Deep learning for in vivo near-infrared imaging

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Optical Microscopy

Background:

• Near-infrared (NIR) fluorescence imaging offers deep tissue penetration but faces challenges with light scattering.
• NIR-IIb window (1,500-1,700 nm) provides superior imaging but requires toxic nanoparticle probes.
• NIR-I/IIa windows (700-1,300 nm) use safer probes but yield lower image quality due to scattering.

Purpose of the Study:

• To develop a deep learning method for enhancing NIR-I/IIa fluorescence images to achieve NIR-IIb image quality.
• To enable high-performance in vivo imaging using biocompatible, FDA-approved fluorescent probes.
• To improve resolution and signal-to-background ratios in noninvasive NIR imaging modalities.

Main Methods:

• Artificial neural networks were trained to translate fluorescence images from NIR-I/IIa to NIR-IIb.
• Deep learning was applied to in vivo lymph node imaging using indocyanine green (ICG).
• The method was tested on preclinical imaging of PD-L1 and EGFR using IRDye-800 and on NIR-II light-sheet microscopy (LSM).

Main Results:

• Deep learning translation achieved a signal-to-background ratio >100 for in vivo lymph node imaging with ICG.
• Tumor-to-normal tissue ratios were enhanced up to ~20 from ~5 for PD-L1/EGFR imaging, improving tumor margin localization.
• Deep learning significantly improved resolution and signal/background in noninvasive NIR-II LSM.

Conclusions:

• Deep learning effectively enhances NIR fluorescence imaging quality, mimicking NIR-IIb performance with safer probes.
• This approach facilitates high-resolution, deep-tissue in vivo imaging for biomedical research.
• The technology holds promise for improved clinical diagnostics and image-guided surgery.