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

In vivo hyperspectral fluorescence imaging (fHSI) has transformed biomedical research by enabling qualitative/quantitative analysis of multiplexed molecular interactions. However, constraints for in vivo imaging and division of photons across numerous spectral channels create extreme photon-limited conditions, where deep signal-to-noise coupling compromises fidelity and prevents accurate analysis. Here, we present a co-designed confocal line-scanning hyperspectral light-sheet microscopy and dual-stream residual attention network with non-negative matrix factorization (DsRAN-NMF), achieving high-fidelity in vivo fHSI with up to three orders-of-magnitude improvement in photon efficiency. Advanced illumination and optical sectioning provide higher-quality initial signals, restored by our DsRAN-NMF with improved spatial and spectral fidelity, which simultaneously comprehends noise physics, high-dimensional data geometry, and hyperspectral unmixing objectives to recover biologically interpretable spectral contributions. This approach resolves highly spectrally overlapping fluorophores at micron-scale resolution in whole live zebrafish and enables visualization of nanoplastic uptake and circulation, establishing a pathway toward 4D hyperspectral imaging of living systems and nanoplastic toxicology.