Indocyanine green-based fluorescence imaging improved by deep learning

Abstract

Intraoperative identification of malignancies using indocyanine green (ICG)-based fluorescence imaging could provide real-time guidance for surgeons. Existing ICG-based fluorescence imaging mostly operates in the near-infrared (NIR)-I (700-1000 nm) or the NIR-IIa' windows (1000-1300 nm), which is not optimal in terms of spatial resolution and contrast as their light scattering is higher than the NIR-IIb window (1500-1700 nm). It is highly desired to achieve ICG-based fluorescence imaging in the NIR-IIb window, but it is hindered by its ultra-low NIR-IIb emission tail of ICG. Herein, we employ a generative adversarial network to generate NIR-IIb ICG images directly from the acquired NIR-I ICG images. This approach was investigated by in vivo imaging of sub-surface vascular, intestine structure, and tumors, and their results demonstrated significant improvement in spatial resolution and contrast for ICG-based fluorescence imaging. It is potential for deep learning to improve ICG-based fluorescence imaging in clinical diagnostics and image-guided surgery in clinics.

Keywords: ICG; NIR-I; NIR-IIb; deep learning; surgical navigation.