Ultrasonic beamforming algorithm based on deep learning and optimal model selection

This paper proposes a method for ultrasonic beamforming based on deep neural network (DNN) with model selection based on contrast-to-noise ratio to suppress the degradation sources of image quality. The contrast-to-noise ratio (CNR), one of the image quality evaluation indicators, is combined with the loss function of fidelity in the network model to form a new loss function: CNR-LOSS, which is expected to improve the correlation between loss function and ultrasonic image quality. The training data comes from the ultrasonic simulation signals of cysts and point targets, and the input of DNN is the channel signal and its corresponding wavelet coefficients. DNN divides the parallel input into two types: on-axis signal and off-axis signal, and expects to retain only on-axis signal and clear off-axis scattering. In addition, the performance of DNN beamformer with and without CNR-LOSS is compared, and the effect of loss functions with different CNR weights on image quality is analyzed. Compared with the DNN beamformer without CNR-LOSS, DNN beamformer with CNR-LOSS and appropriate CNR weights achieves higher image quality in the experiment.