Viscoelastic parameter estimation using simulated shear wave motion and convolutional neural networks.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Ultrasound shear wave elastography (SWE) is valuable for assessing tissue rheology.
• Robust evaluation of viscoelastic properties using SWE faces challenges.
• Current methods often involve complex mathematical analyses.

Purpose of the Study:

• To investigate the capability of convolutional neural networks (CNNs) for retrieving viscoelastic parameters from simulated SWE data.
• To assess CNN performance against traditional Fourier-based analysis for elasticity and viscosity estimation.
• To evaluate the robustness of CNN models to noise and varying imaging parameters.

Main Methods:

• Generated simulated shear wave motion data using staggered-grid finite difference simulations based on a Kelvin-Voigt model.
• Varied shear elasticity (1-25 kPa) and viscosity (0-10 Pa·s) with different push profiles (f-numbers 1 and 2).
• Trained CNN architectures using mean squared error loss to retrieve viscoelastic parameters and compared with 2D Fourier transform analysis.

Main Results:

• CNNs successfully retrieved both elasticity and viscosity with high accuracy (R² > 0.99).
• CNN models demonstrated robustness to noise and vertical position, and partial robustness to f-number.
• CNNs outperformed Fourier-based analysis in accuracy across the full viscoelastic parameter range.

Conclusions:

• CNNs show significant potential as an alternative to complex mathematical analyses for SWE viscoelastic parameter estimation.
• Trained CNN architectures can be robust to multiple push profiles.
• This AI-driven approach offers a promising avenue for more accurate and potentially simpler viscoelastic tissue property analysis.