A Machine Learning Framework to Overcome Displacement Underestimation in Acoustic Radiation Force Impulse (ARFI) Imaging

ARFI imaging-induced displacement estimates tissue stiffness qualitatively (i.e., target region stiffness compared to reference region). Generally, it has better resolution, CNR, and penetration depth than shear wave-based methods. ARFI methods require an accurate estimation of the tissue displacements. However, ARFI methods underestimate the true displacement due to the lateral and elevational shearing of the ultrasound beam’s point-spread function. The pulse repetition frequency (PRF) of the tracking pulse, depending on the imaging depth, contributes to the displacement underestimation. Underestimation can significantly impact image quality, leading to a lower contrast-to-noise ratio (CNR). The objective of this work is to develop a convolutional neural network (CNN)-based method to overcome the displacement underestimation in ARFI imaging. The model was trained on simulated 66,000 displacement profiles from ARFI imaging of homogeneous viscoelastic materials (shear modulus: 1 to 37 kPa, steps: 4 kPa, and viscosity: 0.1-4.1 Pa•s, steps: 0.5 Pa•s) using LS-DYNA FEM and Field II simulation package. FEM displacement was generated with a pulse repetition frequency of 15 kHz and then down-sampled at a PRF of 12 to 2 kHz: steps : 2 kHz to reflect the depth-dependent PRF. Additive white Gaussian noise was introduced to the beam-formed RF data to achieve signal-to-noise ratio (SNR) levels ranging from 0 to 35 dB in 5 dB increments. The model was tested on two heterogeneous materials with inclusion elasticity of 12 and 6 kPa embedded in 3 kPa background. The model improves the mean CNR of the 12 kPa and 6 kPa inclusion embedded in 3 kPa background by 36.7% and 9.96% across all PRFs, respectively. Future work will validate the model in experiments.