acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Dual Velocity Field Learning along Reconstruction Trajectories for Insufficient-Angle CT

Abstract

Diffusion models and flow-based generative models, most notably flow matching (FM), have recently delivered strong performance in computed tomography (CT) reconstruction. However, for the same scanning system whose acquisition geometry varies in the number of projection angles or in the angular range, existing methods still require retraining a dedicated model for each configuration and thus offer limited cross-geometry adaptability. Moreover, we observe that the quantity learned during intermediate sampling—the noise predictor or score function in diffusion or the velocity field in flow models—is essentially independent of the current state along the sampling trajectory: in linear FM, for instance, the conditional velocity is a constant determined only by the two endpoints, . Such a position-independent field is fundamentally implausible for the intrinsically nonlinear transport from projection data to images. In this paper, we therefore organize reconstructions obtained under different numbers of projections as an ordered reconstruction trajectory, and learn along this trajectory a dual velocity field consisting of a current (geometry-transition) velocity that describes the local evolution between successive angle settings and a target velocity that points toward the clean image. An adaptive velocity-learning mechanism further conditions the transport on the current state and its position along the trajectory, and the two velocities drive a progressive two-stage reconstruction network. A single trained model thereby generalizes across sparse- and limited-angle settings without retraining, achieving high-precision quantitative imaging on LoDoInd, Mayo, and PCB benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Computed tomography reconstruction, dual velocity field, adaptive velocity learning, quantitative imaging

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AXWYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a method to adapt flow matching for sparse-angle or limited-angle CT to use the number of angles or the angular range as time variable during velocity field estimation training. The resulting method has two stages, the first producing two velocity fields (one directly pointing towards the ground-truth image, and one pointing towards the reconstruction one time-step further, i.e. with a few more angles or slightly larger angular range). Results show good performance on three different datasets, including medical and industrial datasets.

Critical strengths

Training a Flow Matching algorithm with 'number of angles' or 'angular range' as the time variable is interesting, and the results included show that predicting two velocity fields can indeed make it work. These could be valuable contributions to the field.

Critical weaknesses
  • The main motivation of the paper is unclear and not valid in my opinion. The authors claim that existing flow-based and diffusion-based methods for CT are trained for a fixed geometry, but this is not true as far as I know. These method often pretrain a general generative model (denoising over time) using high-quality references and then use additional data fitting terms during generation to steer the generated image towards realistic images that fit the data. This approach is not trained for a fixed geometry, which is exactly one of the advantages of it. The second motivation, stating that linear trajectories are problematic, is unclear and not valid as well (and only true for rectified flow in any case), as the data fitting steps in between the diffusion or flow steps push the trajectory off its linear path. The paper should make a clearer case for why existing methods fall short, and how their method differs.
  • The experimental setup is severely lacking in comparison methods. As a minimum, there should be comparisons with diffusion and flow-based methods that include data fitting, as these directly compare with the proposed method (see, for example, DM4CT for an overview of diffusion-based methods). Additionally, results should be compared with more baseline methods, including for example INR-based methods and/or gaussian splatting based methods,
  • Overall the structure of the paper makes it difficult to read. Section 2.1 is almost a copy of the introduction, with a lot of repetitive text. Figure 2 is difficult to read without a clear caption that helps the reader through it. In Figure 1 it is not clear what is shown in the figure (what kind of trajectories, and in which space?), and the caption does not clarify things. Some claims are unclear: for example, the paper claims that the method generalizes across sparse- and limited-angle settings, but from the experiments it is not clear whether this is the case.
  • Experimental details are missing, and sometimes not optimal. The used hyperparameters, how these where chosen, and how they affect results are not clearly stated. The Mayo dataset train/test split seems to be over slices and not over patients, as would be more correct. The experiments are all based on simulated noise-free data, which makes it difficult to assess how the method would perform in practice. Computation times and training times are not reported.
Questions for authors
  • How does your approach compare with diffusion and flow-based method that include data-fidelity terms?
  • How does your approach handle measurement noise?
  • How does your method perform on real-world data?
Reviewer_qZo8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is motivated by the two main challenges in existing CT reconstruction methods: (1) the need to retrain models for different acquisition geometries, and (2) the mismatch between the Flow Matching straight-line conditional paths and the CT nonlinear evolution. To address these challenges, the paper proposes Adaptive Dual Velocity Field (ADVF) learning, which defines a nonlinear evolving trajectory on a series of CT reconstructions obtained under different angular sampling configurations. Along this trajectory, the authors introduce two velocity components: a target velocity pointing toward the ground-truth image, and a local velocity describing the transition between consecutive geometry levels. A transition module then uses these predicted velocities to synthesize the missing intermediate reconstruction states and aggregates the completed trajectory to produce the final reconstructed image. Experiments on three datasets report improvements over several baseline methods.

Critical strengths

The observation that CT reconstructions evolve nonlinearly as angular sampling changes is interesting, and provides an intuitive motivation for trajectory-aware reconstruction.

Critical weaknesses

Imprecise presentation of Flow Matching. The paper argues that FM learns position-independent velocity fields because the conditional velocity along a linear interpolation is constant. This confuses the conditional velocity target with the learned marginal velocity field. Although the velocity along an individual linear conditional path can be constant, the learned field (v_\theta(x_t,t)) generally depends on both state x_t and time t, and its ODE trajectories need not be straight lines. Furthermore, the original FM framework is not restricted to linear conditional paths. Therefore, the theoretical motivation that the conventional FM cannot model nonlinear transport is not justified.

Unconvincing motivation on nonlinear trajectories. Figure 1 suggests that FM is unsuitable for CT reconstruction, because CT images evolve nonlinearly as angular sampling changes, and FM is defined on a linear transport trajectory. However, the evolution of reconstructions across acquisition geometries is fundamentally different from probability transport trajectories in FM. FM can operate in either image or latent space, and a nonlinear trajectory in image space could be a linear transport trajectory in latent space. Therefore, nonlinear angular reconstruction trajectories do not inherently imply a limitation of FM.

Necessity of v1 is not clear. The local velocity (v_i^{(2)}) already defines a progressive evolution. It is unclear what additional information the target velocity (v^{(1)}) contributes. Although global target guidance may be beneficial, this requires empirical ablation analysis. A strong ablation comparing local-only, target-only, and dual-velocity formulations is necessary.

Necessity of f_k is not clear. Since the model already predicts velocities, why are additional convolutional modules (f_k) needed to evolve the trajectory? A natural alternative is to directly update the reconstruction using locally predicted velocities. The authors should justify this design and compare it with direct velocity-based iterative updates.

Essential ablation missing. The experiments do not isolate the contributions of the two velocities, the transition modules, or the progressive architecture. At minimum, the authors should compare:

  • Target velocity only vs. local velocity only vs. dual velocity.
  • Reconstruction with and without the learned transition modules f_k.

Presentation issues. The paper contains too many dashes, bold and italics, and unnecessarily long sentences. Figure 1 is overloaded and does not clearly distinguish different types of trajectories. Overall, the paper is difficult to follow.

Questions for authors

See above.

Reviewer_24K1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ADVF for sparse- and limited-angle CT reconstruction. It organizes reconstructions across acquisition geometries into an ordered trajectory, predicts both local geometry transitions and target-directed residuals, and combines trajectory completion with final reconstruction.

Critical strengths
  1. Practical significance. Supporting different angular acquisition settings with one reconstruction model addresses a meaningful deployment challenge.

  2. Coherent methodological intuition. Geometry-dependent reconstruction trajectories provide physically motivated intermediate supervision. Combining local transitions with target-directed recovery is a clear and potentially useful idea.

Critical weaknesses
  1. The central argument about flow matching is incorrect as stated. A constant target velocity along an individual conditional path does not imply a state-independent learned velocity field or parallel transport.

  2. Inference-time input availability is unclear. ADVF consumes a prefix of reconstructions from multiple geometries. These must all be obtainable from the currently acquired measurements. Different sparse-view angle grids are not necessarily nested; generating the prefix from unmeasured projections would introduce additional information and undermine the evaluation.

  3. The source of the gains is unresolved. Multiple reconstruction inputs, intermediate supervision, progressive refinement, and final fusion could explain the improvements without requiring dual velocities.

  4. Unseen-geometry generalization is not demonstrated clearly. The reported results do not distinguish held-out acquisition geometries from configurations included in training.

Questions for authors
  1. Are all inference-time prefix reconstructions generated exclusively from the currently acquired projections? Please clarify angle-set nesting, particularly for sparse-view CT, and whether all baselines receive equivalent measurement information.

  2. Under matched inputs, capacity, and training budgets, how much benefit comes specifically from dual-velocity learning—particularly the local velocity—over ordinary progressive reconstruction with intermediate supervision?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.