Bi-Directional Semi-Supervised Training of Convolutional Neural Networks for Ultrasound Elastography Displacement...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Ultrasound

Background:

• Ultrasound elastography (USE) performance relies on accurate displacement estimation.
• Convolutional neural networks (CNNs) show promise but require adaptation for ultrasound data due to domain gaps.
• Existing methods struggle with unknown ground-truth displacements and domain shifts in simulated data.

Purpose of the Study:

• To develop an improved semisupervised deep learning method for displacement estimation in USE.
• To address the limitations of transfer learning and simulated data in adapting CNNs for ultrasound.
• To enhance the accuracy and robustness of displacement estimation in USE.

Main Methods:

• Employed a semisupervised learning approach using first- and second-order displacement derivatives for regularization.
• Modified network architecture to estimate both forward and backward displacements.
• Introduced consistency between forward and backward strains as an additional regularizer.

Main Results:

• Validated the method on experimental phantom and in vivo ultrasound data.
• Demonstrated that phantom-trained networks generalize well to in vivo data.
• Achieved superior performance compared to current deep learning methods and comparable results to optimization-based algorithms.

Conclusions:

• The proposed semisupervised method significantly improves displacement estimation in USE.
• The technique offers a robust and efficient alternative to existing deep learning and optimization-based approaches.
• This method enhances the clinical applicability of ultrasound elastography through improved accuracy.