Machine learning-facilitated real-time acoustic trapping in time-varying multi-medium environments toward magnetic resonance imaging-guided microbubble manipulation

Magnetic resonance imaging-guided acoustic trapping is expected to manipulate drug carriers (e.g., microbubbles) within the body, potentially improving carrier concentration at tumor sites and thereby enhancing targeted therapy outcomes. However, accurate trap generation remains challenging due to complex wave propagation through multiple tissue materials. Moreover, respiration-induced tissue motion imposes stringent requirements on computational efficiency for rapid phase updates. Here we propose a machine learning-based model and a closed-loop control scheme to modulate phase patterns rapidly. The model delivers precise time-of-flight prediction (mean err. ≤ 0.24 μs) within 26 ms for 196 transducer elements. In proof-of-concept experiments, computer vision feedback permits fast (about 15 frames per second) position adjustment of a trapped polystyrene ball (Ø2.7 mm). This control scheme helps lessen the ball’s spatial drift induced by time-varying multi-medium environments. These experiments on robotic manipulation support our model’s potential for future magnetic resonance imaging-guided targeted therapy. Mengjie Wu and colleagues proposed a machine learning-based model and a closed-loop control scheme to modulate phase patterns of transducers rapidly. This technique enabled the generation of acoustic traps in 3D time-varying environments