pith. machine review for the scientific record. sign in

arxiv: 2604.08781 · v1 · submitted 2026-04-09 · 📡 eess.IV · cs.AI· cs.CV· eess.SP· physics.med-ph

Recognition: unknown

PSIRNet: Deep Learning-based Free-breathing Rapid Acquisition Late Enhancement Imaging

Arda Atalik , Hui Xue , Rhodri H. Davies , Thomas A. Treibel , Daniel K. Sodickson , Michael S. Hansen , Peter Kellman
Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:48 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CVeess.SPphysics.med-ph
keywords deep learningcardiac MRIlate gadolinium enhancementPSIRfree-breathingimage reconstructionmotion correctionneural network
0
0 comments X

The pith

A deep learning network produces diagnostic-quality cardiac MRI images from a single two-heartbeat acquisition.

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

The paper develops PSIRNet, a deep learning method to reconstruct phase-sensitive inversion recovery late gadolinium enhancement images in cardiac MRI from raw data collected in one acquisition spanning two heartbeats. It seeks to replace the standard approach that relies on 8 to 24 motion-corrected signal averages. A sympathetic reader would care because the change could shorten scan duration substantially while keeping image quality at a level experts judge diagnostic. The network was trained on a large split of retrospective data from many patients, sites, and both 1.5T and 3T scanners, then tested with quantitative metrics and blinded ratings by two cardiologists.

Core claim

PSIRNet is a physics-guided deep learning network with 845 million parameters that reconstructs surface-coil-corrected PSIR images from a single interleaved IR/PD acquisition over two heartbeats. On held-out data from different institutions, single-average PSIRNet reconstructions received expert Likert scores superior to or equivalent with the reference MOCO PSIR images built from many averages, while taking roughly 100 ms per slice to compute versus more than 5 seconds.

What carries the argument

PSIRNet, a physics-guided deep learning network trained end-to-end to map single-acquisition k-space data to PSIR LGE images with surface coil correction.

If this is right

  • Reduces acquisition time for free-breathing PSIR LGE by a factor of 8 to 24.
  • Produces images at approximately 100 ms per slice instead of more than 5 seconds.
  • Delivers quality rated superior for dark blood LGE and equivalent or superior for bright blood and wideband variants.
  • Operates on data from both 1.5T and 3T scanners across multiple institutions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar networks could shorten other cardiac MRI protocols that currently depend on extensive averaging for motion robustness.
  • Clinical deployment might raise daily patient throughput in cardiac imaging suites.
  • The multi-year multi-site training set suggests robustness, yet performance on post-2024 scanner hardware would need separate checks.
  • The approach could support protocols for patients unable to perform repeated breath-holds.

Load-bearing premise

A model trained on retrospective multi-site data from 2016-2024 will continue to match or exceed standard image quality on new patients and scanners without retraining.

What would settle it

A prospective study collecting paired single-average and multi-average data on previously unseen patients and scanners, then showing expert ratings of PSIRNet images fall below equivalence margin on the 5-point Likert scale.

Figures

Figures reproduced from arXiv: 2604.08781 by Arda Atalik, Daniel K. Sodickson, Hui Xue, Michael S. Hansen, Peter Kellman, Rhodri H. Davies, Thomas A. Treibel.

Figure 1
Figure 1. Figure 1: PSIRNet reconstruction. (a) Overview of PSIRNet consisting of sensitivity map refinement (SMR), N cascade elements connected in series, and the final PSCC block calculating the PSIR image with surface coil correction (SCC). (b) Components of each element in the cascade—IR data consistency (DC), PD data consistency (DC), and joint refinement (Φ)—are shown along with their inputs, their outputs, and the fina… view at source ↗
Figure 2
Figure 2. Figure 2: Short-axis bright blood reconstructions. At 8× acceleration, PSIRNet maintains image quality on par with the reference standard MOCO PSIR and preserves the conspicuity of the hyperenhanced myocardial infarction. 2,615 of 4,504 (58.1%) men, 1,868 of 4,504 (41.5%) women, and 21 of 4,504 (0.5%) patients with unspecified sex. The median patient age was 56.0 years (interquartile range [IQR], 27.0 years). The me… view at source ↗
Figure 3
Figure 3. Figure 3: Short-axis dark blood reconstructions. PSIRNet delivers 16× acceleration with robust depiction of hyperenhanced myocardial scar and borders against the suppressed blood pool, achieving image quality comparable to or better than the reference standard MOCO PSIR. MOCO PSIR reference ( [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Wideband LGE reconstructions in vertical long-axis, short-axis, and four-chamber views. Wideband LGE is used in patients with implanted cardiac devices, where extended acquisitions requiring up to 48 heartbeats are necessary to compensate for inherently low SNR. PSIRNet achieves image quality comparable to or better than the 24-heartbeat MOCO PSIR reference from a 2-heartbeat acquisition (12× acceleration)… view at source ↗
Figure 5
Figure 5. Figure 5: Joint distributions of PSIRNet versus MOCO PSIR Likert scores representing patient-level means across scored stacks for bright blood (left), dark blood (center), and wideband (right) LGE, shown for Reader 1 (top) and Reader 2 (bottom). Each cell represents a unique combination of PSIRNet and MOCO PSIR scores; color intensity indicates the count of paired ratings at that combination. Cells above the diagona… view at source ↗
read the original abstract

Purpose: To develop and evaluate a deep learning (DL) method for free-breathing phase-sensitive inversion recovery (PSIR) late gadolinium enhancement (LGE) cardiac MRI that produces diagnostic-quality images from a single acquisition over two heartbeats, eliminating the need for 8 to 24 motion-corrected (MOCO) signal averages. Materials and Methods: Raw data comprising 800,653 slices from 55,917 patients, acquired on 1.5T and 3T scanners across multiple sites from 2016 to 2024, were used in this retrospective study. Data were split by patient: 640,000 slices (42,822 patients) for training and the remainder for validation and testing, without overlap. The training and testing data were from different institutions. PSIRNet, a physics-guided DL network with 845 million parameters, was trained end-to-end to reconstruct PSIR images with surface coil correction from a single interleaved IR/PD acquisition over two heartbeats. Reconstruction quality was evaluated using SSIM, PSNR, and NRMSE against MOCO PSIR references. Two expert cardiologists performed an independent qualitative assessment, scoring image quality on a 5-point Likert scale across bright blood, dark blood, and wideband LGE variants. Paired superiority and equivalence (margin = 0.25 Likert points) were tested using exact Wilcoxon signed-rank tests at a significance level of 0.05 using R version 4.5.2. Results: Both readers rated single-average PSIRNet reconstructions superior to MOCO PSIR for dark blood LGE (conservative P = .002); for bright blood and wideband, one reader rated it superior and the other confirmed equivalence (all P < .001). Inference required approximately 100 msec per slice versus more than 5 sec for MOCO PSIR. Conclusion: PSIRNet produces diagnostic-quality free-breathing PSIR LGE images from a single acquisition, enabling 8- to 24-fold reduction in acquisition time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims to develop PSIRNet, a physics-guided deep learning network with 845 million parameters, that reconstructs diagnostic-quality phase-sensitive inversion recovery (PSIR) late gadolinium enhancement (LGE) cardiac MRI images from a single free-breathing interleaved IR/PD acquisition over two heartbeats. Trained end-to-end on a large retrospective multi-site dataset of 800,653 slices from 55,917 patients (2016-2024), with patient-wise splits and different institutions for train/test, it is evaluated against MOCO-averaged references using SSIM, PSNR, NRMSE and independent expert 5-point Likert scoring with Wilcoxon signed-rank tests for superiority and equivalence (margin 0.25), showing superiority or equivalence while reducing acquisition time by 8- to 24-fold.

Significance. If the retrospective performance generalizes, the work has high clinical significance by enabling much shorter free-breathing PSIR LGE scans without loss of diagnostic quality, which would improve throughput and patient tolerance in cardiac MRI. The large-scale multi-institutional patient-split dataset, quantitative metrics, and statistically tested expert ratings provide solid support for the performance claim on the evaluated distribution.

major comments (1)
  1. [Materials and Methods] Materials and Methods: The training and test data are drawn from the same 2016-2024 multi-institutional retrospective distribution (different institutions but no temporal hold-out after 2024 or scanners absent from training). With an 845-million-parameter model, this leaves untested the generalization to prospective acquisitions on new patients or hardware, which is load-bearing for the Conclusion claim that PSIRNet enables routine 8- to 24-fold acquisition-time reduction in clinical practice.
minor comments (2)
  1. [Results] Results: The Likert-scale superiority/equivalence findings are reported with p-values but without explicit mean/median values or confidence intervals per reader and variant; adding these would strengthen interpretability of the clinical equivalence margin.
  2. The inference-time comparison (~100 ms vs >5 s) is useful but lacks hardware specifications or details on whether MOCO time includes all post-processing steps.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and recommendation for minor revision. We address the single major comment below.

read point-by-point responses

  1. Referee: [Materials and Methods] Materials and Methods: The training and test data are drawn from the same 2016-2024 multi-institutional retrospective distribution (different institutions but no temporal hold-out after 2024 or scanners absent from training). With an 845-million-parameter model, this leaves untested the generalization to prospective acquisitions on new patients or hardware, which is load-bearing for the Conclusion claim that PSIRNet enables routine 8- to 24-fold acquisition-time reduction in clinical practice.

    Authors: We agree that the absence of prospective data on new patients and hardware constitutes a genuine limitation for claims of routine clinical deployment. The study is explicitly retrospective, with patient-wise splits and training/testing data drawn from different institutions to reduce data leakage. The dataset spans eight years, multiple sites, and both 1.5 T and 3 T field strengths, providing substantial diversity within the observed distribution. We will add an explicit limitations paragraph in the Discussion acknowledging the retrospective design and the requirement for prospective validation on unseen hardware and acquisition conditions. We will also revise the Conclusion to state that the method demonstrates diagnostic-quality reconstruction with the potential to enable 8- to 24-fold acceleration, subject to confirmation in prospective studies. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents an empirical deep-learning reconstruction method trained end-to-end on a large retrospective multi-site dataset (patient-wise split, different institutions for train/test) to map single-acquisition raw data to PSIR images. Quality is assessed on held-out test data via SSIM/PSNR/NRMSE against independent MOCO references and expert Likert scoring with statistical tests. No equations, ansatzes, or derivations are shown that reduce the claimed output to a fitted parameter, self-referential quantity, or self-citation chain by construction. The central result is an empirical performance comparison on independent data rather than any self-definitional or fitted-input-called-prediction pattern.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on supervised learning from paired single-shot and multi-average data; the network weights are learned from the training set rather than derived from first principles.

free parameters (1)
  • 845 million network parameters
    Learned during end-to-end training on the 640,000-slice dataset to map raw interleaved data to PSIR images.
axioms (1)
  • domain assumption A physics-guided convolutional network can learn an accurate mapping from single-acquisition raw k-space to motion-corrected PSIR images
    Invoked in the training and inference description without further proof.

pith-pipeline@v0.9.0 · 5716 in / 1153 out tokens · 48490 ms · 2026-05-10T16:48:46.497015+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages

  1. [1]

    Cardiac imaging techniques for physicians: late enhancement

    Kellman P, Arai AE. Cardiac imaging techniques for physicians: late enhancement. J Magn Reson Imaging. 2012;36(3):529-542. doi:10.1002/jmri.23605

    work page doi:10.1002/jmri.23605 2012

  2. [2]

    Phase-sensitive inversion recovery for detecting myocardial infarction using gadolinium-delayed hyperenhancement

    Kellman P, Arai AE, McVeigh ER, Aletras AH. Phase-sensitive inversion recovery for detecting myocardial infarction using gadolinium-delayed hyperenhancement. Magn Reson Med. 2002;47(2):372-383. doi:10.1002/mrm.10051

    work page doi:10.1002/mrm.10051 2002

  3. [3]

    Motion corrected free-breathing delayed-enhancement imaging of myocardial infarction using nonrigid registration

    Ledesma-Carbayo MJ, Kellman P, Hsu LY, Arai AE, McVeigh ER. Motion corrected free-breathing delayed-enhancement imaging of myocardial infarction using nonrigid registration. J Magn Reson Imaging. 2007;26(1):184-190. doi:10.1002/jmri.20957

    work page doi:10.1002/jmri.20957 2007

  4. [4]

    Free-breathing late enhancement imaging: phase sensitive inversion recovery (PSIR) with respiratory motion corrected (MOCO) averaging

    Kellman P, Xue H, Hansen MS. Free-breathing late enhancement imaging: phase sensitive inversion recovery (PSIR) with respiratory motion corrected (MOCO) averaging. MAGNETOM Flash. 2016;66:65–73

    2016

  5. [5]

    Deep complex convolutional network for fast reconstruction of 3D late gadolinium enhancement cardiac MRI

    El-Rewaidy H, Neisius U, Mancio J, et al. Deep complex convolutional network for fast reconstruction of 3D late gadolinium enhancement cardiac MRI. NMR Biomed. 2020;33(7):e4312. doi:10.1002/nbm.4312

    work page doi:10.1002/nbm.4312 2020

  6. [6]

    Improvement of late gadolinium enhancement image quality using a deep learning-based reconstruction algorithm and its influence on myocardial scar quantification

    van der Velde N, Hassing HC, Bakker BJ, et al. Improvement of late gadolinium enhancement image quality using a deep learning-based reconstruction algorithm and its influence on myocardial scar quantification. Eur Radiol. 2021;31(6):3846-3855. doi:10.1007/s00330-020-07461-w

    work page doi:10.1007/s00330-020-07461-w 2021

  7. [7]

    Muscogiuri G, Martini C, Gatti M, et al. Feasibility of late gadolinium enhancement (LGE) in ischemic cardiomyopathy using 2D-multisegment LGE combined with artificial intelligence reconstruction deep learning noise reduction algorithm. Int J Cardiol. 2021;343:164-170. doi:10.1016/j.ijcard.2021.09.012

    work page doi:10.1016/j.ijcard.2021.09.012 2021

  8. [8]

    URL https://arxiv.org/abs/2102.09542

    Yaman B, Shenoy C, Deng Z, Moeller S, El-Rewaidy H, Nezafat R, Akçakaya M. Self-supervised physics-guided deep learning reconstruction for high-resolution 3D LGE CMR. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). IEEE; 2021:100–104. doi:10.1109/ISBI48211.2021.9434054

    work page doi:10.1109/isbi48211.2021.9434054 2021

  9. [9]

    Late gadolinium enhancement cardiovascular magnetic resonance with generative artificial intelligence

    Demirel OB, Ghanbari F, Hoeger CW, et al. Late gadolinium enhancement cardiovascular magnetic resonance with generative artificial intelligence. J Cardiovasc Magn Reson. 2025;27(1):101127. doi:10.1016/j.jocmr.2024.101127

    work page doi:10.1016/j.jocmr.2024.101127 2025

  10. [10]

    Dark blood late enhancement imaging

    Kellman P, Xue H, Olivieri LJ, et al. Dark blood late enhancement imaging. J Cardiovasc Magn Reson. 2016;18(1):77. Published 2016 Nov 7. doi:10.1186/s12968-016-0297-3

    work page doi:10.1186/s12968-016-0297-3 2016

  11. [11]

    Clinical impact of cardiovascular magnetic resonance with optimized myocardial scar detection in patients with cardiac implantable devices

    Bhuva AN, Kellman P, Graham A, et al. Clinical impact of cardiovascular magnetic resonance with optimized myocardial scar detection in patients with cardiac implantable devices. Int J Cardiol. 2019;279:72-78. doi:10.1016/j.ijcard.2019.01.005

    work page doi:10.1016/j.ijcard.2019.01.005 2019

  12. [12]

    Tyger v0.13.5

    Microsoft. Tyger v0.13.5. Updated September 25, 2025. Accessed March 7, 2026. https://github.com/microsoft/tyger/releases/tag/v0.13.5

    2025

  13. [13]

    Gadgetron: an open source framework for medical image reconstruction

    Hansen MS, Sørensen TS. Gadgetron: an open source framework for medical image reconstruction. Magn Reson Med. 2013;69(6):1768-1776. doi:10.1002/mrm.24389

    work page doi:10.1002/mrm.24389 2013

  14. [14]

    An iteration formula for Fredholm integral equations of the first kind

    Landweber L. An iteration formula for Fredholm integral equations of the first kind. Am J Math. 1951;73(3):615-624

    1951

  15. [15]

    Learning a variational network for reconstruction of accelerated MRI data

    Hammernik K, Klatzer T, Kobler E, Recht MP, Sodickson DK, Pock T, Knoll F. Learning a variational network for reconstruction of accelerated MRI data. Magn Reson Med. 2018;79(6):3055- 3071

    2018

  16. [16]

    End-to-end variational networks for accelerated MRI reconstruction

    Sriram A, Zbontar J, Murrell T, et al. End-to-end variational networks for accelerated MRI reconstruction. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2020. Springer; 2020:64-73. doi:10.1007/978-3-030-59713-9_7

    work page doi:10.1007/978-3-030-59713-9_7 2020

  17. [17]

    In: Medical Image Compu ting and Computer-Assisted Intervention – MICCAI 2015

    Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Springer; 2015:234-241. doi:10.1007/978-3-319-24574-4_28

    work page doi:10.1007/978-3-319-24574-4_28 2015

  18. [18]

    Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis

    Ulyanov D, Vedaldi A, Lempitsky V. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. IEEE; 2017:6924-6932. doi:10.1109/CVPR.2017.437

    work page doi:10.1109/cvpr.2017.437 2017

  19. [19]

    Rectifier nonlinearities improve neural network acoustic models

    Maas AL, Hannun AY, Ng AY. Rectifier nonlinearities improve neural network acoustic models. In: Proceedings of the 30th International Conference on Machine Learning. 2013;30(1):3

    2013

  20. [20]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process. 2004;13(4):600-612. doi:10.1109/TIP.2003.819861

    work page doi:10.1109/tip.2003.819861 2004

  21. [21]

    PyTorch: an imperative style, high-performance deep learning library

    Paszke A, Gross S, Massa F, et al. PyTorch: an imperative style, high-performance deep learning library. Adv Neural Inf Process Syst. 2019;32:8024-8035

    2019

  22. [22]

    Decoupled weight decay regularization

    Loshchilov I, Hutter F. Decoupled weight decay regularization. In: Proceedings of the 7th International Conference on Learning Representations. 2019

    2019

  23. [23]

    Confidence intervals based on bootstrap percentiles

    Efron B, Tibshirani RJ. Confidence intervals based on bootstrap percentiles. In: An Introduction to the Bootstrap. Chapman and Hall/CRC; 1994:168-177

    1994

  24. [24]

    Biometrics Bulletin , author =

    Wilcoxon F. Individual comparisons by ranking methods. Biometrics Bull. 1945;1(6):80-83. doi:10.2307/3001968

    work page doi:10.2307/3001968 1945

  25. [25]

    A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability

    Schuirmann DJ. A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability. J Pharmacokinet Biopharm. 1987;15(6):657-

    1987

  26. [26]

    doi:10.1007/BF01068419

    work page doi:10.1007/bf01068419

  27. [27]

    Remarks on zeros and ties in the Wilcoxon signed rank procedures

    Pratt JW. Remarks on zeros and ties in the Wilcoxon signed rank procedures. J Am Stat Assoc. 1959;54(287):655-667. doi:10.1080/01621459.1959.10501526

    work page doi:10.1080/01621459.1959.10501526 1959

  28. [28]

    A simple sequentially rejective multiple test procedure

    Holm S. A simple sequentially rejective multiple test procedure. Scand J Stat. 1979;6(2):65-70. doi:10.2307/4615733

    work page doi:10.2307/4615733 1979

  29. [29]

    R: A Language and Environment for Statistical Computing

    R Core Team. R: A Language and Environment for Statistical Computing. Version 4.5.2. R Foundation for Statistical Computing. Updated October 31, 2025. Accessed March 7, 2026. https://www.r-project.org

    2025

  30. [30]

    Statistical Hypothesis Testing in Context: Reproducibility, Inference, and Science

    Fay MP, Brittain EH. Statistical Hypothesis Testing in Context: Reproducibility, Inference, and Science. Cambridge University Press; 2022

    2022