Semantic segmentation of multispectral photoacoustic images using deep learning

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Multispectral photoacoustic (PA) imaging offers rich physiological data for functional medical imaging.
• Clinical adoption of PA imaging is hindered by the complexity of high-dimensional data.
• Interpreting PA images requires robust methods for data conversion into clinically relevant information.

Purpose of the Study:

• To develop a deep learning-based semantic segmentation algorithm for multispectral PA images.
• To improve the interpretability and clinical relevance of PA imaging data.
• To facilitate the translation of PA imaging technology into clinical practice.

Main Methods:

• A supervised deep learning approach was employed for semantic segmentation.
• The algorithm was trained using manually annotated PA and ultrasound imaging data.
• Validation was performed using experimental data from 16 healthy human volunteers.

Main Results:

• The deep learning algorithm successfully performed automatic tissue segmentation of multispectral PA images.
• The segmentation enabled powerful analyses and visualizations of the complex PA data.
• The approach demonstrated the potential for intuitive representation of high-dimensional PA information.

Conclusions:

• Deep learning-based semantic segmentation is a valuable tool for enhancing PA image interpretability.
• Automated segmentation facilitates the analysis and visualization of multispectral PA data.
• This preprocessing method can significantly aid in the clinical translation of PA imaging.