Deep learning-driven quantitative spectroscopic photoacoustic imaging for segmentation and oxygen saturation...

Area of Science:

• Biomedical Imaging
• Optical Physics
• Machine Learning

Background:

• Spectroscopic photoacoustic (sPA) imaging offers noninvasive in vivo estimation of blood oxygenation saturation (sO2).
• Accurate sO2 quantification relies on precise optical fluence estimation, which is challenging in heterogeneous tissues due to varying light absorption and scattering.
• Robust modeling of light transport is crucial for reliable sPA imaging results.

Purpose of the Study:

• To develop a deep neural network (Hybrid-Net) for simultaneous sO2 estimation and vessel segmentation in sPA imaging.
• To improve the accuracy of sO2 measurements by minimizing errors within segmented blood vessels.
• To enable accurate blood oxygenation estimation without requiring explicit optical fluence calculations.

Main Methods:

• Developed Hybrid-Net, a deep neural network for sPA imaging.
• Trained Hybrid-Net on simulated 3D Monte Carlo data of light transport in breast tissue at 700 nm and 850 nm.
• Retrained and validated Hybrid-Net on experimental sPA data from tissue-mimicking phantoms with embedded blood pools.

Main Results:

• Hybrid-Net achieved high segmentation accuracy (≥0.978 in simulations, 0.999 in experiments) across various noise levels.
• Demonstrated low sO2 mean squared error (≤0.048 in simulations, 0.002 in experiments).
• Successfully estimated sO2 in blood vessels without needing optical fluence estimates.

Conclusions:

• Hybrid-Net provides accurate blood oxygenation saturation estimation in sPA imaging.
• The developed method effectively segments blood vessels and quantifies sO2 within them.
• This approach has the potential to significantly advance in vivo sO2 monitoring.