Performance of learned pseudo-CT in transcranial ultrasound simulations using fluid and solid skulls

Transcranial ultrasound (tUS) applications require accurate simulations to predict intracranial acoustic pressure. tUS simulations are usually performed neglecting shear wave propagation in the skull (fluid skull) due to its simplicity. Computed tomography (CT) head scans are the gold standard to extract geometrical and material properties needed in tUS simulations. To minimize ionizing-radiation in patients, pseudo-CT images obtained from magnetic resonance (MR) imaging by deep learning (DL) methods are an attractive alternative to CT. We built a U-net based neural network to map MR images to CT images and simulated the tUS field generated by a 0.5 MHz transducer focused on the cortex, propagating through a fluid- or solid skull. At normal incidence, the maximum error in the DL-simulated lies below 35% compared to the CT-simulation. However, at 40°of incidence the error in the predicted peak transcranial pressure increases up to 60% in DL-simulated solid skulls compared to CT-simulated solid skull. The smaller wavelength of shear waves is much more affected by the fine inner skull structure, which is missing in pseudo-CT images. Thus, our findings suggest that the DL-based pseudo-CT images are not suitable for predicting tUS fields in arbitrary conditions and should only be considered under strict normal incidence.