Cerebral artery stenosis: a neural network can localise it with 99.4 per cent accuracy based on just a few blood flow velocity measurements

Trained on haemodynamic simulations, a model identifies the narrowed artery and the severity of the stenosis by comparing the predicted and observed velocity curves in eight vessels.

A narrowing of a cerebral artery disrupts the flow of blood throughout the vascular network, making it difficult to trace its location and severity from a few scattered measurements. Researchers have designed a conditional neural operator that predicts brain hemodynamics for a candidate stenosis, then uses it backwards: we look for the stenosis whose predicted velocity curves best match those observed.

The model was trained on data from a coupled 0D-1D simulation of pressure, cross-section and velocity in vessels, under various anatomical and physiological conditions. The average relative validation errors are 0.14% for pressure, 0.59% for cross-section and 2.79% for velocity. With measurements on eight vessels, the affected artery is identified among 20 candidates in 99.4% of cases without noise, and again 96% with a noise of 30%.

Still far from the patient

The authors emphasize that the location of the measurement points weighs on the result, and that the communicating and distal arteries remain more ambiguous. Above all, the evaluation remained entirely simulated: generalization to new individuals and to real clinical measurements remains to be demonstrated.