Deep Learning-Based Clutter Suppression for Single-Shot Ultrasound Flow Imaging

Abstract

Clutter filtering is a crucial step in ultrasound flow imaging for eliminating low-frequency signals arising from stationary or slowly moving tissue. Traditional clutter suppression techniques such as high-pass temporal filtering and singular value decomposition (SVD) rely on long temporal ensembles, making them unsuitable for real-time or single-frame processing. In this work, we introduce a deep learning-based method that enables clutter suppression from a single ultrasound frame—no angular compounding or ensembles required. We design an Attention U-Net architecture that incorporates spatial attention mechanisms to focus on flow-related features while attenuating clutter. Our model demonstrates strong clutter suppression and high structural similarity with ground truth filtered outputs. This work opens the door for real-time, single-frame blood flow imaging using deep learning.