Quantitative Viscoelastic Response (QVisR): Direct Estimation of Viscoelasticity With Neural Networks

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Accurate estimation of tissue viscoelasticity is crucial for diagnosing various medical conditions.
• Current ultrasound elastography methods face challenges in precisely quantifying viscoelastic properties.

Purpose of the Study:

• To develop and validate a machine learning method for direct estimation of viscoelastic moduli using ultrasound.
• To assess the accuracy of quantitative viscoelastic response (QVisR) in simulated and real-world scenarios.

Main Methods:

• Utilized a neural network trained on displacement time-series data from viscoelastic response (VisR) ultrasound excitations.
• VisR employs dual acoustic radiation force (ARF) pushes to measure tissue creep and relaxation.
• Input features included displacement profiles, push focal depth, and measurement axial depth.

Main Results:

• The machine learning model accurately mapped VisR displacements to elastic and viscous moduli.
• Quantitative VisR (QVisR) was validated in simulated materials.
• Domain adaptation methods improved phantom VisR displacement analysis.
• Successful in vivo estimates were obtained from clinical ultrasound data.

Conclusions:

• The proposed machine learning method offers a direct and accurate approach to estimate viscoelastic moduli from ultrasound data.
• QVisR, enhanced by machine learning and domain adaptation, shows promise for non-invasive tissue characterization.
• This technique has potential applications in clinical diagnostics and biomechanical research.