Rapid tissue oxygenation mapping from snapshot structured-light images with adversarial deep learning.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Spatial frequency-domain imaging (SFDI) is crucial for wide-field tissue oxygen saturation mapping.
• Current SFDI methods face limitations, including multi-image requirements or accuracy/artifact issues with single-snapshot methods.

Purpose of the Study:

• To introduce OxyGAN, a novel data-driven method for estimating tissue oxygenation from single structured-light images.
• To develop a fast and accurate alternative to conventional SFDI techniques.

Main Methods:

• OxyGAN employs supervised generative adversarial networks for end-to-end estimation of tissue oxygenation.
• Ground truth data was acquired using conventional SFDI on various ex vivo and in vivo tissue samples.
• Performance was benchmarked against single-snapshot optical properties (SSOP) and a hybrid deep learning-physical model approach.

Main Results:

• OxyGAN achieved 96.5% accuracy on human feet and 93% on unseen tissue types, demonstrating robustness.
• It outperformed SSOP and the hybrid model by over 24% in accuracy.
• Optimized inference speed enables video-rate (25 Hz) imaging, approximately 10x faster than prior methods.

Conclusions:

• OxyGAN offers rapid, high-fidelity tissue oxygenation mapping from single images.
• Its speed and accuracy position it for significant potential in diverse clinical applications.
• The method overcomes limitations of existing SFDI techniques, enabling real-time monitoring.