Deep Learning Coronary Pressure Drop Estimation

The patent pre-computes FFR values using accurate 3D CFD models for a comprehensive set of synthetic coronary vessel geometries and stenosis configurations before clinical use. These pre-computed results are stored in a database that serves as training data for a machine learning model. During actual clinical application, the trained ML model instantly predicts FFR for patient-specific CTA images, eliminating the need for time-consuming 3D CFD computations while maintaining accuracy through the pre-established relationship between vessel geometry and FFR values.

The patent creates a comprehensive database by generating synthetic coronary vessel geometries that replicate the diversity and complexity of real patient anatomy through randomized parameters including vessel diameter, length, curvature, and stenosis characteristics. These synthetic copies are then processed through accurate 3D CFD models to establish ground-truth FFR values, which train a machine learning model capable of accurately predicting FFR for real clinical cases without requiring complex 3D computations.

The patent performs the computationally intensive and energy-consuming 3D CFD simulations in advance during an offline training phase using synthetic vessel data. This pre-computation establishes a robust training dataset that captures the complex relationship between vessel geometry and FFR values. Once trained, the machine learning model requires minimal computational energy during clinical deployment, as it only needs to process patient-specific CTA images through the already-established predictive framework, thereby achieving high reliability with low operational energy consumption.