Image-to-Image Translation with Deep Neural Networks for the Enhancement of Monostatic Synthetic-Aperture Ultrasound Images

In synthetic aperture (SA) ultrasound imaging, whether monostatic or multistatic, all elements of the transducer array are used in turn one-by-one during the transmission phase. However, during reception, the monostatic technique activates only the same element used in transmission, whereas the multistatic one employs the entire transducer array. The monostatic approach thus represents a potential solution for the implementation of an ultra-portable device, given that it would require a single-channel system to work. However, a downside would be a significant loss in terms of image contrast and signal-to-noise ratio caused by single-element transmission/reception. In a previous investigation, we introduced a deep neural network (DNN) for reconstructing high-quality images from radio-frequency (RF) signals acquired using a monostatic SA approach. Differently, in the current study, we propose a similar DNN to perform contrast enhancement of the images obtained from the same RF signals but through delay-and-sum (DAS) beamforming, addressing an image-to-image translation task. The dataset used for training the network consisted of pairs of monostatic/multistatic SA images of elliptical targets simulated in Field II. We also evaluated the effects of training the DNN in an adversarial framework and we compared the results in terms of contrast on a test set comprising both simulated and experimental scans. The results show that the DNN effectively enhances the contrast between the targets and the surrounding background. Additionally, the adversarial training introduces an improvement in speckle pattern and fine-grained details reproduction within the output images.