Echocardiogram Synthesis With Myocardium Motion Modeling

The patent combines a diffusion model (for image synthesis) with a neural ODE model (for motion estimation) into a unified framework. The diffusion model generates echocardiogram images while the neural ODE component estimates myocardium motion, and these two components are integrated to produce temporally coherent video sequences with accurate motion modeling, resolving the contradiction between motion accuracy and model complexity.

The neural ODE model acts as an intermediary between the diffusion model and the final video output. It estimates the motion field that transforms the first frame to subsequent frames, serving as a mediator that enables temporal coherence without requiring complex physics-based simulations throughout the entire generation process.

The patent uses the diffusion model to generate synthetic copies of echocardiogram video sequences with realistic myocardium motion. These synthesized datasets serve as artificial copies that augment the limited real annotated data, enabling deep learning models to be trained on larger datasets without requiring additional manual annotations, thus maintaining analysis speed while increasing data availability.

The system performs preliminary synthesis of annotated echocardiogram datasets before actual model training. By pre-generating synthetic training data with accurate motion modeling, the approach prepares sufficient training material in advance, eliminating the bottleneck of data availability that would otherwise slow down the development and deployment of deep learning models.

The diffusion model with neural ODE enables self-service annotation by automatically generating synthetic echocardiogram videos with realistic motion patterns. The system serves itself by producing annotated training data without requiring external expert cardiologists, thereby maintaining data quality while eliminating the time and cost associated with manual annotation.

Instead of manually annotating each real echocardiogram, the system creates synthetic copies with automatic annotations embedded during generation. These copied datasets provide the necessary training labels without the time-consuming manual process, preserving annotation accuracy through the physics-informed motion modeling while dramatically reducing annotation time.