Distribution-informed and wavelength-flexible data-driven photoacoustic oximetry

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Photoacoustic imaging (PAI) shows potential for blood oxygen saturation measurement.
• Current spectral unmixing methods lack accuracy and robustness for PAI.
• Accurate blood oxygenation estimation has significant clinical applications, including cancer detection and inflammation quantification.

Purpose of the Study:

• To address the inflexibility of existing data-driven methods for blood oxygenation estimation in PAI.
• To introduce a novel recurrent neural network architecture for improved PAI data analysis.

Main Methods:

• Developed a wavelength-flexible recurrent neural network architecture using a long short-term memory network.
• Created 25 simulated training dataset variations to evaluate network performance.
• Proposed the Jensen-Shannon divergence to identify the most suitable training dataset for specific applications.

Main Results:

• The proposed network architecture flexibly handles varying input wavelengths.
• The recurrent neural network outperforms linear unmixing and prior learned spectral decoloring methods.
• Jensen-Shannon divergence correlates with estimation error, enabling prediction of optimal training datasets.

Conclusions:

• A flexible, data-driven network architecture combined with Jensen-Shannon divergence offers a promising approach for robust photoacoustic oximetry.
• This method has the potential to enable clinical applications of data-driven photoacoustic oximetry.
• Further development could lead to more accurate and reliable blood oxygenation measurements in clinical settings.