Deep Learning Based Real-Time Phase-Amplitude Correction for Phased Array Transducers in Transcranial Focused Ultrasound

Abstract

Transcranial focused ultrasound (tFUS) is a non-invasive therapeutic approach for various brain disorders. However, acoustic distortions during ultrasound propagation through the skull make accurate targeting challenging. Time-reversal (TR) simulations can correct these distortions but are computationally expensive for real-time applications. The present study proposes a deep learning model that predicts the phase and amplitude of each phased-array transducer element in real time, enabling accurate patient-specific focusing at the target point. Using CT-derived skull geometry as input, the proposed model replaces computationally intensive TR simulations with direct prediction from skull geometry and spatial configuration. The model was trained using TR simulations of a 1,024-element phased array transducer system and evaluated on two unseen test skulls excluded from training. A dual-head architecture separately handles amplitude regression and phase classification, using circular soft labeling to capture phase periodicity. The proposed 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. The reconstructed acoustic focus achieved 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 % across three target regions. Compared to conventional TR simulations, the proposed method achieves a 4,069× speedup in pure inference and a 48× wall-clock acceleration, supporting real-time clinical deployment. These results demonstrate that skull-induced acoustic distortions can be efficiently approximated using physics-informed surrogate modeling, providing a scalable foundation for real-time patient-adaptive transcranial ultrasound systems.