Impact of acoustic properties and material heterogeneity on the acoustic radiation force impulse anisotropy imaging

Recently, we developed a regression neural network (RNN) to quantify shear modulus ratios (SMR) of anisotropic tissues from ARFI-induced peak displacement (PD) ratios using finite element simulations that contain a single inclusion and no ultrasonic tracking. This study assesses the impact of acoustic properties and material heterogeneity on the RNN. The network was trained with PD ratios from incompressible transversely isotropic elastic materials (SMR = 0.9–4.9; focal depths = 20, 22, 24, 26 mm; F-numbers = 1.25, 1.5, 2; c = 1540 m/s; α = 0.5 dB/cm/MHz), and tested on homogeneous (SMR = 2.25, 3.25, 4.25; depth = 22 mm; F-number = 1.5; c = 1500 and 1540 m/s; α = 0.5 & 1 dB/cm/MHz) and heterogeneous materials (SMR, inclusion = 5, 2, 1; background = 1, 1, 5), simulated using LSDYNA3D and Field II with an L7-4 transducer. Mean absolute percent errors (MAPEs) were 11%, 13%, and 8% for SMR = 2.25, 3.25, and 4.25 (axial range 18–32 mm, focal depth 22 mm), rising to average MAPEs of 18% and 19% for changes in sound speed (1540–1500 m/s) and attenuation (0.5– to 1 dB/cm/MHz). The RNN identified all inclusions with MAPEs of 6%, 5%, and 10% despite being trained on homogeneous material. Overall, the RNN robustly predicts SMR without being confounded by acoustic property mismatch or material heterogeneity.