EchoFlow: A Foundation Model for Cardiac Ultrasound Image and Video Generation

Preprint: https://arxiv.org/abs/2503.22357

Dataset Generation Pipeline

This demo showcases EchoFlow's ability to generate synthetic echocardiogram images and videos while preserving patient privacy. The pipeline consists of four main steps:

1. Latent Image Generation: Draw a mask to indicate the region where the Left Ventricle should appear. Select the desired cardiac view, and click "Generate Latent Image". This outputs a latent image, which can be decoded into a pixel space image by clicking "Decode to Pixel Space".
2. Privacy Filter: When clicking "Run Privacy Check", the generated image will be checked against a database of all training anatomies to ensure it is sufficiently different from real patient data.
3. Latent Video Generation: If the privacy check passes, the latent image can be animated into a video with the desired Ejection Fraction.
4. Video Decoding: The video can be decoded back to pixel space by clicking "Decode Video".

⚙️ Parameters

• Sampling Steps: Higher values produce better quality but take longer
• Ejection Fraction: Controls the strength of heart contraction in the animation
• CFG Scale: Controls how closely the animation follows the specified conditions