Deep Convolutional Neural Networks for Displacement Estimation in ARFI Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound elasticity imaging relies on precise micron-level displacement estimation from raw ultrasound data.
• Current methods for displacement estimation can be computationally intensive and may lack accuracy for complex soft tissue structures.

Purpose of the Study:

• To implement and evaluate a fully convolutional neural network (CNN) for accurate ultrasound displacement estimation.
• To develop a novel method for generating synthetic ultrasound training data for CNNs.
• To assess the performance of the CNN against conventional algorithms using simulated, phantom, and in vivo data.

Main Methods:

• A fully convolutional neural network (CNN) was developed for displacement estimation.
• A novel training dataset was generated using simulated 3-D displacement volumes with ellipsoids and Field II ultrasound simulation.
• The CNN was validated on simulated, experimental ARFI phantom, and human in vivo prostate ARFI datasets.

Main Results:

• The CNN achieved comparable root-mean-square error to Loupas's algorithm on simulated data ([Formula: see text] vs. 0.73 [Formula: see text]).
• In phantom studies, the CNN demonstrated similar contrast-to-noise ratio (CNR) for stiff inclusions compared to Loupas's algorithm (2.27 vs. 2.21).
• The trained CNN successfully visualized prostate cancer and anatomy in in vivo data, with computation times comparable to conventional methods.

Conclusions:

• Deep neural network-based displacement estimation is feasible for ultrasound elasticity imaging.
• The proposed CNN approach offers comparable accuracy and speed to existing time-delay estimation methods.
• The novel synthetic data generation method facilitates robust training for deep learning models in ultrasound imaging.