Dense Temporal Contrast Synthesis via Conditioned Latent Transport

TL;DR - A conditioned latent transport model synthesizes virtual contrast-enhanced breast MRI in a single forward pass, potentially reducing or eliminating gadolinium contrast agents. It matters because it targets a real clinical bottleneck with external-cohort validation and a radiologist reader study.

• Anchors the latent trajectory to pre-contrast anatomy with continuous time conditioning, enabling patient-specific enhancement at any acquisition time without slow iterative sampling.
• Outperforms baselines and prior SOTA on spatial, perceptual, temporal, and distributional metrics; robust to scanner noise and differing acquisition protocols on an independent external cohort.
• Downstream tumor segmentation improved 22.4% relative Dice (0.60 vs 0.49 pre-contrast, p < 0.01) with >39% lower boundary error.
• Reader study with four breast radiologists over 40 cases: 70% of synthesized sequences supported the same management decisions as real DCE-MRI.