Pith. sign in

REVIEW 1 major objections 1 minor 41 references

Nonperiodic dynamic CT reconstruction using backward-warping INR with regularization of diffeomorphism (BIRD)

T0 review · 1 major / 1 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read BIRD reconstructs nonperiodic cardiac CT from one projection per frame, removing motion artifacts without pre-scans.

desk verdict BIRD is a plausible, well-engineered INR framework for dynamic CT, but the paper's central accuracy claim is unsubstantiated because it never compares against the prior INR methods it criticizes. read the letter →

arxiv 2505.03463 v1 pith:KFPNQFBR submitted 2025-05-06 cs.CV physics.med-ph

Muge Du , Zhuozhao Zheng , Wenying Wang , Guotao Quan , Wuliang Shi , Le Shen , Li Zhang , Liang Li
show 2 more authors
Yinong Liu Yuxiang Xing
This is my paper · ORCID
classification cs.CVphysics.med-ph
keywords dynamicCTreconstructionimplicitneuralrepresentationdiffeomorphicdeformationfieldbackwardwarpingnonperiodicmotioncardiacartifactreductionlimited-angle
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

BIRD is a reconstruction method aimed at CT scans of anatomy that moves nonperiodically, where every projection view is treated as its own instant in time. The paper's central claim is that a backward-warping implicit neural representation, kept approximately diffeomorphic, can turn such one-projection-per-frame data into a full four-dimensional image with sharply reduced motion artifacts. If correct, cardiac CT would no longer need to bin projections across several heartbeats, and routine scans spoiled by breathing or other motion could be reconstructed without a repeat acquisition or a patient-specific pre-scan. The paper shows on simulated phantoms, a moving heart-and-lung phantom, and a retrospective patient scan that the complete method beats conventional FDK and compressed-sensing baselines and that each added component improves detail and DVF quality.

What carries the argument

The load-bearing object is the backward-warping deformation model. A dynamic coordinate $(x_d,t)$ maps to a static coordinate $x_s = x_d + D_{d\to s}(x_d,t)$, so one ray sample needs one query of the deformation network and one query of the static image network; forward-warping alternatives need the whole field. A paired forward field $D_{s\to d}$ makes a cycle, and the method penalizes the Jacobian of the round-trip residual $x_s + D_{s\to d}(x_s,t) + D_{d\to s}(x_s + D_{s\to d}(x_s,t), t) - x_s$, which keeps the deformation approximately diffeomorphic and anatomically plausible. A free-form feature encoded from the raw 4D coordinate extends the model to intensity or topological changes a warp cannot express. Motion-compensated partial-angle analytical reconstructions, aligned by the forward field and summed into a static space, feed high-frequency detail into the final image and modulate its texture through the reconstruction kernel.

What would settle it

A decisive test is to simulate a dynamic phantom with known ground-truth motion that includes an abrupt topological change—for example, a contrast agent splitting into two blobs—and known projections. If the free-form feature plus the diffeomorphic DVF cannot reproduce the split region while a method with a denser 4D prior can, the claimed representation capacity is insufficient; conversely, if BIRD reproduces the region but its DVF is wrong, the inverse-consistency loss alone is not enforcing true deformation.

Watch

Extended reading notes

Core claim

The central claim is that a nonperiodic dynamic attenuation field can be represented as two complementary features. The first is a topology-preserving feature: a static reference image queried at a static coordinate $x_s = x_d + D_{d\to s}(x_d,t)$, so each dynamic voxel is produced by one backward-warping DVF query and one static-image query. The second is a free-form feature built directly from the 4D coordinate, which absorbs intensity and topological changes that a pure warp cannot represent. The paper further claims that pairing the backward DVF with a forward DVF and penalizing the Jacobian of their round-trip keeps deformations physiologically plausible, and that motion-compensated analytical reconstructions built from partial-angle images inject high-frequency detail into the final image without any pre-scan. In the experiments, the full method removes cardiac and respiratory motion artifacts that blur FDK and PICCS reconstructions, preserves small coronary vessels, and remains the best variant when the gantry rotation is slow enough that each cardiac phase is seen from an extreme limited-angle set.

Load-bearing premise

The load-bearing premise is that the true 4D attenuation field can be decomposed into a smoothly warped static reference image plus a free-form feature, and that minimizing per-ray projection fidelity selects that decomposition uniquely; if the real motion cannot be expressed this way, or if several decompositions fit the data equally well, the reconstruction is not guaranteed to be the true image.

Editorial extensions

If this is right

  • One-beat cardiac imaging becomes possible: a single cardiac cycle's worth of projections can yield a full 4D cardiac image without retrospective phase binning.
  • Conventional scans with nonperiodic motion (failed breath-hold, peristalsis, patient movement) can be reconstructed cleanly, potentially avoiding repeat scans.
  • Computational cost no longer scales with whole-volume DVF evaluation: each ray queries the backward DVF and static image exactly once.
  • No patient-specific historical CT is needed for detail preservation, because high-frequency information is recovered from the current scan's own projections through motion-compensated analytical reconstruction.
  • The diffeomorphism constraint gives a principled way to balance DVF complexity and anatomical plausibility in an ill-posed, one-projection-per-view setting.

Reading between the lines

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

  • Editorial inference: this dual-feature decomposition is a template for other one-shot dynamic inverse problems—4D cone-beam CT on C-arms, dynamic PET, or cine MRI—where each measurement is tied to a single instant.
  • Editorial inference: the paper leaves identifiability unproven; a simple stress test is to re-run BIRD from several random initializations on one projection dataset and measure whether the reconstructed 4D volumes agree away from the measured rays.
  • Editorial inference: the free-form feature is the least constrained part of the model, so a useful extension would be a minimal-explanation penalty that forces the DVF to carry as much of the motion as possible when topology is known to be preserved.
  • Editorial inference: because the analytical-reconstruction kernel can be adjusted to control final noise texture, a deployment-focused extension would be to select kernels to match a clinical site's preferred image appearance without retraining the network.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 1 minor

Summary. The paper proposes BIRD, a self-supervised implicit neural representation (INR) framework for nonperiodic dynamic CT reconstruction. The method models a dynamic object as the sum of a topology-preserving feature obtained by backward-warping a static image with a learned deformation vector field (DVF), and a free-form feature over the 4D coordinates. A bidirectional DVF module applies a diffeomorphism regularization loss adapted from GradICON, and a motion-compensated analytical reconstruction module generates a partial-angle-reconstruction (PAR) image that is fed into the free-form image prediction to enhance fine details. The framework is evaluated on simulated XCAT phantom data, simulated clinical cardiac data, a physical heart/lung phantom, and retrospective patient data, comparing against FDK, PICCS, and three self-ablation variants (INR-base, INR-DM, INR-DM-AR). The paper claims more accurate dynamic CT reconstruction than existing techniques, particularly for one-beat cardiac imaging and motion artifact removal in conventional scans.

Significance. If the claims are substantiated, BIRD addresses a clinically relevant problem: reconstructing high-resolution 4D images from one projection per temporal frame under nonperiodic motion, without patient-specific pre-scans. The framework has several commendable design elements: backward-warping reduces per-voxel computation compared to forward-warping; the diffeomorphism regularization is a principled way to constrain DVF plausibility; and the analytical reconstruction module attempts to recover high-frequency details that pure INR optimization tends to over-smooth. The paper also includes real projection data from commercial CT systems, which is valuable. However, the central comparative claim is currently not supported by the experiments: the closest prior INR-based dynamic CT methods criticized in the introduction are never included as baselines. The registration loss and analytical reconstruction module also raise circularity concerns that are not addressed. As a result, the work is promising but requires additional validation and analysis before its accuracy claims can be accepted.

major comments (1)
  1. [Section II-D; Section II-B.3] The number of PAR groups K is a central hyperparameter in the analytical reconstruction module (Eqs. (20)-(21)), but it is never specified in the implementation details or in any experimental description. Likewise, the PAR reconstruction kernel and angular width of each projection group are not reported. Since the ablation INR-DM-AR versus INR-DM is used to justify the analytical module, the results are not reproducible without these settings. Please state K and the PAR reconstruction parameters for each experiment, and provide a sensitivity analysis or at least a short study of K's effect on PSNR and artifact level.
minor comments (1)
  1. [Section II-D] The total loss in Eq. (26) uses λDM = 1.0 and λRGT = 0.1 as 'recommended for first attempt,' but the paper does not report how sensitive the results are to these weights. A short sensitivity study or a statement that the results are robust over a reasonable range would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reconstruction is grounded in measured projections via the fidelity loss, and the auxiliary losses are internal consistency regularizers rather than fitted predictions.

full rationale

The central reconstruction is not circular. Both topology-preserving and free-form images, mu_TP and mu_FF, are directly optimized by the data fidelity loss LFDL (Eqs. 13-14) against measured post-log projections p, so the final images are anchored to external data. The diffeomorphism loss LDM (Eq. 18) and the forward-DVF registration loss LRGT (Eq. 19) are self-consistency regularizers. In Eq. 19, LRGT compares two evaluations of the same INR field mu_TP, but mu_TP is itself constrained by projection fidelity, and the loss only shapes Ds->d to be inverse-consistent with Dd->s; it does not replace measured data with the network's own output as the reconstruction target. The analytical reconstruction mu_AR is formed from measured partial-angle reconstructions and the estimated DVF and is then supplied to mu_FF (Eq. 12); because mu_FF is still held to the measured projections by LFDL, the enhancement loop does not reduce the reconstruction to a fitted constant or to its own output. The framework invokes no load-bearing self-citation and no uniqueness theorem from the authors; GradICON and HyperNeRF are external prior works. The comparative claim against prior INR methods is under-supported, since STINR and PMF-STINR are not run on the same data, but that is an experimental evidence gap, not a definitional circularity. Overall the derivation chain is self-contained and grounded in projection measurements.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The method rests on a standard linear CT model and several domain assumptions about deformability and identifiability. The main nonstandard postulates are the dual-feature decomposition and the sufficiency of one-projection-per-frame with INR regularization. Loss weights are hand-selected per scenario. No new physical entities are introduced; the latent features are internal representations without external handles.

free parameters (5)
  • lambda_TP (fidelity balance weight) = 0.9 for topology-invariant, 0.5 general, 0.1 for free-form-dominated
    Controls the balance between topology-preserving and free-form fidelity losses in Eq. 14; the paper says it can be tuned for different scenarios, so it is a hand-chosen parameter per dataset.
  • lambda_DM = 1.0
    Weight of the diffeomorphism regularization loss in Eq. 26, given as a recommended first-attempt value in Section II-D.
  • lambda_RGT = 0.1
    Weight of the forward-DVF registration loss in Eq. 26, recommended in Section II-D.
  • K (number of PAR projection groups) = not reported
    K controls the temporal window width in the Analytical Reconstruction Module (Eqs. 20-21) and is never given a value in the experiments, though it directly affects the detail and artifact trade-off.
  • Hashgrid and training hyperparameters = base_res 16, hashmap 19, 16 levels, 2 feats, MLP 32, lr 3e-2/3e-3, 1024 rays
    Implementation choices specified in Section II-D; they are hand-chosen defaults and may affect the reported results.
assumptions (5)
  • standard math The CT forward model is accurately discretized as a linear system per view (Eq. 3).
    Standard assumption in CT reconstruction; needed for the data fidelity loss.
  • ad hoc to paper Dynamic attenuation can be decomposed into a topology-preserving deformed static image plus a free-form feature over 4D coordinates.
    Introduced in Section II-B.1 to model both deformation and topological or intensity changes; not derived from physics and not independently validated.
  • domain assumption The backward DVF is approximately diffeomorphic, so inverse-consistency regularization produces anatomically plausible motion.
    Section II-B.2, based on GradICON. Biological motion is generally invertible, but nonperiodic or contrast-driven changes may violate this.
  • ad hoc to paper One projection per temporal frame, together with the INR regularizers, is sufficient to identify the true dynamic image.
    The extreme limited-angle problem is the central ill-posedness; the paper assumes the losses and low-dimensional encoding select the correct solution, without a uniqueness guarantee.
  • domain assumption The motion-compensated analytical reconstruction, the PAR sum, provides useful high-frequency information rather than mainly artifacts.
    Section II-B.3; PARs from narrow angular ranges are severely limited-angle, and the paper relies on their sum to sharpen the final image.
invented entities (3)
  • Topology-preserving feature (FTP)
    purpose: Latent feature of the warped static reference image used to constrain the reconstruction to topology-preserving motion.
    An internal network representation, not observable; no independent falsifiable handle.
  • Free-form feature (FFF)
    purpose: Hyper-space latent feature meant to capture topological changes, contrast changes, and other variations not representable by warping.
    An internal latent variable inspired by HyperNeRF; no independent handle outside the model.
  • Intermediate static coordinate space (canonical frame)
    purpose: A common reference coordinate space that backward DVFs map into and forward DVFs map out of.
    A modeling construct, not a physical observable; its definition changes during training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonperiodic dynamic CT reconstruction using backward-warping INR with regularization of diffeomorphism (BIRD)." pith.science (2026). https://pith.science/paper/KFPNQFBR

@misc{pith2026250503463,
  author       = {Pith},
  title        = {Pith review of: Nonperiodic dynamic CT reconstruction using backward-warping INR with regularization of diffeomorphism (BIRD)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KFPNQFBR}},
  note         = {Machine review of arXiv:2505.03463}
}
read the original abstract

Dynamic computed tomography (CT) reconstruction faces significant challenges in addressing motion artifacts, particularly for nonperiodic rapid movements such as cardiac imaging with fast heart rates. Traditional methods struggle with the extreme limited-angle problems inherent in nonperiodic cases. Deep learning methods have improved performance but face generalization challenges. Recent implicit neural representation (INR) techniques show promise through self-supervised deep learning, but have critical limitations: computational inefficiency due to forward-warping modeling, difficulty balancing DVF complexity with anatomical plausibility, and challenges in preserving fine details without additional patient-specific pre-scans. This paper presents a novel INR-based framework, BIRD, for nonperiodic dynamic CT reconstruction. It addresses these challenges through four key contributions: (1) backward-warping deformation that enables direct computation of each dynamic voxel with significantly reduced computational cost, (2) diffeomorphism-based DVF regularization that ensures anatomically plausible deformations while maintaining representational capacity, (3) motion-compensated analytical reconstruction that enhances fine details without requiring additional pre-scans, and (4) dimensional-reduction design for efficient 4D coordinate encoding. Through various simulations and practical studies, including digital and physical phantoms and retrospective patient data, we demonstrate the effectiveness of our approach for nonperiodic dynamic CT reconstruction with enhanced details and reduced motion artifacts. The proposed framework enables more accurate dynamic CT reconstruction with potential clinical applications, such as one-beat cardiac reconstruction, cinematic image sequences for functional imaging, and motion artifact reduction in conventional CT scans.

Figures

Figures reproduced from arXiv: 2505.03463 by the authors.

Figure 1
Figure 1. Illustration of the proposed framework. its efficiency as analyzed in Section II-A-3 in high resolution reconstruction. Given a dynamic coordinate (xd, t), the static coordinate xs is computed using (8). Next, the static coordinate xs is encoded using an encoding block e µ TP in the INR-based Reconstruction Module: FTP = e µ TP (xs) (9) FTP is the topology-preserving feature. A topology￾preserving dynamic image µTP … view at source ↗
Figure 2
Figure 2. Illustration of the losses in the proposed framework. To address this issue, we encourage the estimated backward￾warping DVF to be approximately diffeomorphic by intro￾ducing a complementary forward DVF INR and a diffeo￾morphism regularization loss enforcing inverse-consistency between these bidirectional DVFs. The resulted Bidirectional DVF Estimation Module is visualized in the green region of [PITH_FULL_IMAGE:fi… view at source ↗
Figure 1
Figure 1. a) Bidirectional DVF estimation In Bidirectional DVF Estimation Module, there are a back￾ward DVF INR Dd→s that aligns the original dynamic coor￾dinate space with the intermediate static space as in (8), and a forward DVF INR Ds→d for the inverse alignment as in (7). Following the general architecture of INR in (6), both backward and forward DVF INRs consist of an encoding block for spatial-temporal input, and a pre… view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: Illustration of the dimensional reduction structure for encoding block with 4D input. Here, γx and γt represent the 3D and 1D multi-resolution hashgrid encoding functions, while φγx and φγt are single￾hidden-layer MLPs. hx and ht are the hierarchical-aligned encoded fe…
Figure 4
Figure 4. Figure 4: The violin-plot of PSNR counted on 10 phases of the XCAT phantom in the 1st simulation study. The comparison methods include conventional FDK and PICCS [4], and three variants of the proposed framework, as detailed in Table-II, that evaluated the contribution of the pr…
Figure 7
Figure 7. Figure 7: The visualization of estimated backward DVFs in different methods (the 1st to 3rd columns), and estimated AR (analytical recon￾struction) in INR-DM-AR (the 4th column), in the 1st simulation study. The gantry rotation speed is 0.5 s/rot (the 1st row) and 0.25 s/rot (th…
Figure 6
Figure 6. Figure 6: A reconstructed slice at the 1st phase of the XCAT phantom in the 1st simulation study, with the gantry rotation speed set to 0.5 s/rot, and the heart rate set to 120 bpm. The display window is [-600,300] HU. both INR-DM and INR-DM-AR yield optimal result. At the slowe…
Figure 8
Figure 8. Figure 8: The violin-plot of PSNR counted on 20 phases of the cardiac image in the 2nd simulation study. linearly combined in each up-sampled frame, resulting in a dynamic image significantly affected by free-form variations in addition to the cardiac motion. With these 2500 fra…
Figure 10
Figure 10. Figure 10: demonstrates improved performance across all [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 13
Figure 13. Figure 13: A reconstructed slice of the dynamic heart and lung phantom from an axial coronary CTA scan in the 1st practical study, with the gantry rotation speed set to 0.25 s/rot. The display window is [-1024,512] HU for the 1st row and [-250, 450] HU for the 2nd row. C. Practi…
Figure 12
Figure 12. Figure 12: The visualization of estimated backward DVFs in different methods (the 1st to 3rd columns), and estimated AR (analytical recon￾struction) in INR-DM-AR (the 4th column), in the 2nd simulation study. The heart rate is 80 bpm (the 1st row) and 120 bpm (the 2nd row), resp…
Figure 14
Figure 14. Figure 14: A reconstructed slice of the thoracic region from a retrospective helical abdominal CT scan in the 2nd practical study, with the gantry rotation speed set to 0.5 s/rot. The display window is [-1250,250] HU. We therefore applied our method to reconstruct the thoracic r…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages

  1. [1]

    Iterative 4D cardiac micro-CT image reconstruction using an adaptive spatio- temporal sparsity prior,

    L. Ritschl, S. Sawall, M. Knaup, A. Hess, and M. Kachelrieß, “Iterative 4D cardiac micro-CT image reconstruction using an adaptive spatio- temporal sparsity prior,” Physics in Medicine & Biology , vol. 57, no. 6, p. 1517, Mar. 2012

    work page 2012

  2. [2]

    Four-dimensional cone beam CT reconstruction and enhancement using a temporal non- local means method,

    X. Jia, Z. Tian, Y . Lou, J.-J. Sonke, and S. B. Jiang, “Four-dimensional cone beam CT reconstruction and enhancement using a temporal non- local means method,” Medical Physics , vol. 39, no. 9, pp. 5592–5602, 2012

    work page 2012

  3. [3]

    4D cone beam CT via spatiotem- poral tensor framelet,

    H. Gao, R. Li, Y . Lin, and L. Xing, “4D cone beam CT via spatiotem- poral tensor framelet,” Medical Physics, vol. 39, no. 11, pp. 6943–6946, 2012

    work page 2012

  4. [4]

    Prior image constrained compressed sensing (PICCS): A method to accurately reconstruct dynamic CT images from highly undersampled projection data sets,

    G.-H. Chen, J. Tang, and S. Leng, “Prior image constrained compressed sensing (PICCS): A method to accurately reconstruct dynamic CT images from highly undersampled projection data sets,” Medical physics, vol. 35, no. 2, pp. 660–663, Feb. 2008

    work page 2008

  5. [5]

    On-the- fly motion-compensated cone-beam CT using an a priori model of the respiratory motion: On-the-fly motion-compensated cone-beam CT,

    S. Rit, J. W. H. Wolthaus, M. van Herk, and J.-J. Sonke, “On-the- fly motion-compensated cone-beam CT using an a priori model of the respiratory motion: On-the-fly motion-compensated cone-beam CT,” Medical Physics , vol. 36, no. 6Part1, pp. 2283–2296, May 2009

    work page 2009

  6. [6]

    Motion-map constrained image reconstruction (MCIR): Application to four-dimensional cone-beam computed tomography,

    J. C. Park, J. S. Kim, S. H. Park, Z. Liu, B. Song, and W. Y . Song, “Motion-map constrained image reconstruction (MCIR): Application to four-dimensional cone-beam computed tomography,” Medical Physics , vol. 40, no. 12, p. 121710, 2013

    work page 2013

  7. [7]

    Simultaneous motion estimation and image recon- struction (SMEIR) for 4D cone-beam CT,

    J. Wang and X. Gu, “Simultaneous motion estimation and image recon- struction (SMEIR) for 4D cone-beam CT,” Medical Physics , vol. 40, no. 10, p. 101912, 2013

    work page 2013

  8. [8]

    General simultaneous motion estimation and image reconstruction (G-SMEIR),

    S. Zhou, Y . Chi, J. Wang, and M. Jin, “General simultaneous motion estimation and image reconstruction (G-SMEIR),” Biomedical physics & engineering express , vol. 7, no. 5, pp. 10.1088/2057–1976/ac12a4, Jul. 2021

    doi:10.1088/2057 1976

Show all 41 references
  1. [9]

    Reduction of motion artifacts in cardiac CT based on partial angle reconstructions from short scan data,

    J. Hahn, H. Bruder, T. Allmendinger, K. Stierstorfer, T. Flohr, and M. Kachelriess, “Reduction of motion artifacts in cardiac CT based on partial angle reconstructions from short scan data,” in Medical Imaging 2016: Physics of Medical Imaging , vol. 9783. SPIE, Mar. 2016, pp. 321–329

    2016

  2. [10]

    Motion compensation in the region of the coronary arteries based on partial angle reconstructions from short- scan CT data,

    J. Hahn, H. Bruder, C. Rohkohl, T. Allmendinger, K. Stierstorfer, T. Flohr, and M. Kachelrieß, “Motion compensation in the region of the coronary arteries based on partial angle reconstructions from short- scan CT data,” Medical Physics , vol. 44, no. 11, pp. 5795–5813, Nov. 2...

    2017

  3. [11]

    Motion compensation for aortic valves using partial angle CT recon- structions,

    S. Lebedev, E. Fourni ´e, J. Maier, K. Stierstorfer, and M. Kachelrieß, “Motion compensation for aortic valves using partial angle CT recon- structions,” Medical Physics , vol. 49, no. 3, pp. 1495–1506, 2022

    2022

  4. [12]

    High-quality four-dimensional cone-beam CT by deforming prior images,

    J. Wang and X. Gu, “High-quality four-dimensional cone-beam CT by deforming prior images,” Physics in Medicine and Biology , vol. 58, no. 2, pp. 231–246, Jan. 2013

    2013

  5. [13]

    A technique for esti- mating 4D-CBCT using prior knowledge and limited-angle projections,

    Y . Zhang, F.-F. Yin, W. P. Segars, and L. Ren, “A technique for esti- mating 4D-CBCT using prior knowledge and limited-angle projections,” Medical Physics , vol. 40, no. 12, p. 121701, 2013

    2013

  6. [14]

    Motion-aware temporal regulariza- tion for improved 4D cone-beam computed tomography,

    C. Mory, G. Janssens, and S. Rit, “Motion-aware temporal regulariza- tion for improved 4D cone-beam computed tomography,” Physics in Medicine and Biology , vol. 61, no. 18, pp. 6856–6877, 2016

    2016

  7. [15]

    Self-contained deep learning-based boosting of 4D cone-beam CT reconstruction,

    F. Madesta, T. Sentker, T. Gauer, and R. Werner, “Self-contained deep learning-based boosting of 4D cone-beam CT reconstruction,” Medical Physics, vol. 47, no. 11, pp. 5619–5631, 2020

    2020

  8. [16]

    4D-AirNet: A temporally- resolved CBCT slice reconstruction method synergizing analytical and iterative method with deep learning,

    G. Chen, Y . Zhao, Q. Huang, and H. Gao, “4D-AirNet: A temporally- resolved CBCT slice reconstruction method synergizing analytical and iterative method with deep learning,” Physics in Medicine &amp$\mathsemicolon$ Biology , vol. 65, no. 17, p. 175020, 2020

    2020

  9. [17]

    En- hancement of 4-D Cone-Beam Computed Tomography (4D-CBCT) Using a Dual-Encoder Convolutional Neural Network (DeCNN),

    Z. Jiang, Z. Zhang, Y . Chang, Y . Ge, F.-F. Yin, and L. Ren, “En- hancement of 4-D Cone-Beam Computed Tomography (4D-CBCT) Using a Dual-Encoder Convolutional Neural Network (DeCNN),” IEEE Transactions on Radiation and Plasma Medical Sciences , vol. 6, no. 2, pp. 222–230, 2022

    2022

  10. [18]

    TT U-Net: Temporal Transformer U-Net for Motion Artifact Reduction Using PAD (Pseudo All-Phase Clinical-Dataset) in Cardiac CT,

    Z. Deng, W. Zhang, K. Chen, Y . Zhou, J. Tian, G. Quan, and J. Zhao, “TT U-Net: Temporal Transformer U-Net for Motion Artifact Reduction Using PAD (Pseudo All-Phase Clinical-Dataset) in Cardiac CT,” IEEE Transactions on Medical Imaging , vol. 42, no. 12, pp. 3805–3816, Dec. 2023

    2023

  11. [19]

    U-net-based deformation vector field estimation for motion-compensated 4D-CBCT reconstruc- tion,

    X. Huang, Y . Zhang, L. Chen, and J. Wang, “U-net-based deformation vector field estimation for motion-compensated 4D-CBCT reconstruc- tion,” Medical Physics , vol. 47, no. 7, pp. 3000–3012, 2020

    2020

  12. [20]

    Respiratory deformation registration in 4D-CT/cone beam CT using deep learning,

    X. Teng, Y . Chen, Y . Zhang, and L. Ren, “Respiratory deformation registration in 4D-CT/cone beam CT using deep learning,” Quantitative Imaging in Medicine and Surgery , vol. 11, no. 2, pp. 73 748–73 748, Feb. 2021

    2021

  13. [21]

    DeepOrganNet: On-the-Fly Recon- struction and Visualization of 3D / 4D Lung Models from Single- View Projections by Deep Deformation Network,

    Y . Wang, Z. Zhong, and J. Hua, “DeepOrganNet: On-the-Fly Recon- struction and Visualization of 3D / 4D Lung Models from Single- View Projections by Deep Deformation Network,” IEEE Transactions on Visualization and Computer Graphics , vol. 26, no. 1, pp. 960–970, 2020

    2020

  14. [22]

    Estimate and compensate head motion in non-contrast head CT scans using partial angle reconstruction and deep learning,

    Z. Chen, Q. Li, and D. Wu, “Estimate and compensate head motion in non-contrast head CT scans using partial angle reconstruction and deep learning,” Medical Physics , vol. 51, no. 5, pp. 3309–3321, 2024

    2024

  15. [23]

    A Spatiotemporal V olumetric Interpolation Network for 4D Dynamic Medical Image,

    Y . Guo, L. Bi, E. Ahn, D. Feng, Q. Wang, and J. Kim, “A Spatiotemporal V olumetric Interpolation Network for 4D Dynamic Medical Image,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) . Seattle, W A, USA: IEEE, Jun. 2020, pp. 4725–4734

    2020

  16. [24]

    Deep learning-based motion compensation for four-dimensional cone-beam computed tomography (4D-CBCT) reconstruction,

    Z. Zhang, J. Liu, D. Yang, U. S. Kamilov, and G. D. Hugo, “Deep learning-based motion compensation for four-dimensional cone-beam computed tomography (4D-CBCT) reconstruction,” Medical Physics, p. mp.16103, Dec. 2022

    2022

  17. [25]

    Four- Dimensional Cone Beam CT Imaging Using a Single Routine Scan via Deep Learning,

    P. Yang, X. Ge, T. Tsui, X. Liang, Y . Xie, Z. Hu, and T. Niu, “Four- Dimensional Cone Beam CT Imaging Using a Single Routine Scan via Deep Learning,” IEEE Transactions on Medical Imaging , vol. 42, no. 5, pp. 1495–1508, May 2023

    2023

  18. [26]

    Fast four- dimensional cone-beam computed tomography reconstruction using de- formable convolutional networks,

    Z. Jiang, Y . Chang, Z. Zhang, F.-F. Yin, and L. Ren, “Fast four- dimensional cone-beam computed tomography reconstruction using de- formable convolutional networks,” Medical Physics , vol. n/a, no. n/a, 2022

    2022

  19. [27]

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,” in Computer Vision – ECCV 2020 , A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer Interna- tional...

    2020

  20. [28]

    CoIL: Coordinate-Based Internal Learning for Tomographic Imaging,

    Y . Sun, J. Liu, M. Xie, B. Wohlberg, and U. S. Kamilov, “CoIL: Coordinate-Based Internal Learning for Tomographic Imaging,” IEEE Transactions on Computational Imaging , vol. 7, pp. 1400–1412, 2021

    2021

  21. [29]

    IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction,

    G. Zang, R. Idoughi, R. Li, P. Wonka, and W. Heidrich, “IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction,” in 2021 IEEE/CVF International Conference on Com- puter Vision (ICCV) . Montreal, QC, Canada: IEEE, Oct. 2021, pp. 1940–1950

    2021

  22. [30]

    NeRP: Implicit Neural Representation Learning with Prior Embedding for Sparsely Sampled Image Recon- struction,

    L. Shen, J. Pauly, and L. Xing, “NeRP: Implicit Neural Representation Learning with Prior Embedding for Sparsely Sampled Image Recon- struction,” arXiv:2108.10991 [cs, eess] , Aug. 2021

    2021 arXiv

  23. [31]

    NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction,

    R. Zha, Y . Zhang, and H. Li, “NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2022 , ser. Lecture Notes in Computer Science, L. Wang, Q. Dou, P. T. Fletcher, S. Speidel, and S. Li, Eds...

    2022

  24. [32]

    MedNeRF: Medical Neural Radiance Fields for Reconstructing 3D-aware CT-Projections from a Single X- ray,

    A. Corona-Figueroa, J. Frawley, S. Bond-Taylor, S. Bethapudi, H. P. H. Shum, and C. G. Willcocks, “MedNeRF: Medical Neural Radiance Fields for Reconstructing 3D-aware CT-Projections from a Single X- ray,” arXiv:2202.01020 [cs, eess] , Feb. 2022

    2022 arXiv

  25. [33]

    Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields,

    A. W. Reed, H. Kim, R. Anirudh, K. A. Mohan, K. Champley, J. Kang, and S. Jayasuriya, “Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , Oct. 20...

    2021

  26. [34]

    Dynamic cone-beam CT reconstruction using spatial and temporal implicit neural representation learning (STINR),

    Y . Zhang, H.-C. Shao, T. Pan, and T. Mengke, “Dynamic cone-beam CT reconstruction using spatial and temporal implicit neural representation learning (STINR),” Physics in Medicine & Biology , vol. 68, no. 4, p. 045005, Feb. 2023

    2023

  27. [35]

    Dynamic CBCT imaging using prior model-free spatiotemporal implicit neural representation (PMF-STINR),

    H.-C. Shao, T. Mengke, T. Pan, and Y . Zhang, “Dynamic CBCT imaging using prior model-free spatiotemporal implicit neural representation (PMF-STINR),” Physics in Medicine & Biology , vol. 69, no. 11, p. 115030, May 2024

    2024

  28. [36]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,” ACM Transactions on Graphics, vol. 41, no. 4, pp. 1–15, Jul. 2022

    2022

  29. [37]

    HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields,

    K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz, “HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields,” ACM Transactions on Graphics , vol. 40, no. 6, pp. 1–12, Dec. 2021

    2021

  30. [38]

    GradICON: Ap- proximate Diffeomorphisms via Gradient Inverse Consistency,

    L. Tian, H. Greer, F.-X. Vialard, R. Kwitt, R. S. J. Est ´epar, R. J. Rushmore, N. Makris, S. Bouix, and M. Niethammer, “GradICON: Ap- proximate Diffeomorphisms via Gradient Inverse Consistency,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Ju...

    2023

  31. [39]

    4D XCAT phantom for multimodality imaging research,

    W. P. Segars, G. Sturgeon, S. Mendonca, J. Grimes, and B. M. W. Tsui, “4D XCAT phantom for multimodality imaging research,” Medical Physics, vol. 37, no. 9, pp. 4902–4915, 2010

    2010

  32. [40]

    V oxelMorph: A Learning Framework for Deformable Medical Image Registration,

    G. Balakrishnan, A. Zhao, M. R. Sabuncu, J. Guttag, and A. V . Dalca, “V oxelMorph: A Learning Framework for Deformable Medical Image Registration,” IEEE Transactions on Medical Imaging , vol. 38, no. 8, pp. 1788–1800, Aug. 2019

    2019

  33. [41]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering,

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3D Gaussian Splatting for Real-Time Radiance Field Rendering,” ACM Transactions on Graphics , vol. 42, no. 4, pp. 1–14, Aug. 2023

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.