acceptodds
Under review as a conference paper at ICLR 2027

KineWorld: Latent Kinetic World Modeling for Contrast-Free DCE-MRI Synthesis

Abstract

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provides clinically informative enhancement trajectories, yet contrast-agent administration is unsuitable for some patients. Synthesizing patient-specific DCE-MRI trajectories from a pre-contrast scan requires preserving anatomical consistency while faithfully modeling tissue-specific contrast kinetics. Existing methods do not explicitly represent evolving kinetic states, often resulting in structural drift and physiologically implausible kinetic trajectories. We propose KineWorld, a latent kinetic world model that anchors cumulative contrast-dependent changes to the pre-contrast latent state, thereby separating kinetic evolution from time-invariant anatomy to preserve anatomical consistency. To achieve kinetic fidelity, a pharmacokinetics-inspired rollout recursively integrates learned influx and efflux velocities, while kinetics-aware retrieval supplies relevant yet diverse long-range patient history to preserve phase-dependent evolution. Relay-guarded self-rollout training further limits accumulated anatomical and kinetic errors by selectively restoring ground-truth states upon excessive deviation. Across three breast and prostate DCE-MRI datasets, KineWorld improves SSIM by over 11% and lesion kinetic curve CCC by over 17% relative to the strongest baselines. A blinded reader study further demonstrates substantially improved agreement with report-derived kinetic categories. Source code, pretrained weights, and our curated internal dataset will be publicly released to facilitate future research.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Synthesis, Dynamic Contrast-Enhanced MRI, World Models, Pharmacokinetic Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SAkTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes KineWorld, a latent kinetic world modeling framework for synthesizing DCE-MRI sequences from pre-contrast scans. It combines pharmacokinetics-inspired latent rollout, kinetics-aware retrieval, and relay-guarded training to model temporal enhancement dynamics while preserving anatomical consistency. Experiments on three DCE-MRI datasets compare the method with image-to-image and video generation baselines in terms of image quality and kinetic fidelity, along with ablation studies and a radiologist reader study.

Critical strengths
  1. The paper focuses on a clinically meaningful problem of synthesizing DCE-MRI sequences from pre-contrast scans, with an emphasis on anatomical consistency and temporal enhancement dynamics.
  2. The method is evaluated on three datasets against multiple image-to-image and video generation baselines, with reported improvements in image quality and kinetic fidelity. Additional ablation studies and a radiologist reader study provide further empirical evidence.
Critical weaknesses
  1. The proposed framework lacks explicit action-conditioned dynamics or controllable interventions. Although temporal embeddings are incorporated, time serves as a temporal coordinate rather than an action that controls how the underlying system evolves. Therefore, the justification for characterizing KineWorld as a world model remains insufficient.
  2. KineWorld essentially performs autoregressive latent state prediction conditioned on historical states and timestamps. The paper does not establish what fundamental modeling capabilities, as a world model, distinguish this approach from conventional time-conditioned video prediction.
  3. The proposed influx and efflux terms are learned in an unconstrained latent space and do not correspond to measurable pharmacokinetic quantities. As acknowledged in Appendix A.1, their relationship to the Tofts model is structural rather than physical. It remains unclear whether the proposed formulation genuinely captures pharmacokinetic mechanisms or simply introduces a flexible parameterization of latent state transitions.
  4. The paper claims to support state prediction at arbitrary timestamps through dense temporal rollout. However, quantitative evaluations are primarily performed at observed acquisition timestamps. Experiments on held-out intermediate timestamps are needed to verify whether the model learns continuous-time dynamics rather than fitting the observed temporal sequence.
  5. The reported PSNR, SSIM, and LPIPS primarily measure reconstruction agreement with reference images rather than directly quantifying anatomical stability across generated frames. More explicit evaluations of lesion morphology preservation and temporal structural drift would better support the claimed anatomical consistency.
  6. Although KineWorld uses physical-time embeddings and integrates latent velocities on a one-second temporal grid, this does not by itself establish a consistent continuous-time dynamical model. The learned transitions are supervised primarily at sparsely observed acquisition timestamps, and the paper does not demonstrate whether the predicted state at a given physical time remains consistent under different integration step sizes or rollout discretizations. Therefore, the claimed continuous-time modeling capability requires stronger justification.
Questions for authors
  1. What fundamentally distinguishes KineWorld from conventional time-conditioned autoregressive video prediction? Without explicit action-conditioned dynamics or controllable interventions, what specific capabilities justify its characterization as a world model?
  2. How can the authors justify that the learned influx–efflux decomposition captures meaningful pharmacokinetic dynamics rather than an arbitrary parameterization of latent state transitions? Would a generic latent residual predictor achieve comparable performance?
  3. Can the authors provide evaluations at held-out intermediate timestamps and analyze the sensitivity to different integration step sizes? These experiments would help clarify whether KineWorld learns consistent continuous-time dynamics rather than simply reproducing observed temporal sequences.
5 comments

Thank you for the insightful questions. To our best understanding, world modeling aims to learn a system’s underlying dynamics and does not necessarily require dense action inputs. In DCE-MRI, these dynamics concern contrast-enhancement pharmacokinetics (PK), motivating our PK-inspired latent representation transition design. In our method, time serves as a coordinate, not a separately learned time-advance action: every queried state follows the same sequence of 1-second PK-Roll updates with its kinetic history retained. This design avoids query-dependent inconsistencies arising from alternative compositions of time-advance actions on the specified grid. The responses below clarify this rationale and its supporting evidence.

AuthorsOpenReview ↗

W1 & W2 & Q1: Definition of the world model terminology

Following CEK-World [1], we formulate contrast-free DCE-MRI synthesis as world modeling of contrast-enhancement kinetics. Contrast-agent injection is the initiating action in the underlying medical process. Unlike autonomous driving involving dense control actions, our task focuses on long-horizon, patient-conditioned latent evolution following a single injection. This shares the predictive representation spirits with V-JEPA 2’s action-free training, learning representations of the world intrinsic structure and dynamics does not require action inputs at every transition.[2]

In our setting, the core regularities are the stability of patient-specific anatomy and the temporal evolution of tissue-specific enhancement dynamics. KineWorld aims to represent both in latent space. To guide the representation learning, we translate these regularities into two core modeling targets, (i) anatomical consistency, keeping anatomy spatially and morphologically stable, (ii) kinetic fidelity, reproducing the temporally ordered enhancement evolution among wash-in, peak-enhancement, and wash-out phases.

[1] Kong, J., et al. ‘MRI contrast enhancement kinetics world model.’ CVPR 2026.

[2] Assran, M, et al. ‘V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.’ Meta AI tech report, 2025.

AuthorsOpenReview ↗

W3 & Q2: What the pharmacokinetic (PK) connection establishes

To achieve these, in contrast to conventional video generation, we design Tofts function inspired PK-Roll module to explicitly modulaet contrast agent velocities over a fixed pre-contrast anatomy. The connection to the Tofts model is structural rather than quantitative. The learned influx and efflux terms are not estimates of measurable concentrations or transfer constants. Our claim is that this structure provides an explicit modeling strategy for learning enhancement kinetics within a latent world-model framework, rather than estimating physical PK parameters.

To test whether this Tofts-inspired structure design offers benefits beyond a generic learned transition, we distinguish two comparisons. (i) Retrained generic-velocity predictor. We remove the efflux term and retrain the remaining predictor while retaining the influx Transformer. This tests whether directly learning the latent velocity achieves comparable performance without the structured damping term. (ii) Efflux disabled after full-model training. We set ke=0\textbf{k}_e=0 only during inference using the fully trained model. These comparisons assess the predictive benefit of the proposed Tofts-inspired structure. The results are updated in Sec. 3.4.

DukeProstateX
SSIM ↑\uparrowPSNR ↑\uparrowCCC ↑\uparrowS-MAE ↓\downarrowSSIM ↑\uparrowPSNR ↑\uparrowCCC ↑\uparrowS-MAE ↓\downarrow
Directly learning velocity vt\mathbf{v}_t0.744824.660.734088.930.828224.710.841669.85
ke=0\mathbf{k}_e=00.741324.320.750270.470.825124.680.884545.76
Ours0.756826.060.801557.890.844026.100.941225.44
Reviewer_pjcbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes KineWorld, a latent kinetic world model, to synthesizes patient-specific DCE-MRI trajectories from a pre-contrast scan. It anchors contrast-dependent changes to the pre-contrast latent state to preserve anatomical consistency. A pharmacokinetics-inspired state rollout and KA-Retrieval to achieve kinetic fidelity. Besides, it also introduces relay-guarded training to reduce error accumulation during autoregressive rollout.

Critical strengths
  1. Modeling DCE-MRI synthesis as a temporal evolution problem rather than independent image-to-image translation is well-motivated. Incorporating insights from the Tofts pharmacokinetic model is also an interesting approach.
  2. The paper trains and evaluates the method on large-scale datasets and commits to publicly releasing the data.
Critical weaknesses
  1. Synthesizing patient-specific DCE-MRI trajectories from a pre-contrast scan is not well-motivated and clinical reliable. DCE-MRI provides additional physiological information that may not be recoverable from pre-contrast MRI alone. For example, fibroadenomas and invasive carcinomas may exhibit similar pre-contrast appearances but different contrast-enhancement kinetics. Without additional patient-specific observations, it is unclear whether the model can recover true enhancement dynamics rather than predict typical patterns from pre-contrast appearance.
  2. Related to last point, similar pre-contrast appearances may correspond to different enhancement trajectories. KineWorld is a deterministic model which means it produces a single deterministic rollout, which may average together possible futures. For example, in Fig. 3(b), the ground-truth curve exhibits a plateau, whereas KineWorld predicts a washout pattern.
  3. For the influx and efflux decomposition, these two components are jointly learned without direct supervision or constraints ensuring their physiological interpretation. Many combination can give the same results. It is unclear whether the proposed PK-inspired decomposition provides benefits beyond directly learning latent velocity.
Questions for authors
  1. Although the paper already evaluated the removal of the efflux term, could the authors compare against a generic latent velocity predictor to better isolate the benefit of the proposed influx–efflux decomposition?
  2. Regarding Weakness 2, how does the model handle this ambiguity, particularly for lesions with atypical kinetics?
  3. The current qualitative examples are most persistent and plateau pattern. Could the authors provide more qualitative examples (kinetic curves) for cases with clear washout? A subtype-stratified quantitative analysis would also help assess whether the model reliably captures different kinetic patterns.
6 comments

Thank you for recognizing the motivation for modeling DCE-MRI synthesis as temporal evolution, our Tofts-inspired design, and our large-scale evaluation and commitment to data release. We try our best to address your comments in detail below.

AuthorsOpenReview ↗

W1: Clinical motivation for contrast-free DCE-MRI synthesis

For DCE-MRI, contrast administration is unsuitable for patients with hypersensitivity reactions, pregnant, and with renal impairment etc. Thus, DCE-MRI synthesizing from pre-contrast MRI is a trend of research direction. For example, studies in Radiology have investigated breast MRI synthesis from non-contrast inputs and contrast-reduction strategies [1,2,3]. The recent MICCAI 2026 MAMA-SYNTH challenge also explicitly benchmarks post-contrast breast MRI synthesis from a single DCE-pre image [4]. These efforts support the clinical relevance of investigating this challenge.

[1] Chung M, et al. ‘Deep Learning to Simulate Contrast-enhanced Breast MRI of Invasive Breast Cancer.’ Radiology, 2023.

[2] Müller-Franzes G, et al. ‘Using Machine Learning to Reduce the Need for Contrast Agents in Breast MRI through Synthetic Images.’ Radiology, 2024.

[3] Zhang, Y, et al. ‘Development and validation of a deep learning–enabled single breath-hold abbreviated MRI protocol for hepatocellular carcinoma diagnosis.’ Radiology: Artificial Intelligence, 2026.

[4] MAMA-SYNTH Challenge. MICCAI 2026.

AuthorsOpenReview ↗

W2 & Q2: Ambiguity of pre-contrast images and atypical kinetics

Thank you for raising the concern that similar pre-contrast appearances may correspond to different enhancement trajectories, causing a deterministic model to produce compromise predictions, particularly for lesions with atypical kinetics. We address this from three perspectives: (1) the predictive imaging features is available in pre-contrast imaging, (2) how our work leverages individual imaging features for patient-conditioned prediction based on our heterogeneous training data, and (3) a linear probing analusis of pre-contrast representations.

(1) The predictive imaging features is available in pre-contrast imaging. We agree that an individual lesion enhancement kinetics depend on all related physiological factors that is not fully captured by a pre-contrast scan. However, incomplete information should not be equated with an absence of predictive information. Non-contrast T1w (DCE-pre phase) radiomics has demonstrated partial discrimination breast lesions diagnosis [1,2], supporting the presence of informative anatomical and textural cues, although this does not establish that individual enhancement trajectories are uniquely recoverable.

[1] Zhang, Y, et al. ‘Development and validation of a deep learning–enabled single breath-hold abbreviated MRI protocol for hepatocellular carcinoma diagnosis.’ Radiology: Artificial Intelligence, 2026.

[2] Liu Y, et al. ‘Research on the application of distinguishing between benign and malignant breast nodules using MRI and US radiomics’. Frontiers in Oncology. 2025.

(2) How our work leverages individual imaging features for patient-conditioned prediction. Visually similar pre-contrast appearances need not be identical in their underlying imaging features, and deterministic prediction does not imply a population-average trajectory. KineWorld conditions each prediction on the individual scan, combining encoded imaging features with a PK-inspired transition structure to learn appearance-dependent enhancement dynamics. Nevertheless, when the available features cannot distinguish different underlying physiological states, a deterministic model may still produce a compromise prediction.

Beside method design, our Internal dataset contains 3,491 DCE-MRI series with paired pre-contrast and post-contrast acquisitions, spanning pathology-confirmed benign and malignant lesions. Its training partition exposes the model to heterogeneous relationships between pre-contrast appearance and subsequent enhancement, providing a basis for learning different patient-conditioned trajectories rather than imposing a single canonical pattern. Table 2 reports the lesion-specific kinetic agreement across all three datasets. Appendix Table 6 further provides lesion-region evaluation separately by pathology. KineWorld achieves CCC of 0.8465 for benign and 0.7462 and malignant lesions, outperforming the evaluated baselines in both groups.

Benign Lesion (37%)Malignant Lesion (63%)
CCC ↑S-MAE ↓CCC ↑S-MAE ↓
CEK-World0.4250110.80.2215175.2
KineWorld0.846535.400.746257.21

The blinded reader study provides complementary evidence of agreement with report-derived kinetic categories. Together, these findings support predictive utility across the evaluated lesion groups.

(3) Representation-level assessment. To assess whether the learned representation retains pathology-associated information, we will additionally perform linear probing with the trained KineWorld parameters frozen. Using only pre-contrast unilateral breast from the internal dataset, we will extract features from frozen encoder features and train a linear logistic-regression classifier to distinguish benign from malignant cases.

AUCAcc
Ours Linear Probing0.720.75
DCE-Pre Diagnosis (Figure 9 - P0) [3]0.74NA

[3] Li, Y, et al. ‘A deep learning system for non-invasive breast cancer diagnosis with multimodal data.’ Nature Biomedical Engineering, 2026.

This result of malignant diagnosis above suggests that i) pre-contrast representations contain discriminative information, and ii) KineWorld indeed has the potential to leverage these imaging cues for patient-conditioned predictions rather than relying solely on population-average patterns.

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

This paper introduces KineWorld, a latent kinetic world model for synthesizing patient-specific DCE-MRI trajectories from a pre-contrast scan without contrast administration. The method combines a pharmacokinetics-inspired rollout operator, long-range kinetics-aware retrieval, and relay-guarded training. KineWorld outperforms image-to-image and video-generation baselines across three datasets.

Critical strengths
  1. The proposed method shows strong empirical results and outperforms the baselines in Tables 1 and 2.
  2. Anchoring the rollout to the pre-contrast latent and predicting only a contrast-dependent residual is a reasonable inductive bias for a setting where the anatomy remains stationary.
  3. The Relay-Guard ablation includes both extremes (no relay and always relay with τ_d = 0), providing a useful pair of controls.
  4. The evaluation is broad and covers a good range of settings.
Critical weaknesses
  1. Relay-Guard supervises corrections only for deviations within the trust region. Beyond τ_d, the drifted state is replaced with the ground truth, and the memory bank is cleared. As a result, states far from the trajectory are never presented to the model as inputs, so the model is not trained to recover from them. Fig. 7(b) shows that roughly 10% of acquisition points still trigger a relay at the end of training. At inference, however, there is no relay, and the paper does not report how often the rollouts leave the region.

  2. The paper cites scheduled sampling, DAgger, and professor forcing but does not clearly compare Relay-Guard with these methods. There is also no control that separates the effect of state-triggered resets from the reset rate itself.

  3. The paper claims that KineWorld "explicitly captures inter-phase dependencies," but no post-contrast phase is ever observed during inference. Every state in the chain is the model’s own prediction, so the rollout is a function of the pre-contrast latent and acquisition times alone. The sequential rollout therefore seems more like an inductive bias than a capability that direct image-to-image methods inherently lack. This should be tested against a direct time-conditioned predictor using the same backbone.

  4. The authors describe "state prediction at arbitrary timestamps" as a core capability of the method. However, this capability is not evaluated in the experiments.

  5. The paper states that the Duke dataset contains 3-4 post-contrast phases. However, Figure 7(a) reports results for the Duke dataset across 5 post-contrast phases. It is unclear how the fifth phase was evaluated given this inconsistency.

Questions for authors
  1. How often do test-time rollouts leave the trust region, and how do these cases perform?
  2. How does Relay-Guard compare with scheduled sampling using a reset rate matched to Fig. 7(b)?
  3. At inference time, is the generated trajectory a deterministic function of only the pre-contrast latent and the acquisition times?
9 comments

Thank you for recognizing our strong empirical performance, the anatomy-anchored inductive bias, the useful Relay-Guard ablation controls, and the breadth of our evaluation. We address your comments in detail as below.

AuthorsOpenReview ↗

W1: Supervision on the relay-triggered states

We clarify that predictions outside the trust region remain supervised, and the memory bank is not completely cleared.

(i) Supervision is retained before replacement. Predictions, including those with excessive deviations, are supervised by the image and latent reconstruction loss (Eq. 4):

Lrecon=1N∑i=1N(∥z^i−zi∥1+∥I^i−Ii∥1)L_{recon}=\frac{1}{N}\sum_{i=1}^{N}(\Vert\hat{z}_i-z_i\Vert_1+\Vert\hat{I}_i-I_i\Vert_1)

The smoothness loss (Eq. 5) regularizes kinetic acceleration throughout the rollout:

Lsmooth=1N−1∑i=1N−1∥(v^i+1−v^i)/Δt∥1L_{smooth}=\frac{1}{N-1}\sum_{i=1}^{N-1}\Vert(\hat{v}_{i+1}-\hat{v}_i)/\Delta t\Vert_1

States exceeding the deviation threshold additionally receive the relay penalty (Eq. 8) before any replacement:

Lrelay=1N∑i=1Nρi deviL_{relay}=\frac{1}{N}\sum_{i=1}^{N}\rho_i\,dev_i

Replacement occurs only after these losses are evaluated, so the transition producing the deviation remains penalized rather than being discarded from supervision.

(ii) The memory bank is selectively updated, not cleared. Only the excessively deviated state and its corresponding memory entry are replaced with the reference state. All other historical entries remain unchanged. This prevents subsequent transitions from conditioning on that severely deviated entry while retaining the remaining self-generated history, including moderate prediction errors.

Thus, the transition producing the deviation remains supervised, while subsequent training is protected from the severely contaminated memory entry.

AuthorsOpenReview ↗

Q1: The rollout behavior during inference.

Using the same threshold τd\tau_d, KineWorld test rollouts yield threshold-exceedance rates on Duke/Internal/PROSTATEx, respectively, compared with terminal training rates of 12.2%/11.9%/9.4%. The modest increases suggest relatively stable behavior from training to autonomous inference under this diagnostic, despite the removal of training-time corrections. Nevertheless, error accumulation remains a challenge: Relay-Guard mitigates, rather than eliminates, excessive deviations. Existing training strategies balance this challenge differently, ranging from fully self-generated histories (“No relay,” on-policy training) to reference-conditioned histories (“Always relay,” teacher-forced training); Relay-Guard instead uses selective, deviation-triggered intervention.

Using the same threshold τd\tau_d, the test-time rollout of KineWorld yield rates of 13.5%/13.4%/9.8% on Duke/Internal/ProstateX, respectively, comparable to the terminal training rate 12.2%/11.9%/9.4% on Duke/Internal/ProstateX. This suggests that removing training-time corrections does not cause a marked increase in excessive latent deviations, supporting relatively stable behavior from training to autonomous inference under this diagnostic. Fig. 7(a) shows that Relay-Guard achieves lower late-phase kinetic errors than both “No relay” and “Always relay,” supporting selective intervention over either extreme. Under the same autonomous inference protocol, we also reported their exceedance rates as in the table below.

DukeInternalProstateX
Ours(Terminal training relay)12.2%11.9%9.4%
Ours(Test time relay)13.5%13.4%9.8%
Always relay(Test time relay)17.5%16.2%15.6%
No relay(Test time relay)26.7%26.4%18.6%

Performance on threshold-exceeding cases. To further alleviate this concern, we assess whether relay ratio correspond to substantial performance degradation. We compare KineWorld on out-of-trust cases with its overall performance and CEK-World, as reported below.

DukeInternalProstateX
SSIM ↑\uparrowPSNR ↑\uparrowCCC ↑\uparrowS-MAE ↓\downarrowSSIM ↑\uparrowPSNR ↑\uparrowCCC ↑\uparrowS-MAE ↓\downarrowSSIM ↑\uparrowPSNR ↑\uparrowCCC ↑\uparrowS-MAE ↓\downarrow
Ours0.756826.060.801557.890.740724.610.783349.140.844026.100.941225.44
Relay triggered cases0.714324.470.732688.750.690822.490.729886.330.809724.730.886533.68
CEK-World0.560321.280.3275166.20.621621.100.2968151.30.758924.090.468286.71

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks for your suggestions and comments.

We thank all reviewers for recognizing the value of our pharmacokinetics-inspired (PK) modeling and strong empirical evaluation. Your constructive comments address complementary aspects of the work, including its additional model analysis, clinical motivation and world-model formulation. These comments provide an opportunity to clarify our contributions, strengthen the supporting evidence, and better define the scope and limitations of our claims. Detailed responses are as follow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.