acceptodds
Withdrawn from ICLR 2027

FlowSeed: Continuous-Time Contrast-Enhanced CT Synthesis via Organ-Aware Gaussian-Weighted Residual Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multiphase contrast-enhanced computed tomography (CECT) sparsely samples a continuous, organ-dependent enhancement process. However, existing synthesis methods treat arterial, venous, and delayed scans as discrete targets and reconstruct anatomy already available in non-contrast CT (NCCT). We introduce FlowSeed, a continuous-time residual contrast-rendering framework that retains NCCT as the anatomical base and renders only its time-dependent enhancement. FlowSeed uses soft organ maps to route time-conditioned affine modulation at each voxel. A slice-aware Gaussian residual decoder then predicts intensity, opacity, axis-aligned scale, and center offset to render enhancement onto queried axial slices. Physiology-inspired ordinal constraints regularize liver and aortic enhancement, while a conditional PatchGAN provides adversarial supervision for local appearance fidelity. On held-out MCT-LTDiag studies, FlowSeed reduces mean absolute error by 6.7–18.8% relative to the strongest comparator across three anatomical groups. On external WAW-TACE data, it achieves the lowest mean absolute error and highest SSIM in all groups. Ablations support the contributions of organ-aware temporal routing and Gaussian-weighted residual decoding. Dense-query visualizations further show that FlowSeed can synthesize continuous enhancement sequences with distinct organ-dependent responses across time.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Contrast-enhanced CT synthesis, Continuous-time generative modeling, Medical image synthesis, Gaussian-weighted residual splatting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PEC2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

Thia paper proposes a FlowSeed framework to synthesize contrast-enhanced CT from non-contrast CT (NCCT) as NCCT plus a non-negative residual, conditioned on scalar post-injection time. Time modulates the features per organ through a predicted soft organ map, and a “Gaussian-weighted residual splatting” decoder renders the residual slice by slice. It adds an ordinal liver/aorta loss and a PatchGAN. It reports the best aggregate metrics on MCT-LTDiag and the best MAE and SSIM on external WAW-TACE.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The residual formulation is sensible. Anatomy is preserved by construction, and time is a more natural conditioning variable than a phase label.
Critical weaknesses
  1. The continuous-time claim is not well validated. There is no ground truth exists at intermediate times. There is no ablation of the same architecture with categorical phase or fixed phase-centre times, so gains from continuous time cannot be separated from gains from architecture.

  2. Enhancement agreement is weak. Pooled Mean-HU CCC is 0.27–0.32, and within-phase values are far lower (abdominal vessels 0.120/0.063/0.132 against 0.320 pooled). The pooled figure therefore mostly reflects between-phase differences, not patient-specific prediction.

  3. Trivial baselines are missing. Parenchymal MAE (42.5 HU) is comparable to liver enhancement itself (about 50 HU in Fig. 3). An identity baseline and an “NCCT + per-organ, per-phase mean enhancement” baseline are needed. Pix2Pix at 163 HU is almost certainly worse than identity, which suggests the baselines were poorly adapted.

  4. The proposed FlowSeed is trained with organ-masked losses and gating and evaluated only inside those organs. It also receives recorded time while the baselines receive only phase.

  5. The external vessel results need explaination. For example, on WAW-TACE, abdominal-vessel MAE is 162.6 HU with +121 HU bias and CCC 0.058; in the delayed phase, MAE is 232 HU. SSIM is higher than on the internal set, and external parenchymal MAE (30.1) beats internal (42.5).

  6. The time labels are questionable. UCSF times are regressed from the HU statistics of the target scans themselves, with a 30s delayed-phase MAE. The regressor’s training data is unspecified, and renal protocols may fall outside its range. More fundamentally, time since injection does not determine enhancement without the injection protocol, bolus tracking and cardiac output.

  7. The split of train/test is provided, it is unclear how the loss weights, margin and stopping point were chosen.

Questions for authors
  1. What are t_A, t_V, t_D in Eq. 11: fixed reference times or per-study recorded times?
  2. Were segmentation masks derived from NCCT or from the enhanced scans?
  3. What were the registration QC criteria, and where were the “preregistered” surrogates and “frozen” protocol registered?
Reviewer_iU5dOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This method proposes the FlowSeed, the first work to explore continuous-time, whole-abdomen CECT synthesis from NCCT, treating post-injection time as a directly queryable input. However, it doens't clearly describe the clinical motivation (why continous-time?) or the method motivation (why gaussian splatting for image generation? why residual?) Also, it uses the incremental innovations such as adding different losses.

Critical strengths
  1. The paper presents ablation experiments, confidence intervals and temporal sensitivity analysis.
  2. The internal and external datasets are used for evaluation.
Critical weaknesses
  1. Insufficient clinical motivation: although the paper states that the method is "restrictive for delayed scans, whose timing can vary by several minutes", a single delayed scan is sufficient for clinicians. Why is continuous-time generation still necessary? Even the arterial phase alone is adequate for diagnosing certain diseases. Are there additional benefits for clinicians in lesion diagnosis or other clinical tasks?

  2. Insufficient method motivation. a. Formulating the image generation task with Gaussian splatting does not follow an established pipeline. Why was Gaussian splatting selected? Does Gaussian splatting resolve certain limitations of conventional image generation pipelines? b. Why incorporate the residual term? What weaknesses would arise if the residual component were removed?

  3. Comparison methods are not efficient or fair. a. It appears that all baseline methods perform discrete-phase generation: "The categorical baselines generate the requested arterial, venous, or delayed phase. MAFormer uses three separately trained phase-specific models", instead of continuous-time conditioning. b. The proposed method leverages segmentation masks during training. Do the other baselines also use segmentation masks? If not, the inputs differ, making the comparison unfair.

Questions for authors

Clealy describe the clinical and method motivation.

Use other fair comparison methods.

Reviewer_eD3kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FlowSeed predicts contrast-enhanced CT images from a non-contrast scan at a chosen time after injection. It uses organ maps and a Gaussian decoder to predict the added contrast, while keeping the input scan as the base. The paper tests the method on MCT-LTDiag and the external WAW-TACE dataset, and checks the effects of its main components.

Critical strengths
  1. FlowSeed has lower MAE and higher SSIM than the tested baselines in all three groups on both datasets. The external test supports its ability to work on data from a different source.
  2. Removing either organ-aware time control or the Gaussian decoder makes the results worse. Confidence intervals computed across patients support these gains. This provides evidence that both components are useful.
Critical weaknesses
  1. Evaluation mainly covers three observed phases. Fixed phase times on the external dataset do not test continuous-time generalization. Quantitative validation at unseen times is needed.
  2. Unclear benefit of exact scan times FlowSeed receives recorded scan times, while the main baselines receive only phase labels. The paper does not test FlowSeed with phase labels alone. It is therefore unclear how much of the improvement comes from the extra time information.
  3. FlowSeed uses 3D features, but the reported scores do not directly measure whether nearby slices form a consistent 3D image. A direct test of consistency across slices is needed.
  4. Removing the discriminator improves SSIM in two groups, but worsens MAE and CCC. The paper gives a possible explanation but does not show training curves or results from different random seeds. These checks would help show whether the gains are stable across training runs.
Questions for authors

See Weakness.

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

Related papers

Discussion (0)

Sign in to comment.