Towards in-vivo ultrasound-histology: Plane-waves and generative adversarial networks for pixel-wise speed of sound reconstruction

Ultrasound imaging is a well-established modality, widely used for in vivo real time examination. Nevertheless, the ability of conventional ultrasound techniques is limited by the fact that different biological tissues are sometimes represented with the same image brightness, thus hindering visual - as well as automatic - identification. Especially valuable for tissue differentiation is the pressure wave velocity, which can be measured with ultrasound. Deep-learning-based methods carry a possibility to overcome such limitations and enable robust signal-based tissue identification. Such methods have been successfully applied to tackle various challenges of medical imaging research. The aim of the present work is to propose a Generative Adversarial Network (GAN) pipeline towards pixel-wise speed of sound (SoS) reconstruction from plane-wave ultrasound raw channel signals corresponding to three firing angles. The network is trained on a novel synthetic dataset focusing on complex geometry, generated with K-Wave. Results demonstrate a promising performance, with average (± STD) absolute SoS reconstruction errors of 38 ±54 m/s in real time at 114 fps. The proposed approach paves the way towards GAN-based ultrasound histology.