformalized Uncertainty-based framework for Time-adaptivE MRI

Christian Baumgartner; Jan Nikolas Morshuis; Paul Fischer; Thomas K\"ustner

REVIEW 3 major objections 4 minor 1 cited by

CUTE-MRI: Conformalized Uncertainty-based framework for Time-adaptivE MRI

T0 review · 3 major / 4 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 This paper claims that an MRI scan can be stopped per patient, at the moment the scan has collected enough data for a specific clinical measurement, instead of using a fixed acceleration factor for everyone.

desk verdict A genuinely useful adaptive-MRI idea whose headline statistical guarantee does not survive its own data-dependent stopping rule; the paper's coverage numbers show it. read the letter →

arxiv 2508.14952 v1 pith:TTTCTF4U submitted 2025-08-20 eess.IV cs.LG

Paul Fischer , Jan Nikolas Morshuis , Thomas Küstner , Christian Baumgartner This is my paper

classification eess.IVcs.LG

keywords medicalimageanalysisdeeplearningMRIuncertaintyquantificationconformalpredictionadaptiveacquisitionearlystoppingsegmentation

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 claims that MRI scans can be stopped per patient, at the moment the scan has collected enough data for a specific clinical measurement, instead of using a fixed acceleration factor for everyone. The idea is to estimate uncertainty in the reconstructed image with a probabilistic model, push that uncertainty through segmentation to a scalar endpoint (knee cartilage volume or cardiac ejection fraction), and turn it into a calibrated confidence interval using conformal prediction. The scan terminates when the interval is narrow enough. The authors report that this adaptive rule cut scan time relative to fixed protocols while keeping downstream errors at the calibrated stop point low. The payoff, if the claim holds, is that the one-size-fits-all trade-off between scan speed and diagnostic confidence becomes a per-subject decision.

What carries the argument

The engine is the conformalized uncertainty interval. PHiRec generates a Monte Carlo distribution of reconstructions; segmentation maps each sample to a mask; the downstream metric (volume or ejection fraction) turns those masks into a scalar sample set. Its standard deviation σw is the raw uncertainty signal, and split conformal prediction recalibrates it via q̂, the ⌈(1−α)(n_calib+1)⌉-th order statistic of calibration scores |w_i − ŵ_i|/σ_w,i. The final interval C(y) = [ŵ − q̂σw, ŵ + q̂σw] is both the instrument that carries the argument and the stopping criterion: acquisition iterates only while q̂σw exceeds the user's precision threshold ε.

What would settle it

Take a large held-out cohort, run the full pipeline, and at the moment each scan fires its stopping rule record whether the true metric lies inside the conformal interval; the central claim survives only if the empirical coverage approaches the nominal level (the paper's own values, 61.1% and 85.7% at a 90% target, already mark this as the decisive test).

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the width of a conformalized prediction interval for a downstream clinical metric can serve as a trustworthy, real-time stop signal for MR acquisition. Starting from a heavily undersampled mask, the pipeline draws 20 reconstructions from a hierarchical conditional variational autoencoder, segments each, computes the metric distribution, and forms the interval C(y) = [ŵ − q̂σw, ŵ + q̂σw], where q̂ is the empirical quantile of calibration-set normalized errors. The scan stops when this interval shrinks below a task-specific threshold. The calibrated rule stopped knee scans at an average 4.35× acceleration and cardiac scans at 8.3×, with lower met

Load-bearing premise

The load-bearing premise is that the calibration set remains representative of the test scans in the same probabilistic way even after the stopping rule selects test cases by interval width; if early-stopped cases are systematically unlike the calibration cases, the formal 90% coverage guarantee falls apart.

Editorial extensions

If this is right

If the central claim is right, fixed acceleration protocols become unnecessary: an 'easy' knee or heart case can finish at high acceleration while a 'hard' case automatically gets more k-space, so average scan time drops without a preset quality floor being exceeded.
Clinical endpoints gain an explicit precision target: an exam can be ordered as 'stop when cartilage volume is known to ±0.5 cm³' or 'LVEF to ±15%', and the scanner enforces that bound by construction of the interval.
Raw deep-learning uncertainty alone is unsafe for this job: the uncalibrated rule stopped every knee scan at 32× with average volume error 0.91 cm³, versus 0.42 cm³ after calibration, so the conformal step is load-bearing, not cosmetic.
The roughly 28 ms per-slice cost of the reconstruction-plus-segmentation-plus-calibration loop means the stopping decision can be computed online during a real acquisition.
Because the pipeline is task-agnostic beyond the scalar metric, the same adaptive loop transfers to other reconstruction-to-measurement chains.

Reading between the lines

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

Our inference: the framework is not tied to volume or ejection fraction; any scalar endpoint computed from segmentations—lesion burden, wall thickness, tumor diameter—could drive the same stopping rule, provided a calibration set with paired ground truths exists.
Our inference: the coverage gap the paper reports is best read as a deployment warning: the marginal guarantee only holds when calibration and test data are exchangeable, so a scanner moving to a new population or protocol should re-calibrate or expect the 90% target to slip.
Our inference: a natural next test is to replace the fixed mask ladder with trajectory selection aimed at reducing expected interval width per unit scan time; the paper's design uses only temporal feedback, not spatial feedback, so that improvement remains an open question.

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

What to know up front: the paper does something new. It uses conformal prediction on a downstream clinical metric (cartilage volume, LVEF) to stop MRI acquisition per subject, and validates it on two datasets with raw k-space, a probabilistic reconstruction model, and actual segmentations. That is a sensible, useful direction with substantial experiments. The paper also shows that uncalibrated uncertainty stops far too early, a practical warning worth keeping.

The soft spot is the statistics, and it is load-bearing. Split conformal gives marginal coverage for a fixed acceleration factor. Here the stopping time is chosen from the test subject's own interval width, so the calibration set is no longer exchangeable with the cases that stop at each R. Coverage conditional on stopping can be much lower than nominal. The paper's own numbers show exactly that: 61.1% and 85.7% versus 90% in Sections 4.2 and 5.2. The authors blame distribution shift in Section 5.2, but that is insufficient. The selection effect is present by construction.

A second issue is more mundane but still real: for CINE they use ncalib=5 with alpha=0.1. The textbook quantile index ceil((1-alpha)(n+1)) = 6 does not exist among five scores. If they take the maximum, the finite-sample guarantee is at most 5/6 or about 83.3%, so even without the stopping rule, the claimed 90% guarantee is unattainable. This looks like an implementation error rather than a deep conceptual flaw, but it reinforces the same conclusion: the stated formal guarantees are not delivered.

A third, smaller issue: the paper says the adaptive approach reduces scan time versus fixed protocols, but no fixed-protocol baseline is reported. The comparison shown is uncalibrated versus calibrated stopping, which is not the same claim.

What deserves credit: the framework is modular, the experiments are substantial, and tying acquisition to downstream confidence is the right direction. The statistical flaw is fixable—nested or online conformal methods, or directly targeting coverage at the stopping time, would be concrete steps. As written, the central claim is overstated and the paper needs major revision. I would not cite it in its current form, and I would not bring it to reading group until the guarantee issue is addressed. But it is a serious attempt, not a toy, and a serious referee should spend time on it. If the authors fix the calibration and reframe the claims, this could become a solid contribution.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes CUTE-MRI, an adaptive MRI acquisition framework that iteratively acquires k-space, reconstructs with the authors' PHiRec probabilistic model, propagates uncertainty through a segmentation pipeline to a clinical metric, and stops when a conformal-calibrated interval for the metric becomes narrower than a threshold. The method is evaluated on SKM-TEA (patellar cartilage volume) and an in-house CINE dataset (LVEF). The central claim is that this adaptive approach reduces scan time compared to fixed protocols while providing formal statistical guarantees on the precision of the final result.

Significance. If the formal guarantee were valid, the contribution would be significant: patient-specific scan termination with calibrated confidence intervals is a timely and clinically relevant goal. The paper is also commendable for its concrete two-dataset validation, runtime estimates, and clear demonstration that uncalibrated uncertainty is overconfident. However, the core statistical claim is not supported by the reported evidence. The coverage at stopping is 61.1% (SKM-TEA) and 85.7% (CINE) against a 90% target, and the conformal procedure as implemented cannot provide the advertised guarantee under the data-dependent stopping rule, nor even for the CINE calibration set of size five. The work currently amounts to a promising heuristic adaptive acquisition method, not a method with the stated formal guarantees.

major comments (3)

[§2.3, Eq. (8)–(10); §3.1] The split-conformal coverage guarantee in Eq. (8) is valid for a fixed test sample and a pre-specified acceleration factor R, under exchangeability. In §3.1 the scan stops at τ = min{R: width(C_R) ≤ ε}, so τ is a function of the test subject's own calibrated interval width. The marginal guarantee for each fixed R does not imply coverage conditional on τ = R; selecting low-uncertainty cases can preferentially select poorly covered cases. The paper's own coverage at stopping is 61.1% (SKM-TEA) and 85.7% (CINE) versus the 90% target (§4.2). Section 5.2 attributes the deficit to distribution shift, but the data-dependent selection effect is a separate, unaddressed violation of the exchangeability premise. The abstract's claim of 'formal statistical guarantees' on the final stopped image is therefore unsupported. A valid guarantee would require a procedure that accounts for the stopping rule
[§3.4.3; §4.2] For CINE, n_calib = 5 and α = 0.1. The required quantile index ⌈(1−α)(n_calib+1)⌉ = ⌈5.4⌉ = 6 does not exist among the five calibration scores. If q is taken as the maximum of the five scores, the finite-sample marginal coverage of a fixed-R interval is only 5/6 ≈ 83.3% (for continuous scores), not 90%. Thus even before considering the stopping rule, the CINE calibration procedure cannot deliver the claimed 90% guarantee. This is not a distribution-shift issue; it is an implementation error in the calibration step.
[Abstract; §4.2] The abstract claims that the adaptive approach 'reduces scan times compared to fixed protocols', but no fixed-protocol baseline is reported. Section 4.2 compares calibrated with uncalibrated stopping only; it does not compare the adaptive protocol against a standard fixed acceleration factor (e.g., R = 8 or R = 4) in terms of scan time, metric error, and coverage. Without such a comparison, the efficiency claim cannot be evaluated. This is particularly important because the calibrated procedure stops at an average acceleration of only 4.35× for SKM-TEA, i.e., it often acquires near fully sampled data, and the claimed time saving relative to a fixed protocol needs direct evidence.

minor comments (4)

[§4.2] The paragraph describing the stopping behavior and the effect of calibration is repeated almost verbatim. Please remove the duplication.
[Fig. 3 caption] The caption states 'Each point represents a single reconstruction'; in context each point appears to represent a subject or volume. Please clarify.
[§2.2 vs §3.4] The downstream mapping is denoted T(·) in §2.2 and f(·) in §3.4 and Eq. (6). Unify the notation.
[§3.4.2] The LVEF calculation uses all 20×20 pairings of ED and ES samples. Please clarify whether ED and ES acquisitions are independent or paired; naive pairing may inflate the uncertainty estimate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the adaptive stopping framework is an engineering composition of PHiRec, conformal calibration, and empirical evaluation, not a derivation that reduces to its inputs.

full rationale

The paper's central pipeline takes PHiRec samples, propagates them through a fixed segmentation/metric function, calibrates the resulting sigma with split-conformal quantiles fit on a separate calibration set, and stops when the calibrated interval width crosses epsilon. Each component is independently specifiable and tested: the uncertainty source (PHiRec) is used as an off-the-shelf generative model from prior work and is validated against external ground truth; the conformal correction q_R is fit per acceleration factor on dedicated calibration subjects and then applied to held-out test subjects; the reported summary statistics (stopping acceleration, volume/LVEF error, coverage) are measured on the test set, not constructed from the calibration fit. Section 5.2 explicitly acknowledges that test-set coverage (61.1% SKM-TEA, 85.7% CINE) falls below the 90% target, attributing this to distribution shift; this is a correctness/validity limitation, not a circular derivation. The separate CINE calibration-set issue (ncalib=5 with alpha=0.1 cannot yield a 6th order statistic) is also a statistical validity gap, not circularity. No equation in the paper defines a predicted quantity in terms of the target quantity, no fitted parameter is renamed as a prediction, and the self-citations to PHiRec and related uncertainty-propagation work are not load-bearing for the conformal guarantee. Therefore the circularity pass is clean, even though the 'formal statistical guarantee' claim is overstated.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's contribution rests on standard conformal prediction plus PHiRec, a self-cited probabilistic reconstruction model. The free parameters (stopping thresholds, alpha, M) are user-defined and not justified by evidence. The key statistical assumption of exchangeability is explicitly contradicted by the paper's coverage results.

free parameters (4)

Stopping threshold epsilon_v (volume) = 0.5 cm^3
User-defined precision target; the paper does not derive or justify this value beyond calling it user-defined.
Stopping threshold epsilon_LVEF = 15%
User-defined precision target; same unspecified origin.
Conformal error rate alpha = 0.1
Chosen to target 90% coverage; reasonable but not derived or varied.
Number of Monte Carlo samples M = 20 (400 for LVEF)
Used for uncertainty estimation; no sensitivity analysis is provided.

assumptions (5)

standard math Split conformal prediction yields valid marginal coverage under exchangeability
Invoked in Section 2.3; the paper relies on this theorem for its formal guarantee.
domain assumption Test data are exchangeable with calibration data
Required by the conformal guarantee; the adaptive stopping selection likely violates this, as shown by the coverage gap.
domain assumption PHiRec samples approximate the true posterior p(x|y)
The uncertainty signal depends on PHiRec's posterior fidelity; validated only indirectly via final metric error.
domain assumption Segmentation network is accurate and deterministic
Metrics are computed from segmentation masks; errors in segmentation propagate directly to the metric.
standard math Forward model with known coil sensitivities holds
Equation (1), standard in MRI reconstruction, is assumed throughout.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CUTE-MRI: Conformalized Uncertainty-based framework for Time-adaptivE MRI." pith.science (2026). https://pith.science/paper/TTTCTF4U

@misc{pith2026250814952,
  author       = {Pith},
  title        = {Pith review of: CUTE-MRI: Conformalized Uncertainty-based framework for Time-adaptivE MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTTCTF4U}},
  note         = {Machine review of arXiv:2508.14952}
}

read the original abstract

Magnetic Resonance Imaging (MRI) offers unparalleled soft-tissue contrast but is fundamentally limited by long acquisition times. While deep learning-based accelerated MRI can dramatically shorten scan times, the reconstruction from undersampled data introduces ambiguity resulting from an ill-posed problem with infinitely many possible solutions that propagates to downstream clinical tasks. This uncertainty is usually ignored during the acquisition process as acceleration factors are often fixed a priori, resulting in scans that are either unnecessarily long or of insufficient quality for a given clinical endpoint. This work introduces a dynamic, uncertainty-aware acquisition framework that adjusts scan time on a per-subject basis. Our method leverages a probabilistic reconstruction model to estimate image uncertainty, which is then propagated through a full analysis pipeline to a quantitative metric of interest (e.g., patellar cartilage volume or cardiac ejection fraction). We use conformal prediction to transform this uncertainty into a rigorous, calibrated confidence interval for the metric. During acquisition, the system iteratively samples k-space, updates the reconstruction, and evaluates the confidence interval. The scan terminates automatically once the uncertainty meets a user-predefined precision target. We validate our framework on both knee and cardiac MRI datasets. Our results demonstrate that this adaptive approach reduces scan times compared to fixed protocols while providing formal statistical guarantees on the precision of the final image. This framework moves beyond fixed acceleration factors, enabling patient-specific acquisitions that balance scan efficiency with diagnostic confidence, a critical step towards personalized and resource-efficient MRI.

Figures

Figures reproduced from arXiv: 2508.14952 by the authors.

**Figure 2.** Quantitative evaluation of reconstruction and segmentation performance across [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** Performance of uncertainty-guided early stopping with and without calibration. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Patellar cartilage volume estimations along with calibrated uncertainty bounds [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

**Figure 5.** Figure 5: LVEF estimates along with calibrated uncertainty bounds and examples of [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6: Example reconstructions for the SKM-TEA dataset. The top row shows the [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: Example reconstructions for the CINE dataset. The top row shows the un [PITH_FULL_IMAGE:figures/full_fig_p020_7.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.

Pixelwise Uncertainty Quantification of Accelerated MRI Reconstruction
eess.IV 2026-01 conditional novelty 5.0 of 10

Conformalized quantile regression gives pixel-wise uncertainty intervals for accelerated MRI reconstructions that closely track true reconstruction error, with brain test correlations near or above 0.90 at 4–10x acceleration.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages · cited by 1 Pith paper

[1]

Westbrook, J

C. Westbrook, J. Talbot, MRI in Practice, John Wiley & Sons, 2018

work page 2018
[2]

J. B. Andre, B. W. Bresnahan, M. Mossa-Basha, M. N. Hoff, C. P. Smith, Y. Anzai, W. A. Cohen, Toward quantifying the prevalence, severity, and cost associated with patient motion during clinical mr ex- aminations, Journal of the American College of Radiology 12 (7) (2015) 689–695. 22

work page 2015
[3]

Lustig, D

M. Lustig, D. Donoho, J. M. Pauly, Sparse MRI: The application of com- pressed sensing for rapid MR imaging, Magnetic Resonance in Medicine 58 (6) (2007) 1182–1195. doi:10.1002/mrm.21391. URL https://onlinelibrary.wiley.com/doi/10.1002/mrm.21391

doi:10.1002/mrm.21391 2007
[4]

Knoll, K

F. Knoll, K. Hammernik, C. Zhang, S. Moeller, T. Pock, D. K. Sod- ickson, M. Akcakaya, Deep-learning methods for parallel magnetic res- onance imaging reconstruction: A survey of the current approaches, trends, and issues, IEEE signal processing magazine 37 (1) (2020) 128– 140

work page 2020
[5]

O. N. Jaspan, R. Fleysher, M. L. Lipton, Compressed sensing mri: a re- view of the clinical literature, The British journal of radiology 88 (1056) (2015) 20150487

work page 2015
[6]

K. P. Pruessmann, M. Weiger, M. B. Scheidegger, P. Boesiger, Sense: sensitivity encoding for fast mri, Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42 (5) (1999) 952–962

work page 1999
[7]

M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang, B. Kiefer, A. Haase, Generalized autocalibrating partially par- allel acquisitions (grappa), Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47 (6) (2002) 1202–1210

work page 2002
[8]

Hammernik, T

K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, F. Knoll, Learning a variational network for reconstruction of accelerated mri data, Magnetic resonance in medicine 79 (6) (2018) 3055–3071

work page 2018

Show all 41 references

[9]

Schlemper, J

J. Schlemper, J. Caballero, J. V. Hajnal, A. N. Price, D. Rueckert, A deep cascade of convolutional neural networks for dynamic mr image reconstruction, IEEE transactions on Medical Imaging 37 (2) (2017) 491–503

2017
[10]

Hammernik, T

K. Hammernik, T. K¨ ustner, B. Yaman, Z. Huang, D. Rueckert, F. Knoll, M. Ak¸ cakaya, Physics-driven deep learning for computational magnetic resonance imaging: Combining physics and machine learning for im- proved medical imaging, IEEE Signal Processing Magazine 40 (1) (2023)...

2023
[11]

Heckel, M

R. Heckel, M. Jacob, A. Chaudhari, O. Perlman, E. Shimron, Deep learning for accelerated and robust mri reconstruction, Magnetic Reso- nance Materials in Physics, Biology and Medicine 37 (3) (2024) 335–368

2024
[12]

Maynez, S

J. Maynez, S. Narayan, B. Bohnet, R. McDonald, On faithful- ness and factuality in abstractive summarization, arXiv preprint arXiv:2005.00661 (2020)

2005 arXiv
[13]

S. K. Aithal, P. Maini, Z. C. Lipton, J. Z. Kolter, Understanding hallucinations in diffusion models through mode interpolation (2024). arXiv:2406.09358. URL https://arxiv.org/abs/2406.09358

2024 arXiv
[14]

Antun, F

V. Antun, F. Renna, C. Poon, B. Adcock, A. C. Hansen, On instabilities of deep learning in image reconstruction and the potential costs of ai, Proceedings of the National Academy of Sciences 117 (48) (2020) 30088– 30095. arXiv:https://www.pnas.org/doi/pdf/10.1073/pnas.19073771...

2020 doi
[15]

Fischer, K

P. Fischer, K. Thomas, C. F. Baumgartner, Uncertainty estimation and propagation in accelerated mri reconstruction, in: International Work- shop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, Springer, 2023, pp. 84–94

2023
[16]

A. M. Wundram, P. Fischer, S. Wunderlich, H. Faber, L. M. Koch, P. Berens, C. F. Baumgartner, Leveraging probabilistic segmentation models for improved glaucoma diagnosis: A clinical pipeline approach, in: Medical Imaging with Deep Learning, 2024

2024
[17]

Y. Gal, Z. Ghahramani, Dropout as a bayesian approximation: Repre- senting model uncertainty in deep learning, in: international conference on machine learning, PMLR, 2016, pp. 1050–1059

2016
[18]

Lakshminarayanan, A

B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable pre- dictive uncertainty estimation using deep ensembles, Advances in neural information processing systems 30 (2017)

2017
[19]

C. F. Baumgartner, K. C. Tezcan, K. Chaitanya, A. M. H¨ otker, U. J. Muehlematter, K. Schawkat, A. S. Becker, O. Donati, E. Konukoglu, 24 Phiseg: Capturing uncertainty in medical image segmentation (2019). doi:10.48550/ARXIV.1906.04045. URL https://arxiv.org/abs/1906.04045

2019 doi arXiv
[20]

Edupuganti, M

V. Edupuganti, M. Mardani, S. Vasanawala, J. Pauly, Uncertainty quan- tification in deep mri reconstruction, IEEE Transactions on Medical Imaging 40 (1) (2020) 239–250

2020
[21]

Narnhofer, A

D. Narnhofer, A. Effland, E. Kobler, K. Hammernik, F. Knoll, T. Pock, Bayesian uncertainty estimation of learned variational mri reconstruc- tion, IEEE transactions on medical imaging 41 (2) (2021) 279–291

2021
[22]

N. R. Huttinga, T. Bruijnen, C. A. van den Berg, A. Sbrizzi, Gaussian processes for real-time 3d motion and uncertainty estimation during mr- guided radiotherapy, Medical Image Analysis 88 (2023) 102843

2023
[23]

Schlemper, D

J. Schlemper, D. C. Castro, W. Bai, C. Qin, O. Oktay, J. Duan, A. N. Price, J. Hajnal, D. Rueckert, Bayesian deep learning for accelerated mr image reconstruction, in: International workshop on machine learning for medical image reconstruction, Springer, 2018, pp. 64–71

2018
[24]

J. N. Morshuis, M. Hein, C. F. Baumgartner, Segmentation-guided mri reconstruction for meaningfully diverse reconstructions, in: MICCAI Workshop on Deep Generative Models, Springer, 2024, pp. 180–190

2024
[25]

J. N. Morshuis, C. Schlarmann, T. K¨ ustner, C. F. Baumgartner, M. Hein, Mind the detail: Uncovering clinically relevant image de- tails in accelerated mri with semantically diverse reconstructions (2025). arXiv:2507.00670. URL https://arxiv.org/abs/2507.00670

2025 arXiv
[26]

A. M. Wundram, P. Fischer, M. M¨ uhlebach, L. M. Koch, C. F. Baum- gartner, Conformal performance range prediction for segmentation out- put quality control, in: International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, Springer, 2024, ...

2024
[27]

Daud´ e, R

P. Daud´ e, R. Ramasawmy, A. Javed, R. J. Lederman, K. Chow, A. E. Campbell-Washburn, Inline automatic quality control of 2d phase- contrast flow mri for subject-specific scan time adaptation, Magnetic Resonance in Medicine 92 (2) (2024) 751–760. 25

2024
[28]

Pineda, S

L. Pineda, S. Basu, A. Romero, R. Calandra, M. Drozdzal, Active mr k-space sampling with reinforcement learning, in: International Confer- ence on Medical Image Computing and Computer-Assisted Intervention, Springer, 2020, pp. 23–33

2020
[29]

Z. Wang, B. Li, H. Yu, Z. Zhang, M. Ran, W. Xia, Z. Yang, J. Lu, H. Chen, J. Zhou, et al., Promoting fast mr imaging pipeline by full- stack ai, Iscience 27 (1) (2024)

2024
[30]

Huang, J

Z. Huang, J. Duan, Y. Xie, Y. Liu, Udnet: Unified deep network based on transformer and multi-stage fusion for brain tumor classification from undersampled mri, Neurocomputing 619 (2025) 129109

2025
[31]

Z. Wu, T. Yin, Y. Sun, R. Frost, A. van der Kouwe, A. V. Dalca, K. L. Bouman, Learning task-specific strategies for accelerated mri, IEEE Transactions on Computational Imaging (2024)

2024
[32]

Ronneberger, P

O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Spri...

2015
[33]

V. Vovk, A. Gammerman, G. Shafer, Algorithmic learning in a random world, Vol. 29, Springer, 2005

2005
[34]

A. N. Angelopoulos, S. Bates, A gentle introduction to conformal pre- diction and distribution-free uncertainty quantification, arXiv preprint arXiv:2107.07511 (2021)

2021 arXiv
[35]

A. D. Desai, A. M. Schmidt, E. B. Rubin, C. M. Sandino, M. S. Black, V. Mazzoli, K. J. Stevens, R. Boutin, C. R´ e, G. E. Gold, et al., Skm-tea: A dataset for accelerated mri reconstruction with dense image labels for quantitative clinical evaluation, arXiv preprint arXiv:2203...

2022 arXiv
[36]

Ahmad, H

R. Ahmad, H. Xue, S. Giri, Y. Ding, J. Craft, O. P. Simonetti, Variable density incoherent spatiotemporal acquisition (vista) for highly acceler- ated cardiac mri, Magnetic resonance in medicine 74 (5) (2015) 1266– 1278. 26

2015
[37]

D. P. Kingma, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[38]

Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE transac- tions on image processing 13 (4) (2004) 600–612

2004
[39]

F. K. Ciliberti, G. Cesarelli, L. Guerrini, A. E. Gunnarsson, R. Forni, R. Aubonnet, M. Recenti, D. Jacob, J. H. J´ onsson, V. Cangiano, et al., The role of bone mineral density and cartilage volume to predict knee cartilage degeneration, European Journal of Translational Myol...

2022
[40]

T. Wang, X. Xu, J. Xiong, Q. Jia, H. Yuan, M. Huang, J. Zhuang, Y. Shi, Ica-unet: Ica inspired statistical unet for real-time 3d cardiac cine mri segmentation, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru...

2020
[41]

A. N. Angelopoulos, A. P. Kohli, S. Bates, M. Jordan, J. Malik, T. Al- shaabi, S. Upadhyayula, Y. Romano, Image-to-image regression with distribution-free uncertainty quantification and applications in imaging, in: International Conference on Machine Learning, PMLR, 2022, pp. ...

2022

Pith tools

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

[1] [1]

Westbrook, J

C. Westbrook, J. Talbot, MRI in Practice, John Wiley & Sons, 2018

work page 2018

[2] [2]

J. B. Andre, B. W. Bresnahan, M. Mossa-Basha, M. N. Hoff, C. P. Smith, Y. Anzai, W. A. Cohen, Toward quantifying the prevalence, severity, and cost associated with patient motion during clinical mr ex- aminations, Journal of the American College of Radiology 12 (7) (2015) 689–695. 22

work page 2015

[3] [3]

Lustig, D

M. Lustig, D. Donoho, J. M. Pauly, Sparse MRI: The application of com- pressed sensing for rapid MR imaging, Magnetic Resonance in Medicine 58 (6) (2007) 1182–1195. doi:10.1002/mrm.21391. URL https://onlinelibrary.wiley.com/doi/10.1002/mrm.21391

doi:10.1002/mrm.21391 2007

[4] [4]

Knoll, K

F. Knoll, K. Hammernik, C. Zhang, S. Moeller, T. Pock, D. K. Sod- ickson, M. Akcakaya, Deep-learning methods for parallel magnetic res- onance imaging reconstruction: A survey of the current approaches, trends, and issues, IEEE signal processing magazine 37 (1) (2020) 128– 140

work page 2020

[5] [5]

O. N. Jaspan, R. Fleysher, M. L. Lipton, Compressed sensing mri: a re- view of the clinical literature, The British journal of radiology 88 (1056) (2015) 20150487

work page 2015

[6] [6]

K. P. Pruessmann, M. Weiger, M. B. Scheidegger, P. Boesiger, Sense: sensitivity encoding for fast mri, Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42 (5) (1999) 952–962

work page 1999

[7] [7]

M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang, B. Kiefer, A. Haase, Generalized autocalibrating partially par- allel acquisitions (grappa), Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47 (6) (2002) 1202–1210

work page 2002

[8] [8]

Hammernik, T

K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, F. Knoll, Learning a variational network for reconstruction of accelerated mri data, Magnetic resonance in medicine 79 (6) (2018) 3055–3071

work page 2018