Reconstructing Undersampled Photoacoustic Microscopy Images Using Deep Learning.

Area of Science:

• Biomedical optics
• Medical imaging
• Machine learning in imaging

Background:

• Photoacoustic microscopy (PAM) faces a trade-off between spatial resolution and imaging speed.
• Reconstructing high-resolution images quickly is crucial for in vivo applications.

Purpose of the Study:

• To develop a deep learning model for reconstructing undersampled photoacoustic microscopy images.
• To overcome the inherent limitations in PAM imaging speed and resolution.

Main Methods:

• Utilized a Fully Dense U-net (FD U-net) convolutional neural network architecture.
• Artificially downsampled fully-sampled PAM images of mouse brain vasculature to simulate undersampling.
• Trained and tested the deep learning model under various downsampling ratios.

Main Results:

• The FD U-net model demonstrated robust performance in reconstructing PAM images from minimal data.
• Reconstruction was successful even with as few as 2% of the original pixels.
• The method effectively shortened imaging time without substantial loss of image quality.

Conclusions:

• Deep learning, specifically the FD U-net, offers a powerful solution to the resolution-speed trade-off in PAM.
• This technique enables faster PAM acquisition while maintaining diagnostic image quality.
• The developed model has significant potential for improving in vivo imaging of biological structures like brain vasculature.