In synthetic aperture (SA) ultrasound imaging, either monostatic or multistatic approaches can be employed. In both cases, in transmission, a single element of the transducer array is used at each time. In reception, the same element is used for the monostatic approach, while the whole array is used for the multistatic one. Thus, the monostatic approach could be implemented using a simpler single-channel architecture, however at the expense of image quality, while the multistatic one provides a high quality image but requires a more complex N-channel system. In this work, we show that a deep neural network can be trained to reconstruct images with a high contrast, as in the multistatic SA case (considering a 128-element array), but starting from the pre-beamforming signals acquired through the monostatic SA approach. We implemented a U-net and trained it using 27200 simulated signal-sets and the corresponding target images generated with Field II, considering numerical phantoms with random elliptical targets. The deep neural network (DNN) output image quality was evaluated in terms of contrast on a test set made of 500 simulated images, and on experimental scans of a commercial phantom and of the carotid artery. The results show that, after training over 39 epochs, the DNN is able to provide images with a good quality starting from the radiofrequency signals obtained with a simple monostatic SA approach, potentially requiring a single-channel only.
Improving the Quality of Monostatic Synthetic-Aperture Ultrasound Imaging Through Deep-Learning-Based Beamforming
Spairani, E;Matrone, G
2022-01-01
Abstract
In synthetic aperture (SA) ultrasound imaging, either monostatic or multistatic approaches can be employed. In both cases, in transmission, a single element of the transducer array is used at each time. In reception, the same element is used for the monostatic approach, while the whole array is used for the multistatic one. Thus, the monostatic approach could be implemented using a simpler single-channel architecture, however at the expense of image quality, while the multistatic one provides a high quality image but requires a more complex N-channel system. In this work, we show that a deep neural network can be trained to reconstruct images with a high contrast, as in the multistatic SA case (considering a 128-element array), but starting from the pre-beamforming signals acquired through the monostatic SA approach. We implemented a U-net and trained it using 27200 simulated signal-sets and the corresponding target images generated with Field II, considering numerical phantoms with random elliptical targets. The deep neural network (DNN) output image quality was evaluated in terms of contrast on a test set made of 500 simulated images, and on experimental scans of a commercial phantom and of the carotid artery. The results show that, after training over 39 epochs, the DNN is able to provide images with a good quality starting from the radiofrequency signals obtained with a simple monostatic SA approach, potentially requiring a single-channel only.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.