Deep learning-enabled high-speed, multi-parameter diffuse optical tomography.

Area of Science:

• Medical Imaging
• Biomedical Optics
• Computational Imaging

Background:

• Frequency-domain diffuse optical tomography (FD-DOT) offers potential for enhanced clinical breast tumor characterization.
• Conventional diffuse optical tomography (DOT) reconstruction algorithms are computationally intensive and require expert tuning, limiting real-time feedback.
• Deep learning (DL) can front-load computational costs, enabling faster and more accurate DOT reconstructions.

Purpose of the Study:

• To demonstrate simultaneous 3D absorption and reduced scattering coefficient reconstruction using DL-FD-DOT.
• To achieve real-time imaging capabilities with a handheld probe for breast imaging.

Main Methods:

• A DL model was trained to solve the DOT inverse problem.
• The model utilized a simulated FD-DOT dataset emulating handheld probe imaging for human breasts.
• The DL-DOT model was tested using both synthetic and experimental data.

Main Results:

• The DL-DOT model significantly reduced root mean square error and crosstalk compared to model-based tomography.
• Spatial similarity and anomaly contrast accuracy were substantially increased by the DL-DOT model.
• Reconstruction time was dramatically reduced from 3.8 minutes to 0.02 seconds, enabling real-time imaging.
• The model's efficacy was verified using tumor-emulating optical phantoms.

Conclusions:

• DL-FD-DOT shows significant clinical potential for real-time functional imaging of human breast tissue.
• This AI-driven approach overcomes limitations of conventional DOT, paving the way for improved diagnostic tools.