Photoacoustic microscopy with sparse data by convolutional neural networks.

Area of Science:

• Biomedical Imaging
• Machine Learning
• Optics

Background:

• Photoacoustic microscopy (PAM) offers high resolution and contrast but is limited by slow imaging speeds due to its point-by-point scanning mechanism.
• Accelerating PAM image acquisition is crucial for expanding its applications, particularly in dynamic biological processes and clinical settings.

Purpose of the Study:

• To develop and validate a deep learning method using convolutional neural networks (CNNs) to reconstruct high-quality sparse PAM images.
• To significantly increase PAM imaging speed while preserving or improving image quality.

Main Methods:

• A CNN model incorporating attention modules, residual blocks, and perceptual losses was designed.
• The model performs image reconstruction, mapping low-sampling sparse PAM images (1/4 or 1/16 sampling) to a latent fully-sampled representation.
• Training and validation were conducted using PAM images of leaf veins, with further testing on *in vivo* mouse ear and eye blood vessel images.

Main Results:

• The CNN model demonstrated significant quantitative and qualitative improvements in reconstructing sparse PAM images of leaf veins.
• The method effectively enhanced the quality of *in vivo* PAM images of blood vessels, improving clarity and detail.
• The approach successfully accelerated image acquisition while maintaining diagnostic image quality.

Conclusions:

• The proposed CNN-based method offers a viable solution for enhancing sparse PAM images, overcoming the speed limitations of traditional PAM.
• This technique holds promise for facilitating faster PAM acquisition and enabling broader clinical applications, especially for imaging microvasculature.
• The study highlights the potential of deep learning in advancing biomedical imaging technologies.