Ultrasound (US) is one of the most versatile medical imaging modalities. Reconstruction of high-quality images from a limited number of radio-frequency (RF) measurements is highly desired in portable, three-dimensional, and ultrafast ultrasound imaging systems. However, in many real-world imaging situations, access to RF data is limited and acquisition of paired images is infeasible. Inspired by the recent theory of unsupervised learning, the applicability of optimal transport driven CycleGAN (OT-CycleGAN) is investigated for the US artifact removal problems without  atched reference data. Various US artifact removal problems are then addressed using the two types of OT-CycleGAN. Experimental results for various unsupervised US artifact removal tasks confirmed that the proposed unsupervised learning method delivers results comparable to supervised learning in many practical applications.