Predicting transcranial ultrasound insertion loss using skull CT: A deep learning approach

Area of Science:

• Biomedical Engineering
• Neuroscience
• Medical Imaging

Background:

• Transcranial ultrasound (tUS) offers non-invasive brain modulation but faces challenges due to skull-induced ultrasound attenuation.
• Insertion loss (IL) quantifies this signal degradation, critical for effective tUS delivery.
• Current IL prediction methods are computationally intensive and sensitive to variations.

Purpose of the Study:

• To develop a rapid and accurate method for predicting IL based on skull structural features.
• To investigate the correlation between skull anatomy and ultrasound attenuation.
• To leverage deep learning for efficient IL prediction in tUS applications.

Main Methods:

• Collected IL data from 20 human skull specimens at 220 kHz, 650 kHz, and 1000 kHz.
• Developed a modified dual-path Inception-based neural network (mDPI-Net) utilizing skull CT scans.
• Compared mDPI-Net performance against homogeneous pseudo-spectral methods and inhomogeneous simulations.

Main Results:

• mDPI-Net significantly outperformed homogeneous methods in accuracy (Peak Pressure Error: 26.6% vs. 34.3%).
• mDPI-Net showed comparable accuracy to complex simulations (IL Deviation: 2.47 dB vs. 1.69 dB).
• Computational efficiency dramatically improved, reducing prediction time from 15 min/sample to 0.5 s/sample.

Conclusions:

• Skull CT scans contain inherent structural information crucial for IL prediction.
• Deep learning models like mDPI-Net offer a computationally efficient and accurate approach to IL prediction.
• This technique holds potential for real-time pre-operative IL assessment, enhancing precision in tUS therapies.