Improvement in phase velocity reconstruction using attention based deep denoising of shear wave velocity fields

Noise in ultrasound shear wave elastography (SWE) signals often hinders accurate estimation of shear wave phase velocity, particularly in deep or fatty tissues where conventional filtering techniques fail. To address this, we propose a deep learning based denoising method that transforms particle velocity data into a time-frequency representation and applies an attention-enhanced convolutional encoder-decoder network with residual connections. This architecture effectively separates noise from the desired signal, substantially improving the signal-to-noise ratio (SNR) under high-noise conditions and enabling more reliable reconstruction of shear wave phase velocity maps. The model was trained on simulated phantoms (185,570 samples, 80-20% train-validation split) and tested on experimental phantoms and ex vivo goat liver. Results show SNR improvements from 1.47 dB to 19 dB in CIRS phantoms and from 5 dB to 9.4 dB in ex vivo goat liver tissue, with reduced variation in reconstructed phase velocity maps. These findings demonstrate that deep denoising can enhance phase velocity estimation in SWE and potentially improve tissue characterization in challenging imaging conditions.