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

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Signal Processing

Background:

• Ultrasound localization microscopy (ULM) requires precise microbubble (MB) tracking for high-resolution microvessel imaging.
• Tissue clutter filtering is crucial for ULM accuracy, with deep learning methods showing promise but facing efficiency limitations.
• Existing deep learning approaches often use B-mode images, limiting their efficiency in ULM applications.

Purpose of the Study:

• To enhance tissue clutter filtering efficiency and performance in ultrasound localization microscopy (ULM) using knowledge distillation.
• To develop a faster and more effective filtering method for ULM by transferring knowledge from a complex-valued network to a real-valued network.
• To improve the accuracy of microvessel reconstruction in ULM through optimized clutter filtering.

Main Methods:

• A lightweight 2-D complex-valued convolutional neural network (CL-UNet) using I/Q signals was developed as a teacher model.
• A 2-D real-valued convolutional neural network (UNet-T) using envelope data was designed as a student model.
• Feature-based knowledge distillation was employed to transfer filtering knowledge from CL-UNet to UNet-T, creating Guided UNet-T.

Main Results:

• The CL-UNet model demonstrated superior filtering performance over B-mode image-based methods on both simulated and in vivo data.
• Guided UNet-T significantly outperformed traditional Singular Value Decomposition (SVD) and Random SVD (RSVD) methods.
• Guided UNet-T achieved the optimal balance between filtering speed and effectiveness, surpassing existing deep learning approaches in efficiency.

Conclusions:

• Knowledge distillation offers a viable strategy to enhance the efficiency of deep learning-based tissue clutter filtering in ULM.
• The proposed Guided UNet-T method provides a computationally efficient and highly effective solution for ULM imaging.
• This approach advances ULM by enabling faster, more accurate microvessel visualization through improved clutter filtering.