Self-supervised ultrasound B-mode strain elastography using SMURF.

Area of Science:

• Medical imaging
• Biomedical engineering
• Machine learning

Background:

• Ultrasound strain elastography (USE) is crucial for tissue characterization, but lateral displacement estimation remains challenging.
• Clinical systems often provide only B-mode images, limiting the application of existing USE methods.
• Unsupervised deep learning networks (DLN) require large datasets, and B-mode data can address this for clinical translation.

Purpose of the Study:

• To explore the use of Self-Teaching Multi-Frame Unsupervised Recurrent All-Pairs Field Transform (SMURF) for ultrasound strain elastography (USE).
• To estimate axial and lateral strain tensor components using unsupervised deep learning on B-mode image loops.
• To evaluate SMURF's performance against traditional methods for in vivo imaging.

Main Methods:

• Retrained the RAFT network with baseline supervision on simulated and experimental B-mode datasets.
• Applied SMURF unsupervised training on experimental B-mode and in vivo datasets.
• Utilized four-dimensional (4D) cost volumes for displacement and strain estimation.

Main Results:

• The fine-tuned RAFT model with unsupervised SMURF achieved comparable accuracy and precision in displacement and strain estimation to traditional methods, particularly in the lateral direction.
• SMURF demonstrated significantly faster processing times, with a 78% improvement over the GPU-based Lagrangian Carotid Strain Imaging (LCSI) method.
• The method showed potential for real-time Lagrangian USE in clinical studies.

Conclusions:

• Unsupervised deep learning techniques, like SMURF, are effective for USE on B-mode datasets.
• SMURF offers a promising solution for accurate and efficient lateral displacement and strain estimation in USE.
• This approach facilitates wider clinical application of USE, potentially enabling real-time imaging.