Ultrafast Cardiac Imaging Using Deep Learning for Speckle-Tracking Echocardiography.

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• High-quality ultrafast ultrasound imaging relies on compounding multiple wave transmissions, which can reduce frame rates and introduce motion artifacts.
• Existing deep learning methods excel at static image reconstruction but haven't been fully assessed for dynamic cardiac motion tracking.

Purpose of the Study:

• To evaluate a deep learning approach for ultrafast ultrasound image reconstruction that maintains cardiac motion tracking capabilities.
• To assess the performance of a complex-weighted convolutional neural network (CNN) combined with speckle tracking for echocardiography.

Main Methods:

• Developed a complex-weighted CNN for image reconstruction and integrated it with a state-of-the-art speckle-tracking method.
• Validated the approach using a custom simulation framework for artifact-free cardiac images.
• Tested performance on experimental in vitro data with a spinning disk phantom and in vivo human datasets.

Main Results:

• The CNN approach, using only three diverging waves (DWs), achieved image quality and motion accuracy comparable to compounding 31 DWs without motion artifacts.
• The method demonstrated high-quality reconstruction and motion estimation across various velocities in phantom experiments.
• Outperformed a state-of-the-art motion-compensated (MoCo) approach at high velocities and showed consistent improvements on in vivo data.

Conclusions:

• Deep learning reconstruction is feasible and effective for ultrafast speckle-tracking echocardiography.
• The proposed CNN-based method enhances both image quality and motion estimation accuracy in dynamic ultrasound imaging.
• This approach offers a significant advancement over standard compounding and MoCo techniques, particularly for high-velocity scenarios.