Breast DCE-MRI · Generative Modeling · Clinical Validation

Dense Temporal Contrast Synthesis via Conditioned Latent Transport

Smriti Joshia,*, Apostolia Tsirikoglouc,†, Daniel M. Langd,e,†, Richard Osualaa,g,†, Noah MΓ‘rquez Varaa,h, Alejandro Guzmana, Grzegorz Skorupkoa, Sebastian Ibarra Arreguia, Lidia Garruchoa, Akane Ohashii,j, Dimitra Ntoulak, Eugen Divjakl,m, Oğuz LafcΔ±n, Jan C. Peekeng, Julia A. Schnabeld,e,f, Fredrik Strandc, Oliver Diaza, Karim Lekadira,b

*Corresponding author  ·  These authors contributed equally to this work.

Affiliations
  1. a Departament de MatemΓ tiques i InformΓ tica, Universitat de Barcelona, Barcelona, Spain
  2. b InstituciΓ³ Catalana de Recerca i Estudis AvanΓ§ats (ICREA), Barcelona, Spain
  3. c Department of Oncology-Pathology, Karolinska Institutet, Stockholm, Sweden
  4. d Institute of Machine Learning in Biomedical Imaging, Helmholtz Munich, Munich, Germany
  5. e School of Computation, Information and Technology, Technical University of Munich, Munich, Germany
  6. f School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom
  7. g Department of Radiation Oncology, TUM University Hospital Rechts der Isar, TUM School of Medicine and Health, Technical University of Munich, Munich, Germany
  8. h Department of Computer Science and Engineering, Chalmers University of Technology, Gothenburg, Sweden
  9. i Department of Translational Medicine, Diagnostic Radiology, & CIRCE – the Center for Interdisciplinary Research on Cancer and Equity in Women, Lund University, Lund, Sweden
  10. j Department of Imaging and Physiology, SkΓ₯ne University Hospital, MalmΓΆ, Sweden
  11. k Department of Radiology, Karolinska University Hospital, Stockholm, Sweden
  12. l University of Zagreb, School of Medicine, Zagreb, Croatia
  13. m University Hospital Dubrava, Zagreb, Croatia
  14. n Department of Biomedical Imaging and Image-Guided Therapy, Medical University of Vienna, Vienna, Austria

Abstract

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is essential for breast cancer management, but reliance on gadolinium-based contrast agents (GBCAs) restricts use in contraindicated populations, prolongs scan protocols, and presents environmental toxicity concerns. Contrast synthesis offers a non-invasive alternative; however, existing approaches struggle to balance spatial realism with temporal continuity, suffer from slow iterative sampling, underutilize structural priors, and lack clinical validation. We propose a novel conditioned latent transport framework that predicts contrast enhancement in a single forward pass. By anchoring the latent trajectory to the pre-contrast anatomy and applying continuous time conditioning, the model synthesizes patient-specific contrast evolution at any given acquisition time. The proposed approach outperforms baseline and the state-of-the-art models across spatial, perceptual, temporal, and distributional metrics. Evaluated on an independent external cohort, the method demonstrates robustness to domain shifts induced by scanner noise as well as differing acquisition protocol. Furthermore, our synthetic contrast enhancement significantly improved downstream tumor segmentation performance, yielding a 22.4% relative increase in Dice coefficient (0.60 vs. 0.49 baseline pre-contrast, p < 0.01), reducing boundary segmentation error by over 39%, while outperforming all other generative model baselines. Finally, a reader study involving four breast radiologists evaluated the image quality, kinetic fidelity, and diagnostic viability of our synthesized sequences across 40 randomly selected cases. The results demonstrated that in 70% of cases, synthesized images provided sufficient clinical information to support the same management decisions as real DCE-MRI, suggesting a path toward safer and faster contrast-free or contrast-reduced imaging workflows.

Method

Contrast enhancement is predicted in one forward pass by transporting a latent that stays anchored to the patient's own pre-contrast anatomy.

  1. 1

    Anchor the anatomy

    A frozen encoder compresses the pre- and post-contrast images to latents. zpre becomes a fixed structural anchor, so the network never re-synthesizes baseline anatomy.

  2. 2

    Predict only the change

    Conditioned on zpre, a sinusoidal time embedding τ, and the interpolation step t, a U-Net regresses the latent subtraction map αΊ‘enh ≈ zpost − zpre.

  3. 3

    Decode with two-fold supervision

    A frozen decoder returns to pixel space, where LPIPS enforces perceptual fidelity and a log-amplitude Fourier loss preserves fine micro-vascular detail.

Architecture diagram of the proposed latent generative model showing the frozen encoder, latent interpolation between pre- and post-contrast states, the conditioned U-Net target predictor, the frozen decoder, and the MSE, LPIPS and Fourier losses.
Overview of the proposed latent generative model architecture. Continuous conditioning on acquisition time τ allows synthesis at any point along the enhancement trajectory, not only at the 4–5 phases acquired in standard clinical protocols.

Result 1 of 3

Synthesis

Anchoring the latent trajectory to pre-contrast anatomy and conditioning on continuous acquisition time produces contrast evolution that is both spatially faithful and temporally smooth, in a single forward pass, with no iterative sampling. The method leads baseline and state-of-the-art comparisons across pixel-level, perceptual, distributional and temporal metrics, and holds up on an independent external cohort whose pharmacokinetic wash-in is substantially faster than the training distribution.

A Dense trajectories from sparse acquisitions

A clinical protocol acquires only 4–5 post-contrast phases. Because τ is a continuous condition rather than a phase index, the model renders the enhancement curve at any requested time, turning a handful of discrete acquisitions into a dense trajectory.

B Temporal continuity under stochasticity

Injecting latent noise restores realistic texture, but sampling it independently per phase makes the sequence flicker non-physiologically. Holding a single patient-level noise map constant across phases leaves the temporal condition as the only driver of change, and yields smooth trajectories for both CCNet and our method. The difference is what happens across seeds: anchored to the pre-contrast latent, our anatomy stays fixed and the noise models epistemic uncertainty, whereas CCNet, initialized from pure noise, varies in tumor extent and hallucinates anatomy. That distinction is what makes the per-pixel uncertainty maps used in the reader study meaningful.

Result 2 of 3

Segmentation

Does the synthesized uptake carry real biological signal, or only the appearance of it? We feed synthetic images to a downstream tumor segmentation network and compare against real pre-contrast (Baseline) and real post-contrast (Upper Bound) inputs on the MAMA-MIA internal validation set.

0.60Dice
vs. 0.49 on real pre-contrast, a +22.4% relative gain, against an upper bound of 0.63 from real post-contrast acquisitions
−39%
Boundary error: HD95 falls from 71.5 to 43.4, closer to the ground truth than the real post-contrast upper bound (47.5)
Bestof 5
Outperforms U-Net, pix2pix, CCNet and TeNCA under both training paradigms (paired Wilcoxon, p < 0.01)
Qualitative downstream segmentation results per patient case across real and synthesized images, with kernel density estimation of intensities inside the ground truth mask.
Qualitative downstream segmentation results. Each row is a distinct patient case. KDE maps (far right) show the intensity distribution within the ground truth mask. pix2pix artificially spikes intensity around the tumor; CCNet hallucinates texture and overestimates extent; U-Net underestimates it. Our model aligns closest with the ground truth in both morphological boundary and intensity distribution.
Quantitative segmentation performance
MethodDice ↑HD95 ↓
Segmentation model trained on post-contrast
Baseline0.17 (0.30)160.69 (104.48)
U-Net0.44 (0.36)82.20 (104.68)*
pix2pix0.22 (0.33)147.79 (108.24)
CCNet0.45 (0.35)76.81 (102.32)
TeNCA0.30 (0.35)122.31 (110.69)
Ours0.51 (0.37)68.78 (99.62)
Upper Bound0.68 (0.33)38.59 (77.73)
Segmentation model trained on pre-contrast
Baseline0.49 (0.37)71.48 (99.90)
U-Net0.56 (0.35)53.05 (89.01)
pix2pix0.44 (0.35)70.10 (96.55)
CCNet0.44 (0.35)78.51 (102.96)
TeNCA0.51 (0.36)62.59 (93.86)
Ours0.60 (0.33)43.38 (80.12)
Upper Bound0.63 (0.34)47.46 (84.74)

Mean (standard deviation). Significance by paired Wilcoxon signed-rank test (p < 0.01); results that did not reach significance are marked *. Success on the post-contrast-trained network shows the synthesized uptake mirrors the ground truth; success on the pre-contrast-trained network shows the synthesis preserves the underlying morphology and texture rather than overwriting it.

Result 3 of 3

Clinical Reader Study

Four board-certified breast radiologists from three academic centers (5, 8, 13 and 15 years of sub-specialty experience) evaluated 40 cases stratified by tumor size, shape, and enhancement pattern, yielding 2,927 analyzable responses across four study sections. The full study platform, with the case presentation and questionnaires exactly as the readers saw them, is public.

Which synthesis did readers prefer?

Share of 160 blinded evaluations (40 cases Γ— 4 readers) selecting each method as the superior image.

Ours 83.8% U-Net 8.1% TeNCA 8.1%

Preference for our method was consistent across all four readers and highly significant (p < 0.001); U-Net and TeNCA were statistically indistinguishable from one another (p = 1.0).

Data table
MethodVotesShare
Ours13483.8%
U-Net138.1%
TeNCA138.1%
Total160100%

What happens to clinical management?

Impact of relying on the synthetic image instead of the real acquisition, across all side-by-side assessments.

70.0%: clinical management unchanged 35.6% 34.4% 30.0% No change Minor change, no impact on management Major change

Perceived synthetic appearance (OR 3.60) and diagnostic complexity (OR 2.14) independently predicted greater clinical impact; reader experience was protective (OR 0.54). Mean characterization agreement between synthetic and real images was 69.8% (median 66.7%). For every categorical task, GT-to-synthetic agreement (Cohen's κ) exceeded the inter-reader agreement (Fleiss' κ) already present among radiologists reading the real images.

Data table
Impact on clinical managementShare of assessments
No change35.6%
Minor change, no impact on management34.4%
Major change, affects management or diagnosis30.0%

What predicts a change in clinical management?

Proportional-odds ordinal logistic regression. Odds ratios above 1 indicate a higher likelihood of altered patient management; below 1 is protective. Log scale.

OR = 1 Perceived synthetic appearance 3.60 Diagnostic complexity 2.14 Image quality 0.86  n.s. Reader work experience 0.54 0.5 1 2 4 Odds ratio (log scale) Increases impact Protective Not significant

Image quality was not independently predictive once perceived appearance and diagnostic complexity were accounted for. A 95% confidence interval is drawn only for reader work experience, the one predictor for which the manuscript reports it (0.38–0.77).

Data table
PredictorOdds ratio95% CIp
Perceived synthetic appearance3.60n/a< 0.001
Diagnostic complexity2.14n/a< 0.001
Image quality0.86n/a0.51
Reader work experience0.540.38–0.77< 0.001

How do uncertainty maps shift reader confidence?

Share of evaluations in which the per-pixel uncertainty map raised or lowered diagnostic confidence, by case complexity. The balance in each row is "no change".

0 All cases 17% 32% Low complexity 11% 39% High complexity 41% 14% Confidence decreased Confidence increased

The map was informative in 49% of evaluations overall, peaking at 55% on the most complex cases and 67% on non-mass enhancement lesions. Its direction inverts with complexity (Kendall's τ = 0.218, p = 0.003). Downstream, this behaved as a safety mechanism: when it lowered confidence readers requested true contrast in 88% of instances; when it raised confidence, 82% proceeded without real contrast. Overall, readers would avoid or defer physical contrast injection in 64% of cases.

Caveat. Inter-reader agreement on the map's effect was low (Fleiss' κ = 0.02–0.10), and readers flagged cases where the map showed high confidence despite a clinically inaccurate synthetic image. It is a useful adjunct in aggregate, not a fail-safe.

Data table
Case complexityConfidence increasedConfidence decreasedInformative rate
All cases32%17%49%
Low complexity39%11%50%
High complexity14%41%55%
Reader study figure contrasting a low-complexity case and a high-complexity case, each showing pre-contrast baseline, predicted synthetic time series, pixel-wise uncertainty maps, ground truth series, and kinetic enhancement plots.
Clinical evaluation across varying case complexities. A representative low-complexity case (top) and high-complexity case (bottom). Panels show the pre-contrast baseline, predicted synthetic time series, pixel-wise uncertainty maps (σ), and ground truth series. Kinetic plots compare predicted mean tumor intensity against ground truth over time, with the consensus characterization and diagnostic impact summarized per case.

Data Availability

BibTeX

@misc{joshi2026densetemporalcontrastsynthesis,
      title={Dense Temporal Contrast Synthesis via Conditioned Latent Transport},
      author={Smriti Joshi and Apostolia Tsirikoglou and Daniel M. Lang and Richard Osuala and Noah MÑrquez Varaa and Alejandro Guzman and Grzegorz Skorupko and Sebastian Ibarra Arregui and Lidia Garrucho and Akane Ohashi and Dimitra Ntoula and Eugen Divjak and Oğuz Lafcı and Jan C. Peeken and Julia A. Schnabel and Fredrik Strand and Oliver Diaz and Karim Lekadir},
      year={2026},
      eprint={2607.29394},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.29394},
}