Quantitative Viscoelastic Response (QVisR) Imaging

Quantitative Viscoelastic Response (QVisR) is a machine learning method for quantifying viscoelastic properties of tissue from on-axis acoustic radiation force (ARF) displacements. QVisR models utilize the same displacement profiles measured in Viscoelastic Response (VisR) as input to multi-layer perceptron models. Datasets are generated by finite element method (FEM) and ultrasound simulations using LS-DYNA and Field II.

QVisR models can be applied to acquisition VisR data via transfer learning using data acquired from calibrated elasticity ultrasound phantoms (tissue mimicking devices with known elastic and acoustic properties). Calibration of the simulation data to match acquisition data is vital for applying QVisR to real-world data, particularly because materials of known viscoelasticity for transfer learning are rare.

Shown below are QVisR elastic modulus predictions for a spherical inclusion phantom. The true elasticity values are shown in the top row, with the simulation-trained QVisR predictions in the middle row. We observe that strictly simulation-trained models can distinguish relatively soft and stiff materials, although with poor detection of the inclusion boundary. After transfer learning on an elasticity phantom, we see in the bottom row that QVisR is able to quantify elasticity in acquisition.

Selected Publication

J. B. Richardson, C. J. Moore, and C. M. Gallippi, “Quantitative Viscoelastic Response (QVisR): Direct Estimation of Viscoelasticity with Neural Networks,” IEEE Trans. Ultrason., Ferroelect., Freq. Contr., pp. 1–1, 2024, doi: 10.1109/TUFFC.2024.3404457.