A new AI system can map how blood flows through a blocked heart artery using nothing but two ordinary X-ray images.
Researchers built a deep learning pipeline that takes dual-view coronary angiography - the standard X-ray procedure cardiologists already use - and reconstructs a 3D model of the artery, then predicts blood velocity and pressure at every point along it. An attention-enhanced CNN builds the geometry, while a second network layers in physics equations for fluid flow so the predictions stay physically consistent rather than just statistically plausible. Tested on 32 patients across four different flow conditions, the system's pressure-drop estimates were off by an average of just 2.02%, and it matched hospital-measured fractional flow reserve readings in 30 of 32 cases, for 93.8% diagnostic accuracy. The full process, from raw angiogram to complete hemodynamic map, runs in under 20 minutes per patient.
Today's standard test, fractional flow reserve, gives doctors one number for how much a blockage restricts flow but nothing about the 3D picture behind it. This approach promises that fuller picture without the cost of the wire-based FFR measurement itself, and without the hours of computing time traditional fluid-dynamics simulations require.
It's still a 32-patient research paper, not a cleared clinical tool, so the real test is whether these numbers hold up outside a controlled study - and whether hospitals want another model helping decide who needs a stent.