Convolutional Neural Network-Based Speckle Tracking for Ultrasound Strain Elastography: An Unsupervised Learning...

Area of Science:

• Medical imaging
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Accurate motion estimation is vital for real-time ultrasound strain elastography (USE).
• Supervised deep learning models often rely on simulated data, raising concerns about their efficacy with complex in vivo motion.
• Existing methods face challenges in reliably tracking speckle motion for clinical applications.

Purpose of the Study:

• To develop an unsupervised deep learning network for motion estimation in ultrasound strain elastography.
• To address the limitations of supervised learning approaches using simulated data.
• To improve the accuracy and reliability of strain estimation in USE.

Main Methods:

• An unsupervised motion estimation neural network (UMEN-Net) was developed by adapting the PWC-Net architecture.
• The network utilizes predeformation and postdeformation radio frequency (RF) echo signals as input.
• A novel loss function incorporating signal correlation (using GOCor module), displacement field smoothness, and tissue incompressibility was employed.

Main Results:

• The UMEN-Net demonstrated superior performance on simulated, phantom, and in vivo breast lesion data.
• It achieved higher signal-to-noise ratios (SNRs) and contrast-to-noise ratios (CNRs) for axial strain estimates.
• The network significantly improved the quality of lateral strain estimates compared to state-of-the-art methods.

Conclusions:

• The developed unsupervised CNN model offers a robust solution for motion estimation in ultrasound strain elastography.
• UMEN-Net overcomes the limitations of supervised methods by effectively handling complex in vivo motion.
• This advancement holds promise for more accurate and reliable clinical diagnosis using USE.