Predicción de la pérdida de inserción por ultrasonido transcraneal utilizando TC de cráneo: un enfoque de aprendizaje profundo

Transcranial ultrasound (tUS) is a non-invasive neuromodulation technique with applications in brain disorders. However, ultrasound attenuation induced by the skull significantly affects focal energy transmission. The loss of ultrasound intensity can be quantified by insertion loss (IL). Accurate IL prediction is crucial for optimizing ultrasound delivery. Conventional grid-based numerical methods for IL prediction are computationally expensive and highly sensitive to parameter variations. To address these challenges, we hypothesized that skull structural features are inherently correlated with IL and can be effectively captured through deep learning method. In this study, we conducted transmission experiments on 20 human skull specimens to measure IL at three frequencies of 220 kHz, 650 kHz, and 1000 kHz. We proposed a modified dual-path Inception-based neural network (mDPI-Net) for IL prediction based on skull computed tomography (CT) scan. Comparison results showed that mDPI-Net outperformed homogeneous pseudo-spectral methods (Peak Pressure Error: 26.6% vs. 34.3%, IL Deviation: 2.47 dB vs. 4.64 dB), and is comparable to the inhomogeneous simulations (Peak Pressure Error: 26.6% vs. 21.0%, IL Deviation: 2.47 dB vs. 1.69 dB), while achieving higher computational efficiency, increasing from 15 min/sample to 0.5 s/sample. The proposed approach demonstrated that the skull CT scan could inherently encode structural information relevant to IL. Under a well-fixed experimental setup, deep learning has the potential to enable real-time or rapid pre-operative IL predictions, and achieve more precise dose control in tUS applications such as neuromodulation, transcranial drug delivery, and non-invasive brain stimulation.