In ultrasound scanning videos, noise and sub-stationary artifacts can severely impair the detection of the intended targets. This paper advances an unsupervised learning framework for noise suppression in ultrasound videos based on eigen filter concept. Specifically, the study focuses on the removal of artifacts in Synthetic Aperture Focusing Technique (SAFT) beamformed videos created by a moving transducer array in a scanning mode. The specific application is in-motion imaging of internal flaws in railroad tracks using a transducer array mounted in a Rolling Search Unit (RSU). By employing Singular Value Decomposition (SVD) analysis, an ultrasound video is decomposed into principal components through their time-space coherence/correlation. Artifacts can be eliminated by projecting out the associated principal components from the video (i.e. eigen filtering). A novel recursive algorithm with signal rectification is proposed to allow SVD filters to better capture the subtle movement of artifacts in ultrasound tomography. As opposed to the prior knowledge of a stopband in conventional SVD filters, the recursive SVD filter only removes the primary eigen mode until convergence. The proposed approach is validated through experiments using an RSU hosting a 25-element linear transducer array in defected rail sections in the laboratory and in the field. The results demonstrate a remarkable filtering ability, where artifacts (e.g. reflections at the wheel-rail interface, reverberations within the wheel, reflections from the bottom flange of the rail head) are successfully eliminated and only the internal flaws are imaged with high accuracy. This approach is applicable to any ultrasound imaging test where “loud” artifacts are present in addition to the intended targets.
Huang et al. (2025) studied this question.