Tissue Clutter Filtering Methods in Ultrasound Localization Microscopy Based on Complex-Valued Networks and Knowledge Distillation

Abstract

Ultrasound localization microscopy (ULM) is a blood flow imaging technique that utilizes micrometer-sized microbubbles (MBs) as contrast agents to achieve high-resolution microvessel reconstruction through precise localization and tracking of MBs. The accuracy of MB localization is critical for producing high-quality images, which makes tissue clutter filtering an essential step in ULM. Recent advances in deep learning have led to innovative methods for tissue clutter filtering, particularly those based on 3-D convolution, which effectively capture the spatiotemporal features of MBs. These methods significantly improve upon traditional approaches by addressing issues such as lengthy inference time and limited flexibility. However, many deep learning techniques primarily focus on B-mode images and demonstrate lower efficiency. To overcome these limitations, this study proposes knowledge distillation for tissue clutter filtering to enhance filtering efficiency while maintaining performance. This study first develops a lightweight 2-D complex-valued convolutional neural network (CCNN) (CL-UNet) as the teacher model, utilizing I/Q signal input. Subsequently, a 2-D real-valued convolutional neural network (CNN) (UNet-T) is developed as the student model, which uses envelope data as input. Feature-based knowledge distillation is applied to transfer knowledge from the teacher model to the student model (Guided UNet-T). All models are trained on simulated data and fine-tuned on in vivo data. The experimental results show that CL-UNet (I/Q, ours) demonstrates better filtering performance compared to the B-mode image-based approach on both simulated and in vivo data. Guided UNet-T outperforms both singular value decomposition (SVD) and random SVD (RSVD) in terms of both performance and speed, offering the best balance between filtering efficiency and effectiveness.