FUS-Net: U-Net-Based FUS Interference Filtering.

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• High intensity focused ultrasound (FUS) applications in imaging and therapy generate artifacts that obscure data.
• These artifacts intensify with increasing ultrasound power, posing a significant challenge for diagnostic accuracy and treatment efficacy.

Purpose of the Study:

• To develop and evaluate FUS-net, a deep learning method for removing FUS-induced artifacts from radiofrequency (RF) data.
• To demonstrate FUS-net's superiority over conventional filtering methods and stacked autoencoders (SAEs) in artifact reduction and image quality enhancement.

Main Methods:

• A Convolutional Neural Network (CNN)-based U-net autoencoder, termed FUS-net, was trained end-to-end on clean and corrupted RF data.
• The network was implemented using Tensorflow 2.3, learning to map corrupted RF data to clean RF data by representing artifacts in latent space.

Main Results:

• FUS-net demonstrated a 15% performance improvement over SAEs on test datasets.
• Beamformed B-mode images from FUS-net processed RF data exhibited superior speckle quality and contrast-to-noise ratio (CNR) compared to notch-filtered and adaptive least means squares filtered data.
• FUS-net filtered images showed reduced errors and higher similarity to clean images across all pressure levels, and enabled displacement mapping when combined with speckle-tracking algorithms.

Conclusions:

• FUS-net effectively removes high-pressure FUS interference from RF data, outperforming current filtering techniques and SAEs.
• The proposed method holds potential applicability for all FUS-based imaging and therapeutic modalities, enhancing diagnostic and therapeutic outcomes.