Hybrid transformer-CNN network-driven optical-scanning undersampling for photoacoustic remote sensing microscopy

Area of Science:

• Biomedical Optics
• Microscopy
• Artificial Intelligence

Background:

• Imaging speed is crucial for photoacoustic microscopy (PAM) to capture dynamic biological processes and enable real-time clinical applications.
• Current methods to increase imaging speed often use high-repetition-rate lasers, risking thermal damage to samples.

Purpose of the Study:

• To develop a deep-learning-driven method for accelerating photoacoustic remote sensing (PARS) microscopy imaging acquisition.
• To maintain a constant laser repetition rate and reduce laser dosage during accelerated imaging.

Main Methods:

• A hybrid Transformer-Convolutional Neural Network (HTC-GAN) was developed to handle nonuniform sampling and motion misalignment in optical-scanning undersampling.
• A mouse ear vasculature image dataset was acquired using a customized galvanometer-scanned PARS system for training and validation.
• HTC-GAN was trained and validated on undersampled PARS data (1/2 and 1/4).

Main Results:

• HTC-GAN successfully restored high-quality images from undersampled data, closely matching ground truth.
• The method demonstrated superior performance over basic misalignment compensation algorithms and standalone CNN or Transformer networks.
• Three-dimensional imaging results confirmed the method's robustness and versatility across multiscale scanning modes.

Conclusions:

• The proposed optical-scanning undersampling method achieves a fourfold improvement in PARS imaging speed without requiring hardware upgrades.
• This deep-learning approach offers a viable solution for enhancing imaging speed in various optical-scanning microscopic systems.
• The technique reduces laser dosage, mitigating potential thermal damage to biological samples.