pith. sign in

arxiv: 2601.06166 · v2 · submitted 2026-01-07 · 💻 cs.CV

B-FIRE: Binning-Free Diffusion Implicit Neural Representation for Hyper-Accelerated Motion-Resolved MRI

Di Xu , Hengjie Liu , Yang Yang , Mary Feng , Jin Ning , Xin Miao , Jessica E. Scholey , Alexandra E. Hotca-cho
show 6 more authors
William C. Chen Michael Ohliger Martina Descovich Huiming Dong Wensha Yang Ke Sheng
This is my paper

Pith reviewed 2026-05-16 17:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords B-FIREdiffusion implicit neural representationhyper-accelerated MRIbinning-free reconstructionmotion-resolved 4DMRInon-Cartesian k-spaceabdominal imaginginstantaneous anatomy
0
0 comments X

The pith

B-FIRE reconstructs instantaneous 3D abdominal anatomy directly from single-spoke non-Cartesian k-space data without motion binning.

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

The paper introduces B-FIRE to recover real-time volumetric motion in abdominal MRI from data that is far more sparsely sampled than current techniques allow. Instead of grouping measurements into averaged breathing phases, the method trains on binned pairs but applies the model to raw, un-binned inputs at accelerations reaching one spoke per frame. A reader would care because this removes the blur and loss of instantaneous detail that limit existing 4DMRI for applications needing precise motion trajectories.

Core claim

B-FIRE is a binning-free diffusion implicit neural representation framework for hyper-accelerated MR reconstruction capable of reflecting instantaneous 3D abdominal anatomy from undersampled non-Cartesian k-space data. It employs a CNN-INR encoder-decoder backbone optimized using diffusion with a comprehensive loss that enforces image-domain fidelity and frequency-aware constraints. Motion binned image pairs serve as training references while inference runs on binning-free undersampled data.

What carries the argument

B-FIRE, the binning-free diffusion implicit neural representation using a CNN-INR encoder-decoder backbone that is optimized with diffusion losses to enforce both spatial fidelity and frequency-domain consistency during reconstruction.

If this is right

  • Reconstruction fidelity and motion consistency remain high across acceleration factors from 8 spokes per frame down to 1 spoke per frame on T1-weighted StarVIBE liver data.
  • Inference produces lower latency than direct NuFFT, GRASP-CS, or unrolled CNN baselines while avoiding phase-averaging artifacts.
  • The same trained model supports both binned training references and fully binning-free inference without retraining.
  • Frequency-aware loss terms ensure that reconstructions respect the original k-space measurements even at extreme undersampling.

Where Pith is reading between the lines

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

  • The approach could be tested on cardiac or respiratory-gated sequences where instantaneous phase is clinically more relevant than averaged states.
  • If latency stays low, the framework might support real-time feedback during MRI-guided interventions without requiring breath holds.
  • Extending the loss to enforce temporal smoothness across adjacent frames could further improve trajectory consistency without reintroducing binning.

Load-bearing premise

A model trained exclusively on motion-binned image pairs can generalize to produce accurate instantaneous reconstructions from binning-free, extremely undersampled data while preserving image fidelity and motion trajectory consistency.

What would settle it

Perform reconstruction on RV1 single-spoke data and compare the resulting motion trajectories against simultaneous high-resolution reference scans; large deviations in instantaneous organ positions or increased blurring would falsify the claim.

read the original abstract

Accelerated dynamic volumetric magnetic resonance imaging (4DMRI) is essential for applications relying on motion resolution. Existing 4DMRI produces acceptable artifacts of averaged breathing phases, which can blur and misrepresent instantaneous dynamic information. Recovery of such information requires a new paradigm to reconstruct extremely undersampled non-Cartesian k-space data. We propose B-FIRE, a binning-free diffusion implicit neural representation framework for hyper-accelerated MR reconstruction capable of reflecting instantaneous 3D abdominal anatomy. B-FIRE employs a CNN-INR encoder-decoder backbone optimized using diffusion with a comprehensive loss that enforces image-domain fidelity and frequency-aware constraints. Motion binned image pairs were used as training references, while inference was performed on binning-free undersampled data. Experiments were conducted on a T1-weighted StarVIBE liver MRI cohort, with accelerations ranging from 8 spokes per frame (RV8) to RV1. B-FIRE was compared against direct NuFFT, GRASP-CS, and an unrolled CNN method. Reconstruction fidelity, motion trajectory consistency, and inference latency were evaluated.

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

2 major / 2 minor

Summary. The paper introduces B-FIRE, a binning-free diffusion implicit neural representation framework for hyper-accelerated motion-resolved MRI. It uses a CNN-INR encoder-decoder backbone optimized via diffusion loss with image-domain fidelity and frequency-aware constraints. Training employs motion-binned image pairs as references, while inference targets binning-free, extremely undersampled non-Cartesian k-space data to recover instantaneous 3D abdominal anatomy. Experiments on T1-weighted StarVIBE liver MRI data evaluate performance at accelerations from RV8 to RV1 against baselines including direct NuFFT, GRASP-CS, and an unrolled CNN, focusing on reconstruction fidelity, motion trajectory consistency, and inference latency.

Significance. If the central claims are substantiated, B-FIRE would offer a meaningful advance in 4DMRI by removing respiratory binning requirements and enabling true instantaneous dynamic reconstructions at extreme accelerations. This could benefit applications needing precise sub-bin temporal resolution of abdominal motion. The combination of diffusion models with implicit neural representations for non-Cartesian k-space data represents a technically interesting direction, though its impact depends on demonstrating generalization beyond the binned training distribution.

major comments (2)
  1. [Abstract and Methods] Abstract and Methods: The core claim of binning-free instantaneous reconstruction rests on training exclusively with motion-binned image pairs yet inferring on binning-free data. Binned pairs supply only phase-averaged supervision, and no motion field, phase-continuous simulation, or unpaired instantaneous ground truth is described to bridge the gap. This creates a direct risk that the network converges to averaged outputs satisfying the binned loss while failing to resolve sub-bin anatomical variations, undermining the hyper-accelerated instantaneous reconstruction claim.
  2. [Methods (Training and Loss)] Methods (Training and Loss): The loss term weights and diffusion schedule parameters are free parameters without reported justification, ablation, or sensitivity analysis. It is therefore unclear whether the claimed consistency between training and inference is independent of the binned training distribution or arises from tuning that does not generalize to binning-free inputs.
minor comments (2)
  1. [Abstract] Abstract: No quantitative error metrics, statistical comparisons, or specific numerical results are supplied despite the description of experiments and comparisons to GRASP-CS and unrolled CNN; these should be added for a self-contained summary.
  2. [Experiments] Experiments: The cohort size, exact acquisition parameters, and number of subjects for the T1-weighted StarVIBE liver MRI dataset are not stated, limiting reproducibility assessment.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point-by-point below, providing clarifications and committing to revisions where they strengthen the work without misrepresenting the presented methods or results.

read point-by-point responses

  1. Referee: [Abstract and Methods] Abstract and Methods: The core claim of binning-free instantaneous reconstruction rests on training exclusively with motion-binned image pairs yet inferring on binning-free data. Binned pairs supply only phase-averaged supervision, and no motion field, phase-continuous simulation, or unpaired instantaneous ground truth is described to bridge the gap. This creates a direct risk that the network converges to averaged outputs satisfying the binned loss while failing to resolve sub-bin anatomical variations, undermining the hyper-accelerated instantaneous reconstruction claim.

    Authors: The diffusion-optimized CNN-INR is designed to learn a continuous implicit representation of abdominal anatomy, where the frequency-aware constraints and image-domain fidelity terms encourage recovery of fine temporal details beyond the phase-averaged supervision provided by binned pairs. At inference, the model operates directly on binning-free, hyper-accelerated non-Cartesian k-space without any binning step, and the reported experiments (RV8 to RV1) demonstrate superior motion trajectory consistency and reduced blurring relative to GRASP-CS and unrolled CNN baselines, indicating that sub-bin variations are resolved rather than averaged. We acknowledge that explicit motion fields or unpaired instantaneous ground truth are not used; the generalization arises from the diffusion process modeling the data distribution. In revision we will expand the Methods and Discussion sections to clarify this mechanism and add qualitative examples of instantaneous reconstructions. revision: partial

  2. Referee: [Methods (Training and Loss)] Methods (Training and Loss): The loss term weights and diffusion schedule parameters are free parameters without reported justification, ablation, or sensitivity analysis. It is therefore unclear whether the claimed consistency between training and inference is independent of the binned training distribution or arises from tuning that does not generalize to binning-free inputs.

    Authors: We agree that explicit justification, ablation studies, and sensitivity analysis for the loss weights and diffusion schedule are necessary to demonstrate robustness. These parameters were selected based on preliminary validation on the training cohort to balance fidelity and perceptual quality, but the manuscript does not report the corresponding experiments. In the revised manuscript we will add a dedicated ablation subsection in Methods (and supplementary material) that varies each weight and schedule parameter, quantifies reconstruction metrics on held-out binning-free test cases, and confirms that performance remains stable across reasonable ranges, thereby supporting generalization beyond the binned training distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper trains a CNN-INR encoder-decoder via diffusion loss on motion-binned image pairs as references and performs separate inference on binning-free undersampled non-Cartesian k-space data. This training-inference split is presented as a generalization step rather than a definitional equivalence or fitted parameter renamed as prediction. No equations reduce the instantaneous reconstruction claim to the binned inputs by construction, no load-bearing self-citations are invoked for uniqueness, and external comparisons (NuFFT, GRASP-CS, unrolled CNN) are described. The derivation chain remains self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim depends on the effectiveness of the CNN-INR backbone and the diffusion loss; the ledger records the implicit assumptions and tunable elements required to make the framework function.

free parameters (2)
  • loss term weights
    The comprehensive loss combines image-domain fidelity and frequency-aware constraints; balancing coefficients are not stated and must be chosen or fitted.
  • diffusion schedule parameters
    Diffusion optimization requires noise schedule and step count choices that are not specified in the abstract.
axioms (1)
  • standard math Non-uniform fast Fourier transform accurately maps non-Cartesian k-space to image domain
    Invoked by the comparison to direct NuFFT reconstruction and by the frequency-aware loss term.
invented entities (1)
  • B-FIRE (diffusion implicit neural representation) no independent evidence
    purpose: To enable binning-free reconstruction of instantaneous 3D anatomy from hyper-undersampled data
    Newly introduced framework whose performance is asserted without external independent validation in the abstract.

pith-pipeline@v0.9.0 · 5540 in / 1483 out tokens · 41143 ms · 2026-05-16T17:20:50.832969+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

45 extracted references · 45 canonical work pages · 2 internal anchors

  1. [1]

    Introduction Fully sampled, high -quality magnetic resonance imaging (MRI) necessitates extended acquisition times as a direct consequence of the minimum k -space sampling density dictated by the Nyquist theorem1, the inherently sequential nature of k -space data acquisition, and the pronounced sensitivity of MR signal encoding to physiological motion. Th...

    work page

  2. [2]

    1, with the inference process of DPM shown in Fig

    Materials and Methods 2.1 Conditional Diffusion Probabilistic Modelling Process The architecture of B -FIRE is an end -to-end framework , as illustrated in Fig. 1, with the inference process of DPM shown in Fig. 1(a). Given an under- and fully sampled image pair (𝒙𝑖, 𝒚𝑖), B-FIRE aims to learn a parametric approximation of the data distribution 𝑝(𝒚|𝒙) via ...

    work page

  3. [3]

    Experiments and Results 3.1 T1 StarVIBE Liver Data Cohort The study was approved by the local Institutional Review Board at UCSF (# 14-15452). 225 patients after injecting hepatobiliary contrast (gadoxeric acid; Eovist, Bayer) and 1 healthy volunteer without contrast injection were scanned on a 3T MRI scanner (MAGNETOM Vida, Siemens Healthcare). A prototy...

    work page

  4. [4]

    𝟗𝟔 ± 𝟎. 𝟎𝟖 10 RV5 0.84 ± 0.14 0.031 ± 0.03 30.05 ± 2.34 0.96 ± 0.08 RV3 0.81 ± 0.14 0.034 ± 0.05 29.32 ± 2.53 0.94 ± 0.1 RV2 0.79 ± 0.16 0.035 ± 0.06 28.91 ± 2.67 0.94 ± 0.1 RV1 0.79 ± 0.18 0.036 ± 0.08 28.87 ± 2.72 0.92 ± 0.13 Compressed Sensing RV8 0.52 ± 0.18 0.12 ± 0.06 18.17 ± 4.53 0.45 ± 0.25 31 ± 1.23 RV5 0.47 ± 0.23 0.17 ± 0.09 15.32 ± 4.53 0.41 ±...

    work page

  5. [5]

    The B -FIRE framework combines a CNN encoder and INR decoder w ithin a diffusion -encapsulated paradigm, supported by comprehensive constraints on image and frequency consistency

    Discussion The study presents B -FIRE (Binning -Free diffusion Implicit neural REpresentation), a framework designed for hyper -accelerated, binning -free, and motion -resolved non - Cartesian MRI reconstruction. The B -FIRE framework combines a CNN encoder and INR decoder w ithin a diffusion -encapsulated paradigm, supported by comprehensive constraints ...

    work page

  6. [6]

    It achieves high-fidelity, real-time visualization down to single -spoke sampling

    Conclusion B-FIRE is a binning -free framework for hyper -accelerated non -Cartesian MRI, utilizing a diffusion-optimized CNN–INR backbone to enforce both image and k -space consistency. It achieves high-fidelity, real-time visualization down to single -spoke sampling. Compared to NuFFT, CS, and unrolled CNNs on T1 -weighted liver data, B -FIRE delivers s...

    work page

  7. [7]

    & Sarkovic, V

    Por, E., Van Kooten, M. & Sarkovic, V . Nyquist–Shannon sampling theorem. Leiden Univ. 1, 1–2 (2019)

    work page 2019

  8. [8]

    in Advances in Magnetic Resonance Technology and Applications vol

    Non-cartesian imaging. in Advances in Magnetic Resonance Technology and Applications vol. 6 481–498 (Elsevier, 2022)

    work page 2022

  9. [9]

    L., Hamilton, J

    Wright, K. L., Hamilton, J. I., Griswold, M. A., Gulani, V . & Seiberlich, N. Non‐Cartesian parallel imaging reconstruction. J. Magn. Reson. Imaging 40, 1022–1040 (2014)

    work page 2014

  10. [10]

    Donoho, D. L. Compressed sensing. IEEE Trans. Inf. Theory 52, 1289–1306 (2006)

    work page 2006

  11. [11]

    Xu, D. et al. Accelerated Patient-specific Non-Cartesian Magnetic Resonance Imaging Reconstruction Using Implicit Neural Representations. Int. J. Radiat. Oncol. S036030162506208X (2025) doi:10.1016/j.ijrobp.2025.08.059

    work page doi:10.1016/j.ijrobp.2025.08.059 2025

  12. [12]

    V ., Price, A

    Schlemper, J., Caballero, J., Hajnal, J. V ., Price, A. N. & Rueckert, D. A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction. IEEE Trans. Med. Imaging 37, 491–503 (2018)

    work page 2018

  13. [13]

    & Sheng, K

    Xu, D., Liu, H., Ruan, D. & Sheng, K. Learning Dynamic MRI Reconstruction with Convolutional Network Assisted Reconstruction Swin Transformer. in Medical Image Computing and Computer Assisted Intervention – MICCAI 2023 Workshops (eds Woo, J. et al.) vol. 14394 3–13 (Springer Nature Switzerland, Cham, 2023)

    work page 2023

  14. [14]

    Xu, D. et al. Paired conditional generative adversarial network for highly accelerated liver 4D MRI. Phys. Med. Biol. 69, 125029 (2024)

    work page 2024

  15. [15]

    Xu, D. et al. Rapid reconstruction of extremely accelerated liver 4D MRI via chained iterative refinement. in Medical Imaging 2025: Image Processing (eds Colliot, O. & Mitra, J.) 34 (SPIE, San Diego, United States, 2025). doi:10.1117/12.3034640

    work page doi:10.1117/12.3034640 2025

  16. [16]

    Sarma, M. et al. Accelerating Dynamic Magnetic Resonance Imaging (MRI) for Lung Tumor Tracking Based on Low-Rank Decomposition in the Spatial–Temporal Domain: A Feasibility Study Based on Simulation and Preliminary Prospective Undersampled MRI. Int. J. Radiat. Oncol. 88, 723–731 (2014)

    work page 2014

  17. [17]

    & Sheng, K

    Zhao, N., O’Connor, D., Basarab, A., Ruan, D. & Sheng, K. Motion Compensated Dynamic MRI Reconstruction With Local Affine Optical Flow Estimation. IEEE Trans. Biomed. Eng. 66, 3050–3059 (2019)

    work page 2019

  18. [18]

    Knoll, F . et al. Deep-Learning Methods for Parallel Magnetic Resonance Imaging Reconstruction: A Survey of the Current Approaches, Trends, and Issues. IEEE Signal Process. Mag. 37, 128–140 (2020)

    work page 2020

  19. [19]

    & Prieto, C

    Cruz, G., Atkinson, D., Buerger, C., Schaeffter, T. & Prieto, C. Accelerated motion corrected three‐dimensional abdominal MRI using total variation regularized SENSE reconstruction. Magn. Reson. Med. 75, 1484–1498 (2016)

    work page 2016

  20. [20]

    Holtackers, R. J. & Stuber, M. Free-Running Cardiac and Respiratory Motion-Resolved Imaging: A Paradigm Shift for Managing Motion in Cardiac MRI? Diagnostics 14, 1946 (2024)

    work page 1946

  21. [21]

    Uecker, M. et al. Real‐time MRI at a resolution of 20 ms. NMR Biomed. 23, 986–994 (2010)

    work page 2010

  22. [22]

    Goodfellow, I. et al. Generative adversarial networks. Commun. ACM 63, 139–144 (2020)

    work page 2020

  23. [23]

    Saharia, C. et al. Image Super-Resolution Via Iterative Refinement. IEEE Trans. Pattern Anal. Mach. Intell. 1–14 (2022) doi:10.1109/TPAMI.2022.3204461

    work page doi:10.1109/tpami.2022.3204461 2022

  24. [24]

    & Brox, T

    Ronneberger, O., Fischer, P . & Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (eds Navab, N., Hornegger, J., Wells, W. M. & Frangi, A. F .) vol. 9351 234–241 (Springer International Publishing, Cham, 2015)

    work page 2015

  25. [25]

    Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. Preprint at https://doi.org/10.48550/arXiv.1512.03385 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1512.03385 2015

  26. [26]

    Liu, X. et al. Implicit Diffusion Models for Continuous Super-Resolution. Int. J. Comput. Vis. 133, 6535–6557 (2025)

    work page 2025

  27. [27]

    Enhanced Deep Residual Networks for Single Image Super-Resolution

    Lim, B., Son, S., Kim, H., Nah, S. & Lee, K. M. Enhanced Deep Residual Networks for Single Image Super-Resolution. Preprint at https://doi.org/10.48550/arXiv.1707.02921 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1707.02921 2017

  28. [28]

    Liu, Q. H. & Nguyen, N. An accurate algorithm for nonuniform fast Fourier transforms (NUFFT’s). IEEE Microw. Guid. Wave Lett. 8, 18–20 (1998)

    work page 1998

  29. [29]

    Feng, L. et al. XD‐GRASP: Golden‐angle radial MRI with reconstruction of extra motion‐ state dimensions using compressed sensing. Magn. Reson. Med. 75, 775–788 (2016)

    work page 2016

  30. [30]

    & Yin, F .-F

    Cai, J., Chang, Z., Wang, Z., Paul Segars, W. & Yin, F .-F . Four-dimensional magnetic resonance imaging (4D-MRI) using image-based respiratory surrogate: a feasibility study. Med. Phys. 38, 6384–6394 (2011)

    work page 2011

  31. [31]

    & Sheng, K

    Xu, D., Descovich, M., Liu, H. & Sheng, K. Robust localization of poorly visible tumor in fiducial free stereotactic body radiation therapy. Radiother. Oncol. 200, 110514 (2024)

    work page 2024

  32. [32]

    Xu, D. et al. Mask R-CNN assisted 2.5D object detection pipeline of 68Ga-PSMA-11 PET/CT-positive metastatic pelvic lymph node after radical prostatectomy from solely CT imaging. Sci. Rep. 13, 1696 (2023)

    work page 2023

  33. [33]

    Larson, A. C. et al. Preliminary investigation of respiratory self‐gating for free‐breathing segmented cine MRI. Magn. Reson. Med. 53, 159–168 (2005)

    work page 2005

  34. [34]

    Feng, L. et al. Golden‐angle radial sparse parallel MRI: Combination of compressed sensing, parallel imaging, and golden‐angle radial sampling for fast and flexible dynamic volumetric MRI. Magn. Reson. Med. 72, 707–717 (2014)

    work page 2014

  35. [35]

    Oliver, P . A. K. et al. Influence of intra- and interfraction motion on planning target volume margin in liver stereotactic body radiation therapy using breath hold. Adv. Radiat. Oncol. 6, 100610 (2021)

    work page 2021

  36. [36]

    B., Vogelius, I

    Stick, L. B., Vogelius, I. R., Risum, S. & Josipovic, M. Intrafractional fiducial marker position variations in stereotactic liver radiotherapy during voluntary deep inspiration breath-hold. Br. J. Radiol. 93, 20200859 (2020)

    work page 2020

  37. [37]

    Ehrbar, S. et al. Intra- and inter-fraction breath-hold variations and margins for radiotherapy of abdominal targets. Phys. Imaging Radiat. Oncol. 28, 100509 (2023)

    work page 2023

  38. [38]

    Sung, J., Choi, Y ., Kim, J., Kim, J. W. & Kim, J. Development of in-house software to process real-time cine magnetic resonance images acquired during 1.5 T MR-guided radiation therapy. Sci. Rep. 15, 29515 (2025)

    work page 2025

  39. [39]

    Lewis, B. et al. Evaluating motion of pancreatic tumors and anatomical surrogates using cine MRI in 0.35T MRgRT under free breathing conditions. J. Appl. Clin. Med. Phys. 24, e13930 (2023)

    work page 2023

  40. [40]

    Kurz, C. et al. Medical physics challenges in clinical MR-guided radiotherapy. Radiat. Oncol. Lond. Engl. 15, 93 (2020)

    work page 2020

  41. [41]

    Holyoake, D. L. P ., Aznar, M., Mukherjee, S., Partridge, M. & Hawkins, M. A. Modelling duodenum radiotherapy toxicity using cohort dose-volume-histogram data. Radiother. Oncol. J. Eur. Soc. Ther. Radiol. Oncol. 123, 431–437 (2017)

    work page 2017

  42. [42]

    & Ceberg, S

    Edvardsson, A., Nordström, F ., Ceberg, C. & Ceberg, S. Motion induced interplay effects for VMAT radiotherapy. Phys. Med. Biol. 63, 085012 (2018)

    work page 2018

  43. [43]

    Caravatta, L. et al. Role of upper abdominal reirradiation for gastrointestinal malignancies: a systematic review of cumulative dose, toxicity, and outcomes on behalf of the Re-Irradiation Working Group of the Italian Association of Radiotherapy and Clinical Oncology (AIRO). Strahlenther. Onkol. Organ Dtsch. Rontgengesellschaft Al 196, 1–14 (2020)

    work page 2020

  44. [44]

    & Chetty , I

    Mao, W., Kim, J. & Chetty , I. J. Association Between Internal Organ/Liver Tumor and External Surface Motion From Cine MR Images on an MRI-Linac. Front. Oncol. 12, 868076 (2022)

    work page 2022

  45. [45]

    Kissick, M. W. & Mackie, T. R. Task Group 76 Report on ‘The management of respiratory motion in radiation oncology’ [Med. Phys. 33, 3874-3900 (2006)]. Med. Phys. 36, 5721– 5722 (2009)

    work page 2006