Deep Learning With Physics-Embedded Neural Network for Full Waveform Ultrasonic Brain Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound imaging is a safe, affordable non-invasive technique for soft tissue examination.
• Acoustic impedance mismatch between the skull and brain tissue limits conventional ultrasound's diagnostic utility for the brain.
• Developing advanced imaging techniques is crucial for non-invasive brain diagnostics.

Purpose of the Study:

• To introduce a novel deep learning approach, physics-embedded neural network with full waveform inversion (PEN-FWI), for quantitative brain tissue imaging.
• To overcome the limitations of traditional ultrasound in brain imaging due to skull interference.
• To demonstrate the capability of PEN-FWI for high-resolution brain imaging.

Main Methods:

• Developed a physics-embedded neural network with deep learning based full waveform inversion (PEN-FWI).
• The network comprises a forward convolutional neural network (FCNN) for wavefield prediction and an inversion sub-neural network (ISNN) for model reconstruction.
• Employed an iterative approach embedding the FCNN within the ISNN for wavefield-to-brain model tomography.

Main Results:

• PEN-FWI successfully generated high-quality images of skull and soft tissues, even from a homogeneous water model.
• Achieved excellent imaging of clot models with various velocity distributions (uniform, Gaussian, irregular).
• Demonstrated robust differentiation of brain slices and skulls, with a horizontal cross-sectional brain image generated in just 1.13 seconds.

Conclusions:

• PEN-FWI provides reliable quantitative imaging of brain tissues, overcoming skull-related ultrasound limitations.
• The algorithm offers a significant advancement for ultrasound-based brain tomography.
• PEN-FWI shows potential for broad applications in other imaging fields requiring accurate subsurface visualization.