Deep Neural Network for Navigation of a Single-Element Transducer During Transcranial Focused Ultrasound Therapy: Proof of Concept

Abstract

Transcranial focused ultrasound (tFUS) has gained attention in the field of brain stimulation owing to its non-invasive neurotherapeutic potentials. However, complex interactions between acoustic waves and the cranium may introduce misalignment of the acoustic focus from a geometric target location, thus, necessitate on-site feedback of real-time navigational information of the transducer (spatial coordinates and angular orientation) for the operators to accurately place the acoustic focus to the desired brain area. In this study, we propose a deep-learning-based network model that can provide spatial navigational information of a single-element FUS transducer with respect to the targeted brain region. The training dataset were acquired through forward simulations that reflect the different tFUS transmissions for each skull structure using cranial computed tomography (CT) image data. The performance of the network was evaluated through three ex vivo calvaria. As a result show that the presented neural network-based method can an accurately navigate the FUS transducer with the conformity of ∼99.59% in placement of the transducer and ∼74.49% in the focal volume and an average difference of ∼0.96 mm in the focal point, all capable of real-time operation (∼10 ms).

tFUSFormer: Physics-Guided Super-Resolution Transformer for Simulation of Transcranial Focused Ultrasound Propagation in Brain Stimulation

“…In particular, voxels with ϕ ijk ≤ 0 were assigned the classification of water (including brain region and coupling media), while voxels with 0 < ϕ ijk < 1000 were categorized as trabecular bone, and voxels with 1000 ≤ ϕ ijk were designated as cortical bone, as indicated in Table I. The ultrasound velocity, density, and attenuation coefficient of each type of media were modeled using the parameter defined from previous studies [9], [14], [48], [61], [62]. In [63], sound wave propagation was simulated with greater realism by applying varying attenuation coefficients across different regions.…”

Multivariable-Incorporating Super-Resolution Convolutional Neural Network for Transcranial Focused Ultrasound Simulation

“…To evaluate the performance of the presented SRCNN models, IoU [54] was used as follows: where |R| denotes the number of the coordinate pairs (x i , y j , z k ) ∈ R, R SR and R HR are the set of coordinates of full-width at half-maximum (FWHM) regions of SR and HR simulation data, respectively. In general, the region of the FUS focus is defined by the FWHM [11].…”

“…The Westervelt-Lighthill equation in Eq. ( 2) was discretized by the staggered finite-difference time-domain (FDTD) method as follows [27], [54]:…”

“…The voxels were segmented based on the Hounsfield units (ϕ ijk ) as water ϕ ijk ≤ 0, trabecular bone for 0 < ϕ ijk < 1000, and cortical bone for 1000 ≤ ϕ ijk . The ultrasound velocity, density, and attenuation coefficient of the corresponding media were modeled as [7], [22], [54], [56]:…”

“…A single-element transducer with an operating frequency of 250kHz, a diameter of 96mm, a radius-of-curvature of 52mm, and a focal length of 83mm was used based on our previous work [54]. The ultrasound frequency range is applicable to human tFUS study [19], [20].…”