Deep learning-based real-time estimation of transcranial focused ultrasound acoustic field

Transcranial focused ultrasound (tFUS) techniques have garnered considerable attention as a novel noninvasive brain stimulation modality due to their high spatial specificity and depth penetration. However, estimating the intensity, location, and shape of the ultrasound focus is challenging due to wave distortion through the inhomogeneous skull. Because conventional imaging methods cannot capture low-intensity acoustic foci, numerical simulations are required to estimate intracranial pressure fields. However, such simulations are computationally intensive, limiting real-time use. In this study, we introduce a deep learning-based surrogate model to enable real-time estimation of the intracranial acoustic field distribution of tFUS. The proposed model effectively captures skull computed tomography (CT) features via a pre-trained deep neural network and includes two modules: one predicts acoustic field distributions, and the other estimates peak pressure values to enhance overall accuracy. The model was trained using data from 13 cranial CT scans and validated against direct field measurements from three ex vivo calvaria. The proposed model demonstrated high accuracy in focal point estimation, achieving a peak pressure ratio error of 3.94 % and a focal position error of 2.46 mm, indicating precise localization of the ultrasound focus. For focal volume prediction, the model exhibited a maximum boundary error of 5.90 mm while maintaining a focal volume conformity of 81 %. Notably, the inference time was 16 ms, which is significantly faster than conventional numerical simulations, ensuring feasibility for real-time applications. This method facilitates precise intracranial targeting, significantly enhancing clinical viability of tFUS for therapeutic applications, including functional neuromodulation.