Clinicians often have to assess cardiac function from only a few images, but dense motion data would be far more informative. Reconstructing the full heart cycle from sparse observations is mathematically ill-posed, and while ECG signals provide physiological context, they are hard to align with images because of cross-modal differences and phase offsets. A new paper from Liu et al. introduces CardioFAD, a Focus–Align–Diffuse framework that tackles this by learning a small set of motion-salient keyframes.
CardioFAD's central idea is to treat the problem as finding 'anchor states' that are both sufficient to reconstruct the visual trajectory and reliable for aligning ECG to the image sequence. Two modules implement this: Focus selects keyframes that preserve trajectory-critical motion content, while Align promotes consistent ECG–visual correspondence at those anchors. A diffusion model then generates the keyframes first and interpolates the remaining frames in a second stage.
Experiments on cardiac MRI and echocardiography videos report superior visual quality, temporal coherence, and functional consistency compared with existing methods, particularly in the extreme one-shot regime. The work, accepted at NeurIPS 2026, suggests that combining physiological signals with keyframe-based diffusion could make sparse cardiac imaging more clinically useful.