Learning physical properties in complex visual scenes: An intelligent machine for perceiving blood flow dynamics from...

Area of Science:

• Medical imaging analysis
• Computational fluid dynamics
• Artificial intelligence in cardiology

Background:

• Estimating physical properties from visual data is challenging, especially in complex biological systems.
• Fractional flow reserve (FFR) is crucial for diagnosing myocardial ischemia but traditionally requires invasive pressure measurements.
• Current methods for inferring FFR from medical images are limited.

Purpose of the Study:

• To develop a deep neural network, TreeVes-Net, capable of estimating FFR directly from static coronary CT angiography images.
• To enable machines to perceive complex physiological parameters like FFR from visual input.
• To provide a non-invasive tool for assessing coronary artery stenosis severity.

Main Methods:

• A novel deep neural network, TreeVes-Net, was designed.
• The framework incorporates a coronary representation encoder to capture geometric information.
• A tree-structured recurrent neural network (RNN) was employed to model long-distance spatial dependencies in blood flow.

Main Results:

• TreeVes-Net achieved an area under the ROC curve (AUC) of 0.92 and 0.93 on synthetic and real patient datasets, respectively.
• The model demonstrated superior performance compared to seven other machine learning-based FFR computation methods.
• Experiments were validated on 13,000 synthetic and 180 real coronary trees.

Conclusions:

• TreeVes-Net effectively estimates FFR from coronary CT angiography images.
• The proposed deep learning framework offers a promising non-invasive approach for evaluating coronary artery disease.
• This technology has the potential to improve clinical decision-making in coronary intervention procedures.