Simulation study demonstrates real-time acoustic distortion correction using deep learning across skull targets, indicating potential for adaptive transcranial ultrasound therapy.
Key Points
To develop a deep learning surrogate model that predicts transducer phase and amplitude corrections directly from skull geometry in real time, bypassing computationally demanding time-reversal simulations.
Trained a dual-head deep learning architecture (amplitude regression and circular soft-labeled phase classification) using time-reversal simulations of a 1,024-element phased array transducer and CT-derived skull geometry.
Evaluated model inference, acoustic focus reconstruction, and computational speedup on two unseen test skulls across three target brain regions.
The model achieved a circular mean absolute error of 0.28–0.42 rad in phase prediction and a relative energy error of 1.18–4.35% in amplitude estimation.
Acoustic focus reconstruction achieved a peak location error of 0.61–1.46 mm, mean surface distance of 0.50–1.84 mm, and peak pressure recovery of 33–141%.
The approach yielded a 4,069× speedup in pure inference and a 48× wall-clock acceleration compared to conventional time-reversal simulations.