3D Real-Time Acoustic Simulation for Transcranial Focused Ultrasound Using Conditional Generative Adversarial Network

Abstract

Acoustic simulation is useful to plan the sonication path prior to transcranial application of focused ultrasound. There have been increasing demands for real-time acoustic simulation during the therapy since it is challenging to perfectly match the sonication path to the planned one. However, conventional numerical wave solvers are computationally expensive to implement online treatment planning. In this study, we developed a conditional generative adversarial network (cGAN) model which can predict acoustic pressure distribution in real-time. The input of the cGAN model was made as an image of a transducer and skull structure. For training data, we performed acoustic simulations at the transducer maneuvering space using k-Wave acoustic toolbox under the 200 kHz sonication condition. The cGAN model was trained using the pre-calculated simulation results during 500 epochs. To assess the accuracy, we compared the simulation results between numerical solver and cGAN model. The maneuvering space of the transducer was evenly discretized as 1200 points and used half for training and the other half for assessing. In this study, we considered that the axial axis of the transducer intersected with the target point. The average difference of peak intracranial acoustic pressure was 1.1%, and the distance between two acoustic focus was 3 mm. The cGAN model spends only 0.07s to estimate the pressure distribution while the numerical solver takes 30s. We developed a cGAN model to predict the acoustic pressure distribution in real-time. This makes possible the real-time prediction of the acoustic field passing through the skull during the tFUS treatment.