Moving Beyond Simulation: Data-Driven Quantitative Photoacoustic Imaging Using Tissue-Mimicking Phantoms.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Accurate optical absorption coefficient measurement is crucial for molecular quantification in photoacoustic imaging (PAI).
• The domain gap between simulated and experimental data hinders quantitative PAI in living systems.
• Existing methods struggle with artifacts and biases due to this simulation-experiment gap.

Purpose of the Study:

• To bridge the domain gap in PAI by developing experimentally validated phantoms and their digital twins.
• To enable supervised training of deep learning models (U-Net) directly on experimental PAI data.
• To improve pixel-wise estimation of optical absorption coefficients for enhanced molecular mapping.

Main Methods:

• Creation of a novel dataset of experimentally characterized imaging phantoms and their digital twins.
• Supervised training of a U-Net architecture using experimental PAI data for absorption coefficient estimation.
• Comparison of results with traditional Monte Carlo-based fluence correction methods.

Main Results:

• Training U-Nets on simulated data introduced artifacts and biases, confirming the domain gap.
• Training on experimental data yielded more accurate and robust optical absorption coefficient estimates.
• Deep learning approaches recovered depth-dependent signal loss in a mouse model, outperforming simulation-trained models.

Conclusions:

• Supervised learning with experimentally acquired phantom data significantly improves quantitative PAI.
• This approach shows promise for advancing molecular concentration mapping and clinical insights from PAI.
• Further research is needed to reduce the absolute quantification error but the potential of deep learning is evident.