Non-invasive coronary fractional flow reserve prediction using a neural network with hemodynamic and geometric...

Invasive fractional flow reserve (FFR) is the clinical gold standard for assessing coronary artery stenosis, but its application is limited by its invasive nature. While CFD-based FFR_CT offers a non-invasive alternative, its high computational cost restricts real-time use. Deep learning has emerged as a promising solution. However, many existing methods rely primarily on geometric data and neglect personalized physiological boundary conditions. These limitations hinder the efficient characterization of complex coronary hemodynamics. To address these challenges, this study proposes a method for the rapid, patient-specific prediction of coronary hemodynamics by integrating personalized boundary conditions with geometric information.

In this study, we constructed a patient-specific steady-state coronary flow field dataset (288 patients). A dual-embedding neural network was proposed, which integrates personalized hemodynamic constraints with coronary geometry to enhance predictive accuracy. The network employs dual encoders to separately extract features from personalized hemodynamic boundary conditions and reduced-dimensional coronary geometry. By integrating these encoded representations into a Bi-LSTM architecture, the model learns the relationship between vascular topology and hemodynamic parameters. The predicted FFR_BGE values were validated against both CFD-based FFR_CT and invasive FFR to assess numerical agreement and diagnostic performance.

On the independent test set, the model achieved an RMSE of 7.42% compared to CFD-based FFR_CT. When validated against invasive FFR, the FFR_BGE showed a correlation of 0.87 and an AUC of 0.901. The inference time per case was significantly reduced compared to traditional CFD.

The dual-embedding neural network achieves high consistency with both CFD-based FFR_CT and invasive FFR. By encoding patient-specific coronary structures and boundary conditions, this approach provides a proof of concept that FFR distributions can be computed in real time, even with a relatively small CFD training dataset. These findings demonstrate a robust, high-efficiency solution with significant potential for broader application in diverse cardiovascular scenarios.