Transcranial ultrasound viscoelasticity and fluidity imaging of brain tissue based on multi-scale spatiotemporal deep learning.

Brain tissue mechanics play a critical role in neurological disorders. However, noninvasive characterization is impeded by the tissue's biphasic composition and small shear modulus. We developed a transcranial ultrasound viscoelasticity and fluidity imaging method using multiscale spatiotemporal deep learning. This method integrates multiscale pyramidal convolution and hybrid losses to overcome the limitations of low-SNR signals and small shear displacements through dual-branch processing, capturing high-frequency details (300 Hz) and low-frequency patterns (100 Hz) simultaneously. The proposed method achieves transcranial ultrasound viscoelasticity and fluidity imaging in brain tissue, outperforming the pyramid, warping, and cost volume network (PWC-Net) with a 4% reduction in low-frequency reconstruction errors and enhanced high-frequency signal fidelity. Validation of the method was conducted on simulation data, phantom data, and ex vivo animal data, demonstrating dual-tumor imaging at a 5.5 mm radius with quantitatively superior metrics (SNR=17.43, CNR=5.64) compared to existing methods.