Blood flow visualization using a ConvLSTM-based deep learning clutter filter

View at publisher Summarise with VIRAT
Authors: Hongpeng Wang, Jiayi Li, Masaaki Omura, Ryo Nagaoka, Shangce Gao, Hideyuki Hasegawa
Cited by: 0
References: 30

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Abstract This study investigates the feasibility of applying a convolutional long short-term memory (ConvLSTM) model to high-frame-rate ultrasound data for clutter suppression and blood-flow visualization. The proposed two-layer ConvLSTM model was trained using a Gaussian-masked mean squared error loss to emphasize vascular regions and suppress tissue clutter. The experiments showed that ConvLSTM effectively separated blood-flow and tissue components, achieving higher contrast than conventional singular value decomposition (SVD) filtering. Velocity estimation using block-matching indicated that ConvLSTM closely tracked the temporal velocity trends in the SVD results. A slight increase in the mean velocity during the systolic phase was attributed to delayed temporal accumulation caused by long-term memory. Designed to capture long-term temporal dependencies, the ConvLSTM model can extract hemodynamic features from sequential ultrasound data, enabling effective clutter suppression and stable visualization of blood flow in high-frame-rate imaging.

Fields of Study & Subjects

Ultrasound Imaging and Elastography Medical Image Segmentation Techniques Photoacoustic and Ultrasonic Imaging

Publication details

Journal / source
Japanese Journal of Applied Physics
Volume
65
Issue
7
Pages
07SP31–07SP31
Publication date
23 March 2026
Language
EN