Controllable Echocardiogram Video Generation Driven by Fine-Grained Cardiac Motion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Echocardiogram video synthesis aids AI model training and clinical education.
• Current methods use limited scalar metrics (e.g., LVEF), failing to capture complex cardiac dynamics.
• This limits the clinical applicability of AI models in echocardiography.

Purpose of the Study:

• To develop a novel image-to-video synthesis framework for echocardiograms.
• To generate videos guided by comprehensive left ventricular volume-time curves (VTCs) for improved cardiac function representation.
• To enable controllable and physiologically consistent echocardiogram video generation.

Main Methods:

• A VTC-conditioned diffusion model for controllable echocardiogram video generation.
• A two-stage architecture with a latent optical flow module and an image-to-flow sequence model.
• Adversarial loss for enhanced visual fidelity and a semi-supervised framework for VTC extraction from videos.

Main Results:

• The proposed method achieves state-of-the-art performance in echocardiogram video synthesis.
• Enables fine-grained control over cardiac dynamics.
• Generates temporally coherent and visually faithful echocardiographic sequences.

Conclusions:

• The VTC-guided synthesis framework offers a more comprehensive representation of cardiac function than scalar metrics.
• The method addresses data scarcity and improves the practical applicability of generative models in echocardiography.
• Advances the potential for AI in cardiac diagnostics and education.