Physics-informed deep learning for blood flow quantification

This work is done by my PhD student Hang-Jung Ling and in collaboration with my colleague and friend Damien Garcia, researcher at the French institute INSERM and specialist in ultrasound imaging and fluid dynamics.

We developed a physics-informed deep learning method to estimate vector flow mapping from color Doppler imaging. Our solution is based on an nnU-Net architecture which outputs the radial and angular velocity map of the blood flow. The corresponding spatial derivatives are then automatically computed from an autograd layer to build a physical loss. This strategy generates vector flow mapping from color Doppler imaging, facilitating interpretation and analysis of blood flow.