Singular value decomposition (SVD) has been established as an effective spatiotemporal clutter filter in ultrafast flow imaging, enabling enhanced sensitivity to small vasculature visualization. However, its performance depends on the choice of singular value thresholds assuming clutter, flow, and noise occupy distinct and separable subspaces. Clinically, these signal components often intermingle throughout the entire subspace spanned by the singular vectors, particularly in spatially heterogeneous fields of view. This entanglement undermines the efficacy of pure threshold-based SVD filtering, which frequently results in excessive suppression of weak flow signals and insufficient rejection of clutter and noise. To address this challenge, we reformulate it as an unsupervised clustering problem in the subspace formed by the singular vectors of the Casorati matrix. Instead of filtering singular components based on adaptive cutoff thresholds, the proposed method exploits the full SVD subspace to identify physiologically meaningful but spatially incoherent flow structures through clustering. Validated on human placenta data, this method enables accurate reconstruction of both macrocirculation and microcirculation across extensive imaging depth while achieving robust suppression of clutter and motion artifacts. Moreover, clustering-derived masks offer interpretable visualization and facilitate quantitative assessment of complex hemodynamic patterns. Work supported in part by Carilion Clinic Research Acceleration Program.
Zhu et al. (Wed,) studied this question.