VESTA: Machine Learning-Enabled Estimation of ViscoElastic Ratios from On-Axis Spatio-Temporal ARFI Features

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Tissue viscoelasticity is crucial for cancer diagnosis, as tumors alter elasticity and viscosity.
• Current acoustic radiation force (ARF) methods are computationally intensive and require bias correction.
• Existing techniques struggle to fully capture the diagnostic value of combined elasticity and viscosity changes.

Purpose of the Study:

• To develop a novel, data-driven pipeline named VESTA for estimating elasticity ratio (ER) and viscosity ratio (VR).
• To bypass computationally intensive model fitting and bias correction in ARF-based viscoelastic characterization.
• To enable direct prediction of ER and VR from ARFI displacement features for improved diagnostic efficiency.

Main Methods:

• A two-stage machine learning pipeline: Stage 1 uses an MLP classifier for boundary detection, and Stage 2 employs a dilated Conv1D regression model for ER and VR estimation.
• The pipeline utilizes seven normalized ARFI displacement features, eliminating the need for nonlinear model fitting.
• Trained on 500 simulated inclusion scenarios with diverse parameters and validated on phantoms and an in vivo murine breast cancer model.

Main Results:

• In silico validation showed mean predicted ER and VR within 12% of ground truth.
• The VESTA pipeline demonstrated plausible generalization to heterogeneous real tissue.
• In vivo studies successfully tracked treatment-related changes in mechanical contrast in tumors over 36 days.

Conclusions:

• VESTA provides a computationally efficient and accurate method for estimating tissue viscoelastic ratios.
• The pipeline shows promise for non-invasive tumor monitoring and treatment response assessment.
• This data-driven approach advances the application of ARF elastography in oncology.