CNN-Based Image Reconstruction Method for Ultrafast Ultrasound Imaging

Area of Science:

• Biomedical Imaging
• Medical Ultrasound
• Artificial Intelligence in Medicine

Background:

• Ultrafast ultrasound (US) imaging offers high frame rates (>1 kHz) for advanced applications like elastography and neuroimaging.
• A key limitation of ultrafast US is the presence of significant diffraction artifacts, degrading image quality.
• Current methods to mitigate artifacts often require multiple acquisitions, reducing the effective imaging speed.

Purpose of the Study:

• To develop a real-time, single-acquisition image reconstruction method for ultrafast US.
• To effectively remove diffraction artifacts inherent to ultrafast US imaging.
• To achieve high-quality image reconstruction from single unfocused acquisitions.

Main Methods:

• A two-step convolutional neural network (CNN) based reconstruction approach was developed.
• The first step uses backprojection to generate a low-quality estimate.
• A residual CNN with multiscale/multichannel filtering, trained with a novel MSLAE loss function, refines the image and removes artifacts.

Main Results:

• The proposed CNN method reconstructs high-quality images from single plane-wave (PW) acquisitions.
• Image quality is comparable to gold-standard synthetic aperture imaging, even with a dynamic range over 60 dB.
• Simulated training data proved effective for both in vitro and in vivo experimental data.

Conclusions:

• The developed CNN method enables high-quality, artifact-free ultrafast US imaging from single acquisitions.
• This approach significantly enhances the potential for real-time biomedical imaging applications.
• The method demonstrates robustness, performing well across simulated, in vitro, and in vivo datasets.