Ultrasound Image Reconstruction by Self-Supervised Deep Neural Network A Study on Coherent Compounding Strategy

Sparse regularization (SR) techniques solve the inverse problem of beamforming and can reconstruct high-quality ultrasound image from single plane wave (PW) transmission, but with long reconstruction time (~1-5 mins). We previously proposed a self-supervised deep neural network (DNN) beamformer to accelerate the SR-based reconstruction of single PW to ~10 ms (DNN mapping time > 9 ms). However, when coherent compounding is used, the reconstruction time will increase by a factor equal to the number of PWs used. In this study, a new coherent compounding strategy is proposed to reduce the number of DNN mappings in the case of coherent compounding. Using simulation and experimental data, the proposed method is demonstrated to achieve comparable image quality and significantly reduced reconstruction time (from ~750 ms to ~33 ms) with 75 PWs, compared with our previous method.