Deep-Learning-Driven High-Fidelity In Vivo Hyperspectral Fluorescence Imaging Under Extreme Photon-Limited Conditions

Area of Science:

• Biomedical Imaging
• Optical Microscopy
• Artificial Intelligence

Background:

• In vivo hyperspectral fluorescence imaging (fHSI) is crucial for analyzing molecular interactions but faces photon-limited conditions.
• Signal-to-noise ratio degradation compromises imaging fidelity and accurate analysis in current fHSI techniques.

Purpose of the Study:

• To develop a high-fidelity in vivo fHSI method with improved photon efficiency.
• To enhance the analysis of multiplexed molecular interactions and nanoplastic toxicology in living systems.

Main Methods:

• Co-designed confocal line-scanning hyperspectral light-sheet microscopy.
• Developed a dual-stream residual attention network with non-negative matrix factorization (DsRAN-NMF).
• Integrated advanced illumination and optical sectioning with AI-driven signal restoration.

Main Results:

• Achieved up to a three-orders-of-magnitude improvement in photon efficiency for in vivo fHSI.
• DsRAN-NMF restored signals with enhanced spatial and spectral fidelity, overcoming noise limitations.
• Successfully resolved spectrally overlapping fluorophores at micron resolution in live zebrafish and visualized nanoplastic circulation.

Conclusions:

• The developed DsRAN-NMF approach significantly enhances in vivo fHSI capabilities.
• This method enables high-resolution, quantitative analysis of complex biological processes and environmental toxicology.
• Establishes a foundation for 4D hyperspectral imaging in living systems.