Physics-Guided Neural Networks for Intraventricular Vector Flow Mapping.

Area of Science:

• Cardiovascular Imaging
• Computational Fluid Dynamics
• Artificial Intelligence in Medicine

Background:

• Intraventricular vector flow mapping (iVFM) is crucial for quantifying cardiac blood flow using color Doppler.
• Traditional iVFM optimization methods have limitations in performance and generalizability.

Purpose of the Study:

• To develop and evaluate novel alternatives to traditional iVFM optimization schemes.
• To leverage physics-informed neural networks (PINNs) and a physics-guided nnU-Net approach for improved iVFM.

Main Methods:

• Implementation of PINNs with dual-stage optimization and pre-optimized weights.
• Development of a physics-guided nnU-Net supervised learning model.
• Evaluation on simulated Doppler images from patient-specific CFD models and in vivo Doppler acquisitions.

Main Results:

• Both PINNs and nnU-Net demonstrated comparable reconstruction performance to the original iVFM algorithm.
• nnU-Net exhibited superior generalizability, real-time capabilities, and robustness on sparse/truncated Doppler data.
• PINNs showed boosted efficiency through optimized training strategies.

Conclusions:

• PINNs and nnU-Net are effective alternatives for reconstructing intraventricular vector blood flow.
• nnU-Net offers advantages in robustness and real-time application for cardiac imaging.
• Potential for PINNs in ultrafast Doppler imaging and deriving cardiovascular disease biomarkers from blood flow data.