Ultrasound Beamforming: Investigating Time Series with Sequence to Sequence Approach in Deep Learning

Deep learning takes the lead in developing new ultrasound beamforming techniques, the rapid development of new models in computer vision, image segmentation, and classification has turned ultrasound beamforming into an image processing problem. The CNN and GAN structures achieved noticeable performance compared with the classical methods regarding B-mode image formation time and accuracy; nonetheless, using the time series with a sequence-to-sequence approach still needs to be thoroughly investigated. The challenge of acquiring suitable datasets limits any significant advancement in the time series with the Seq2seq approach, in contrast to the image processing approach delivering good results with less training data, in the other side the successive evolution of Seq2seq models applied in (NLP) show some promising future of this approach. This paper outlines the ultrasound Radio-Frequency time series (USRFTS) and the various architectures that tackle this problem, from RNNs and their derivatives like LSTM and GRU to the attention mechanism, and mainly self-attention the mechanism behind the transformers architecture. We also highlighted some limitations in using each model, some of the common obstacles, and the thriving opportunities to use this method in ultrasound B-mode image beamforming.