Unsupervised Learning for Acoustic Shadowing Artifact Removal in Ultrasound Imaging

Abstract

One of the most prominent artifacts in the ultrasound (US) imaging is the acoustic shadowing artifact. Acoustic shadowing artifacts appear when the transmitted signal encounters with the structure of high attenuation property results in a drop of transmitted energy for deep region and produce dark area in the US image. To overcome this problem, a time gain compensation (TGC) tool has been used, but it cannot automatically produce an improved result. That is to say, the human intervention is necessary. To alleviate this problem, herein a deep learning based algorithm is proposed for the removal of shadowing artifacts. Since, obtaining a paired dataset for this task is difficult, therefore cycle consistency-based generative adversarial network (CycleGAN) is utilized to train the proposed deep learning model in an unsupervised fashion. The model is trained to process delay-and-sum (DAS) beamformer images as the input to generate shadowing-free patches as the target. The method is evaluated on real in vivo samples and shown to produce noticeably improved quality images.