Robust Eigen-Filter Design for Ultrasound Flow Imaging Using a Multivariate Clustering

Abstract

Blood flow visualization is a challenging task in the presence of tissue motion. Unsuppressed tissue clutter produces flashing artefacts in ultrasound flow imaging which hampers blood flow detection by dominating part of the blood flow signal in certain challenging clinical imaging applications, ranging from cardiac imaging (maximal tissue vibrations) to microvascular flow imaging (very low blood flow speeds). Conventional clutter filtering techniques perform poorly since blood and tissue clutter echoes share similar spectral characteristics. Eigen-based filtering was recently introduced and has shown good clutter rejection performance; however, flow detection performance in eigen filtering suffers if tissue and flow signal subspaces overlap after eigen components are projected to a single signal feature space for clutter rank selection.

To address this issue, a novel multivariate clustering based singular value decomposition (SVD) filter design is developed. The proposed multivariate clustering based filter robustly detects and removes non-blood eigen components by leveraging on three key spatiotemporal statistics: singular value magnitude, spatial correlation and the mean Doppler frequency of singular vectors. A better clutter suppression framework is necessary for high-frame-rate (HFR) ultrasound imaging since it is more susceptible to tissue motion due to poorer spatial resolution (tissue clutter bleeds into flow pixels easily). Hence, to test the clutter rejection performance of the proposed filter, HFR plane wave data was acquired from an in vitro flow phantom testbed and in vivo from a subject’s common carotid artery and jugular vein region induced with extrinsic tissue motion (voluntary probe motion). The proposed method was able to adaptively detect and preserve blood eigen components and enabled fully automatic identification of eigen components corresponding to tissue clutter, blood and noise that removes dependency on the operator for optimal rank selection.

The flow detection efficacy of the proposed multivariate clustering based SVD filter was statistically evaluated and compared with current clutter rank estimation methods using the receiver operating characteristic (ROC) analysis. Results for both in vitro and in vivo experiments showed that the multivariate clustering based SVD filter yielded the highest area under the ROC curve at both peak systole (0.98 for in vitro; 0.95 for in vivo) and end diastole (0.96 for in vitro; 0.93 for in vivo) in comparison with other clutter rank estimation methods, signifying its improved flow detection capability. The impact of this work is on the automated as well as adaptive (in contrast to a fixed cut-off) selection of eigen components which can potentially allow to overcome the flow detection challenges associated with fast tissue motion in cardiovascular imaging and slow flow in microvascular imaging which is critical for cancer diagnoses.