Deep Learning-Based Framework for Fast and Accurate Acoustic Hologram Generation.

Area of Science:

• Acoustics
• Artificial Intelligence
• Holography

Background:

• Acoustic holography is crucial for applications like noncontact manipulation and medical imaging.
• Current acoustic hologram generation methods are slow and inaccurate, limiting novel applications.

Purpose of the Study:

• To propose a deep learning framework for fast and accurate acoustic hologram generation.
• To enable unsupervised training of acoustic hologram generation models.

Main Methods:

• Developed a deep learning framework with an autoencoder-like architecture.
• Introduced the holographic ultrasound generation network (HU-Net) for unsupervised learning.
• Proposed a novel loss function for energy-efficient holograms and a physical constraint (PC) layer for device compatibility.

Main Results:

• The framework achieved hundreds of times faster generation speeds than IASA and Diff-PAT in simulations.
• Demonstrated comparable or superior reconstruction quality compared to existing methods.
• Validated experimentally with 3-D printed lenses and a 2-D ultrasound array.

Conclusions:

• The proposed deep learning framework offers a fast and accurate solution for acoustic hologram generation.
• It serves as a valuable alternative tool for existing applications and can expand novel medical applications.