Pith. sign in

REVIEW 3 major objections 7 minor 26 references

Flow matching for sparse-view CT can match diffusion quality while cutting neural-network evaluations by up to 89%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 20:06 UTC pith:6ZU2OYEM

load-bearing objection Solid efficiency claim and reproducible experiments, but the proof for velocity reuse assumes a different data-consistency update than the one actually implemented. the 3 major comments →

arxiv 2603.00205 v2 pith:6ZU2OYEM submitted 2026-02-27 eess.IV cs.AIcs.CV

Efficient Flow Matching for Sparse-View CT Reconstruction

Jiayang Shi , Lincen Yang , Zhong Li , Tristan van Leeuwen , Daniel M. Pelt , K. Joost Batenburg This is my paper
classification eess.IV cs.AIcs.CV
keywords flow matchingCT reconstructionsparse-view CTvelocity reusedata consistencygenerative modelsinverse problemsefficient sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that deterministic flow matching is a better generative backbone for sparse-view CT reconstruction than stochastic diffusion models, because its ODE trajectories do not fight repeated data-consistency corrections. The central claim is that the velocity fields predicted at adjacent integration steps are so correlated that a single network evaluation can be reused for several steps, and that this reuse introduces error of the same order as the Euler discretization itself. With the efficient variant EFMCT, the paper reports reconstruction from 40-view projections at 7 neural-network evaluations and about 0.83 seconds, with PSNR/SSIM competitive with diffusion baselines that need 50 to 1000 evaluations. If this holds, generative CT reconstruction could become practical in time-critical clinical settings.

Core claim

The paper's central claim, stated on its own terms, is that flow-matching-based CT reconstruction is not only competitive with diffusion-based reconstruction but fundamentally cheaper: the deterministic ODE has no stochastic noise to interfere with data consistency, and the high correlation of consecutive velocity fields makes velocity reuse safe. Proposition 1 formalizes this by showing that reusing a velocity for up to M steps yields a local error O(Δt²) and a global error O(Δt) over a fixed time horizon, the same first-order accuracy as the underlying Euler scheme, provided the velocity is Lipschitz and the data-consistency operator is non-expansive. On the AAPM 40-view benchmark, EFMCT a

What carries the argument

The load-bearing component is the velocity-reuse update (Eq. 9): after computing v_t = v_θ(t, x_t), the sampler reuses that same vector for up to M=10 consecutive Euler steps, with an adaptive residual check (Eq. 10) that aborts reuse if the measurement residual grows beyond a factor η=1.05. Data consistency is enforced via a conjugate-gradient correction DC(·) applied at every step (Eqs. 7-8). The theoretical anchor is Proposition 1, which uses Lipschitz continuity of the velocity and κ-non-expansiveness of DC to propagate a one-step deviation δ through a block of M reuse steps and show the accumulated error over the whole trajectory remains O(Δt).

Load-bearing premise

The proof assumes the data-consistency operator is non-expansive, but the conjugate-gradient correction the algorithm actually implements is an additive term whose non-expansiveness is never established; if that correction can enlarge distances between nearby reconstructions, the O(Δt) global error guarantee no longer follows.

What would settle it

Apply the implemented DC update (Eqs. 7-8) to two slightly different estimates of the same image and measure the ratio of output distance to input distance; if any pair gives a ratio greater than 1, the non-expansiveness premise of Proposition 1 fails, and the claimed O(Δt) accumulation bound has no theoretical support.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • EFMCT lowers neural-network evaluations from 50–1000 (diffusion baselines) to 7–11 with no statistically meaningful PSNR/SSIM loss, making generative sparse-view CT reconstruction fast enough for time-critical workflows.
  • Velocity reuse preserves first-order accuracy of the ODE discretization, so the efficiency gain does not come at the cost of an extra error order.
  • Because the method only needs the forward operator A, it can be adapted to other CT geometries (cone-beam, limited-angle) by swapping the projector.
  • The deterministic formulation removes the push-and-pull between denoising and data consistency that diffusion models suffer from, so repeated corrections behave predictably.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the same reuse strategy should transfer to other flow-matching inverse problems (MRI, deblurring, super-resolution) whenever the data-consistency correction is non-expansive; we infer the O(Δt) bound would carry over with minimal changes.
  • Because the adaptive residual check only tests one step ahead, a natural extension would be to make the reuse length M itself adaptive, halting reuse based on residual growth rate; the paper fixes M=10 after an ablation on test images.
  • The theoretical argument assumes DC is non-expansive, but the implemented CG-based additive correction is not proven to be; a concrete next step is to measure its contraction factor numerically to confirm the bound in practice.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes FMCT/EFMCT, flow-matching-based frameworks for sparse-view CT reconstruction. The FMCT method integrates a pretrained rectified-flow model with a conjugate-gradient data-consistency correction at each Euler integration step; the EFMCT variant reuses previously computed velocity fields for up to M consecutive steps to reduce the number of neural-network function evaluations. The authors state a theoretical result, Proposition 1, claiming that the error introduced by velocity reuse is O(Δt) over a fixed time horizon when the velocity is Lipschitz and each step is followed by a non-expansive data-consistency operator. Experiments on AAPM and Decathlon CT datasets report PSNR/SSIM values competitive with diffusion-based baselines (e.g., DDS) at substantially lower NFE and computation time, and the code is open-sourced.

Significance. If the theoretical and empirical claims are made rigorous, the work is practically relevant: a deterministic flow-matching sampler with velocity reuse could provide substantially faster generative CT reconstruction while preserving image quality. The paper evaluates on two external CT datasets, compares against several strong diffusion-based baselines, and openly provides the codebase. However, the paper's central theoretical guarantee is not established for the actual algorithm as implemented: Proposition 1 assumes a non-expansive DC operator applied to the full update, while the algorithm uses an additive CG correction whose composed map is never shown non-expansive. The adaptive reuse check also monitors a pre-correction residual rather than the residual of the final updated image, which is consistent with the markedly worse data-fit values reported for EFMCT in Table 1. These issues are load-bearing for the claim that the reuse strategy is both safe and efficient.

major comments (3)
  1. [Appendix Proposition 1/Eqs. (17)–(18); Section 2 Eqs. (7)–(8); Algorithm 1] The proof of Proposition 1 assumes the data-consistency operation is applied to the full update: x_{k+1}=DC(x_k+Δt v_{t_k}(x_k)) and \hat{x}_{k+1}=DC(\hat{x}_k+Δt v_{t_{k0}}(x_{k0})), with DC κ-non-expansive for 0≤κ≤1. The implemented algorithm is different: it computes v_DC=DC(\hat{x}_0,y,A) from a CG solve on the extrapolated estimate \hat{x}_0 and then sets x_{t-Δt}=x_t-Δt v_t+v_DC. The resulting one-step map T(x)=x-Δt v_t(x)+g(x-t v_t(x)) has no proven non-expansiveness; contractivity of the CG correction g alone does not imply contractivity of T. Since the O(Δt) bound in Eq. (24) relies on κ≤1 at every update, the theoretical guarantee does not currently apply to the algorithm as written. The authors should either align the proof with the additive CG update by stating and proving the required contractivity assumptions, or modify the algorithm so that DC is a true non-expansive proje
  2. [Section 2 Eq. (10); Algorithm 1 lines 5–7; Table 1] The adaptive reuse check in Eq. (10) evaluates \tilde{r}=||A\tilde{x}-y||^2 only on the pre-correction state \tilde{x}=x_t-Δt v_prev and then sets r←\tilde{r}. It never checks the residual of x_{t-Δt}=\tilde{x}+v_DC, which is the image actually stored and used for the next step. Thus the acceptance rule does not verify that the final updated image improves data consistency, and a reuse step can be accepted even if the DC correction worsens the true residual. This is not merely a formal concern: Table 1 reports EFMCT data fit of 182.80 on AAPM 40-view compared with 19.09 for DDS, an order-of-magnitude gap. The acceptance criterion and the reported data-fit metric should be based on the post-DC update, or the paper should justify why the pre-correction residual is the relevant quantity.
  3. [Section 3, Fig. 4 and following paragraph; Section 2 Eq. (10)] The hyperparameters η=1.05 and M=10 are selected from an ablation conducted on '21 randomly selected test CT images' and are then used for all EFMCT results. Selecting hyperparameters on the test set makes the reported EFMCT performance optimistic relative to a fair comparison with baselines whose hyperparameters are not tuned on the same test images. Please move hyperparameter selection to a validation fold, freeze the hyperparameters, and report test-set results separately.
minor comments (7)
  1. [Section 2 heading] 'Flow Maching' should be 'Flow Matching'.
  2. [Appendix, after Eq. (16)] Typo: 'standard eulrx' should be 'standard Euler'.
  3. [Section 2, Eq. (10)] Typo: 'emprically' should be 'empirically'.
  4. [Appendix Eq. (23)] The stated expression κ^2 M(M-1) does not follow from recurrence (22); unrolling (22) gives a sum of the form C Δt^2 Σ_{j=1}^{M-1} j κ^{M-j}. The order O(M^2 Δt^2) remains correct for fixed M, but the constant expression should be corrected.
  5. [Table 1] The 'Data Fit' metric is not defined in the text. Please state explicitly that it is the squared residual ||A x - y||^2 or another defined quantity, and describe how it is averaged over test images.
  6. [Algorithm 1, line 5] The subscript in '\hat{x}_0 ← \tilde{x}_t - t v_prev' is inconsistent with the surrounding notation; clarify whether the extrapolation uses \tilde{x} or the updated pre-correction state.
  7. [Figure 3] The label 'Decathlon 40 views' appears twice in the figure panel text; one of these is probably intended to be 'Decathlon 20 views'. Please correct.

Circularity Check

0 steps flagged

No substantive circularity: the velocity-reuse error bound is a conditional theorem independent of fitted values, and experiments use external benchmarks; minor self-citation and test-set tuning do not make the derivation circular.

full rationale

The central derivation, Proposition 1 (Appendix Eqs. 12-24), is self-contained: it assumes Lipschitz velocity (Eq. 14) and a kappa-non-expansive data-consistency operator (Eq. 17), then derives the local error norm(x_{k+1} - hat x_{k+1}) = O(Delta t^2) and the global O(Delta t) bound. No trained-network output, fitted parameter, or measured PSNR enters the proof; it is a conditional mathematical statement about the reuse scheme. The experimental claims are checked against external datasets (AAPM 2016, Medical Segmentation Decathlon) and external baselines (DPS, MCG, PGDM, DDS, ADMM-TV), so the quality results are not forced by construction. The one same-author citation, DM4CT [18], is used only as background motivation for why diffusion methods are slow; it is not load-bearing in the proof or benchmarks. Two non-circular weaknesses should be flagged: (i) the proof models DC as a non-expansive map applied to the full update (Appendix Eq. 18), whereas Algorithm 1 implements DC as an additive conjugate-gradient correction x_{t-Delta t} = tilde x + v_DC(hat x_0, y, A) (Algorithm 1 lines 7 and 14), and the paper never proves this composed update is non-expansive, so the O(Delta t) bound is not actually established for the implemented algorithm; (ii) the reuse hyperparameters eta=1.05 and M=10 are selected via ablation on 21 test images, so the reported test metrics are partially tuned. These are correctness and validation risks, not circular derivations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper's central claim rests on a trained FM model (no convergence guarantee), a Lipschitz/bounded-velocity assumption in the proof, and a non-expansive DC assumption that is not validated for the implemented additive CG correction. Hyperparameters η and M are hand-set/test-tuned.

free parameters (3)
  • η (relaxation factor) = 1.05
    Empirically set in Section 2; controls how much residual growth is tolerated before recomputing velocity.
  • M (max consecutive reuse steps) = 10
    Selected from ablation in Section 3 on 21 test images; balances NFE and reconstruction quality.
  • N (number of integration steps) = inferred from NFE counts (e.g., 25/50 for FMCT; not stated explicitly)
    Discretization step count; affects accuracy and runtime; not justified beyond 'fixed to 50' in ablation.
axioms (4)
  • standard math Flow matching / rectified flow training objective (Eq. 5) learns a velocity field that transports Gaussian noise to full-view CT images
    Borrowed from Lipman et al. [12] and Liu et al. [15]; assumed correct and applicable.
  • domain assumption Learned velocity field is sufficiently accurate that Euler integration with DC yields valid CT reconstructions
    Underlies all experiments; no convergence guarantee for the learned model.
  • domain assumption Velocity field v is locally Lipschitz in x and t and bounded by Vmax along the trajectory
    Used in the proof of Proposition 1; not verified for the trained network.
  • ad hoc to paper Data-consistency operator is κ-non-expansive, 0≤κ≤1, when applied after each update
    Stated in Appendix Eq. (17); not shown for the implemented additive CG correction; theorem and algorithm diverge.

pith-pipeline@v1.3.0-alltime-deepseek · 9497 in / 14626 out tokens · 125975 ms · 2026-08-02T20:06:00.881871+00:00 · methodology

0 comments
read the original abstract

Generative models, particularly Diffusion Models (DM), have shown strong potential for Computed Tomography (CT) reconstruction serving as expressive priors for solving ill-posed inverse problems. However, diffusion-based reconstruction relies on Stochastic Differential Equations (SDEs) for forward diffusion and reverse denoising, where such stochasticity can interfere with repeated data consistency corrections in CT reconstruction. Since CT reconstruction is often time-critical in clinical and interventional scenarios, improving reconstruction efficiency is essential. In contrast, Flow Matching (FM) models sampling as a deterministic Ordinary Differential Equation (ODE), yielding smooth trajectories without stochastic noise injection. This deterministic formulation is naturally compatible with repeated data consistency operations. Furthermore, we observe that FM-predicted velocity fields exhibit strong correlations across adjacent steps. Motivated by this, we propose an FM-based CT reconstruction framework (FMCT) and an efficient variant (EFMCT) that reuses previously predicted velocity fields over consecutive steps to substantially reduce the number of Neural network Function Evaluations (NFEs), thereby improving inference efficiency. We provide theoretical analysis showing that the error introduced by velocity reuse is bounded when combined with data consistency operations. Extensive experiments demonstrate that FMCT/EFMCT achieve competitive reconstruction quality while significantly improving computational efficiency compared with diffusion-based methods. The codebase is open-sourced at https://github.com/EFMCT/EFMCT.

Figures

Figures reproduced from arXiv: 2603.00205 by Daniel M. Pelt, Jiayang Shi, K. Joost Batenburg, Lincen Yang, Tristan van Leeuwen, Zhong Li.

Figure 1
Figure 1. Figure 1: Left: Cosine similarity of predicted velocity fields at consecutive iterations in FMCT, with and without data consistency correction, showing strong cor￾relation across adjacent steps. Right: Reconstruction performance vs. NFE for DM methods and FMCT/EFMCT, demonstrating that FMCT/EFMCT achieve competitive quality at substantially lower NFEs. DM methods here use the deter￾ministic DDIM sampler for fairness… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed FMCT(left) and EFMCT(right) method. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of reconstructions across different methods, datasets, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation of the velocity reuse strategy. Left: Reuse enabled after differ [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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

26 extracted references

  1. [1]

    Nature communications13(1), 4128 (2022)

    Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nature communications13(1), 4128 (2022)

    2022

  2. [2]

    In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=gv8R88PVEG

    Bajpai, D.J., Agarwal, S., Saxena, A., Kulkarni, K., Mitra, S., Hanawal, M.K.: Flowcast: Trajectory forecasting for scalable zero-cost speculative flow matching. In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=gv8R88PVEG

    2026

  3. [3]

    Foundations and Trends®in Machine learning3(1), 1–122 (2011)

    Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al.: Distributed opti- mizationandstatisticallearningviathealternatingdirectionmethodofmultipliers. Foundations and Trends®in Machine learning3(1), 1–122 (2011)

    2011

  4. [4]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Chen, J., Lin, Y., Qin, Y., Wang, H., Li, X.: Cross-view generalized diffusion model for sparse-view ct reconstruction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 140–150. Springer (2025)

    2025

  5. [5]

    In: The Eleventh International Con- ference on Learning Representations (2023),https://openreview.net/forum?id= OnD9zGAGT0k

    Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Con- ference on Learning Representations (2023),https://openreview.net/forum?id= OnD9zGAGT0k

    2023

  6. [6]

    In: The Twelfth International Conference on Learning Rep- resentations (2024),https://openreview.net/forum?id=DsEhqQtfAG

    Chung, H., Lee, S., Ye, J.C.: Decomposed diffusion sampler for accelerating large- scale inverse problems. In: The Twelfth International Conference on Learning Rep- resentations (2024),https://openreview.net/forum?id=DsEhqQtfAG

    2024

  7. [7]

    In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K

    Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022),https: //openreview.net/forum?id=nJJjv0JDJju

    2022

  8. [8]

    Emergency radiology13(5), 251–257 (2007)

    Fanucci, E., Fiaschetti, V., Rotili, A., Floris, R., Simonetti, G.: Whole body 16- row multislice ct in emergency room: effects of different protocols on scanning time, image quality and radiation exposure. Emergency radiology13(5), 251–257 (2007)

    2007

  9. [9]

    The computer journal7(2), 149–154 (1964)

    Fletcher, R., Reeves, C.M.: Function minimization by conjugate gradients. The computer journal7(2), 149–154 (1964)

    1964

  10. [10]

    SIAM journal on imaging sciences2(2), 323–343 (2009)

    Goldstein, T., Osher, S.: The split bregman method for l1-regularized problems. SIAM journal on imaging sciences2(2), 323–343 (2009)

    2009

  11. [11]

    BioMed research international 2022(1), 5164970 (2022)

    Hussain, S., Mubeen, I., Ullah, N., Shah, S.S.U.D., Khan, B.A., Zahoor, M., Ul- lah, R., Khan, F.A., Sultan, M.A.: Modern diagnostic imaging technique applica- tions and risk factors in the medical field: a review. BioMed research international 2022(1), 5164970 (2022)

    2022

  12. [12]

    In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=PqvMRDCJT9t 10 J

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=PqvMRDCJT9t 10 J. Shi et al

    2023

  13. [13]

    In: Proceedings of the IEEE/CVF international conference on com- puter vision

    Liu, J., Anirudh, R., Thiagarajan, J.J., He, S., Mohan, K.A., Kamilov, U.S., Kim, H.: Dolce: A model-based probabilistic diffusion framework for limited-angle ct re- construction. In: Proceedings of the IEEE/CVF international conference on com- puter vision. pp. 10498–10508 (2023)

    2023

  14. [14]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Liu, W., Ding, H.: Solving low-dose ct reconstruction via gan with local coherence. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 524–534. Springer (2023)

    2023

  15. [15]

    In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=XVjTT1nw5z

    Liu, X., Gong, C., qiang liu: Flow straight and fast: Learning to generate and trans- fer data with rectified flow. In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=XVjTT1nw5z

    2023

  16. [16]

    Medical physics44(10), e339–e352 (2017)

    McCollough, C.H., Bartley, A.C., Carter, R.E., Chen, B., Drees, T.A., Edwards, P., Holmes III, D.R., Huang, A.E., Khan, F., Leng, S., et al.: Low-dose ct for the detection and classification of metastatic liver lesions: results of the 2016 low dose ct grand challenge. Medical physics44(10), e339–e352 (2017)

    2016

  17. [17]

    In: The Fourteenth International Conference on Learning Representa- tions (2026),https://openreview.net/forum?id=QGd34p02mI

    Pourya, M., Rawas, B.E., Unser, M.: FLOWER: A flow-matching solver for inverse problems. In: The Fourteenth International Conference on Learning Representa- tions (2026),https://openreview.net/forum?id=QGd34p02mI

    2026

  18. [18]

    In: The Fourteenth International Confer- ence on Learning Representations (2026),https://openreview.net/forum?id= YE5scJekg5

    Shi, J., Pelt, D.M., Batenburg, K.J.: DM4CT: Benchmarking diffusion models for computed tomography reconstruction. In: The Fourteenth International Confer- ence on Learning Representations (2026),https://openreview.net/forum?id= YE5scJekg5

    2026

  19. [19]

    Song,J.,Meng,C.,Ermon,S.:Denoisingdiffusionimplicitmodels.In:International Conference on Learning Representations (2021),https://openreview.net/forum? id=St1giarCHLP

    2021

  20. [20]

    In: International Conference on Learning Representations (2023),https://openreview.net/forum?id=9_gsMA8MRKQ

    Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion mod- els for inverse problems. In: International Conference on Learning Representations (2023),https://openreview.net/forum?id=9_gsMA8MRKQ

    2023

  21. [21]

    In: Interna- tional Conference on Learning Representations (2021),https://openreview.net/ forum?id=PxTIG12RRHS

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Interna- tional Conference on Learning Representations (2021),https://openreview.net/ forum?id=PxTIG12RRHS

    2021

  22. [22]

    Optics express24(22), 25129–25147 (2016)

    VanAarle,W.,Palenstijn,W.J.,Cant,J.,Janssens,E.,Bleichrodt,F.,Dabravolski, A., De Beenhouwer, J., Joost Batenburg, K., Sijbers, J.: Fast and flexible x-ray tomography using the astra toolbox. Optics express24(22), 25129–25147 (2016)

    2016

  23. [23]

    Insights into imaging4(3), 391–397 (2013)

    Willemink, M.J., Schilham, A.M., Leiner, T., Mali, W.P.T.M., de Jong, P.A., Budde, R.P.: Iterative reconstruction does not substantially delay ct imaging in an emergency setting. Insights into imaging4(3), 391–397 (2013)

    2013

  24. [24]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ying, X., Guo, H., Ma, K., Wu, J., Weng, Z., Zheng, Y.: X2ct-gan: reconstructing ct from biplanar x-rays with generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10619– 10628 (2019)

    2019

  25. [25]

    Advances in Neural Information Processing Systems37, 57389–57417 (2024)

    Zhang, Y., Yu, P., Zhu, Y., Chang, Y., Gao, F., Wu, Y.N., Leong, O.: Flow priors for linear inverse problems via iterative corrupted trajectory matching. Advances in Neural Information Processing Systems37, 57389–57417 (2024)

    2024

  26. [26]

    In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference

    Zhang, Y., Liu, Z., Li, Z., Li, Z., Clark, J.J., Si, X.: Decoupling training-free guided diffusion by admm. In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference. pp. 23292–23302 (2025) Efficient Flow Matching for Sparse-View CT Reconstruction 11 5 Appendix: Proof of Proposition 1 Proof.LetvR d →R d be a velocity field. Define the ex...

    2025