Real-Time Left Ventricular Surrogate

A computational surrogate for rapid estimation and visualization of left ventricular hemodynamics.

Real-Time Left Ventricular Surrogate project preview

Project Overview

The Real-Time Left Ventricular Surrogate addresses the computational cost associated with repeatedly performing high-fidelity computational fluid dynamics simulations for cardiovascular analysis. Although CFD provides detailed information about blood-flow behaviour within the left ventricle, repeated simulations can be computationally demanding and unsuitable for applications requiring rapid prediction.

The project uses deep learning to construct a surrogate model that learns the relationship between simulation conditions and corresponding left ventricular hemodynamic outputs. CFD-generated data is used to train the model to approximate pressure and velocity fields, allowing new predictions to be generated substantially faster than performing a complete numerical simulation for each case.

By combining cardiovascular modelling, CFD-derived datasets, neural surrogate modelling, and scientific visualization, the system explores how data-driven methods can complement conventional numerical simulation and support faster analysis of left ventricular hemodynamics.

Technology Stack

  • Python
  • Deep Learning
  • Neural Networks
  • Computational Fluid Dynamics
  • Scientific Computing
  • Hemodynamic Modelling
  • Scientific Visualization

System Architecture

The system transforms CFD-derived cardiovascular simulation data into a learning pipeline that trains a neural surrogate to rapidly approximate left ventricular hemodynamic fields.

  1. CFD Simulation Data — provides high-fidelity left ventricular hemodynamic data used for surrogate training.
  2. Data Processing Layer — prepares physiological and simulation-derived parameters for model development.
  3. Surrogate Learning Model — learns the relationship between simulation inputs and corresponding hemodynamic outputs.
  4. Prediction Engine — generates rapid approximations of left ventricular pressure and velocity fields.
  5. Visualization Layer — presents predicted hemodynamic fields for analysis and comparison.

System Workflow

  1. Left ventricular CFD simulations are used to generate reference hemodynamic data.
  2. Simulation parameters and corresponding output fields are extracted and prepared for model training.
  3. The dataset is processed and supplied to the neural surrogate model.
  4. The surrogate learns mappings between simulation conditions and resulting hemodynamic fields.
  5. New physiological or simulation parameters are provided to the trained model.
  6. The model rapidly predicts corresponding pressure and velocity fields.
  7. Predicted outputs are visualized and compared with CFD-derived reference results.

Key Features

  • Real-time approximation of computationally expensive CFD outputs.
  • Prediction of left ventricular pressure and velocity fields.
  • Deep-learning-based surrogate modelling.
  • Integration of CFD-generated data with machine learning.
  • Rapid inference for repeated hemodynamic analysis.
  • Scientific visualization of predicted cardiovascular flow fields.

Results

  • Developed a neural surrogate for rapid approximation of left ventricular hemodynamic fields.
  • Reduced reliance on repeated computationally expensive CFD simulations during prediction.
  • Enabled rapid prediction of pressure and velocity distributions from simulation-derived inputs.
  • Integrated cardiovascular simulation data with deep-learning-based surrogate modelling.
  • Created a foundation for interactive and real-time exploration of left ventricular hemodynamics.

GitHub Repository

View source code on GitHub