Hyperspectral LSFM with DMD-only shaping and neural network reconstruction

Quantitative fluorescence imaging in biological samples with autofluorescence or spectrally overlapping fluorophores remains a challenge for conventional light sheet fluorescence microscopy (LSFM). To address this, we introduce DNR-HLSFM, an approach for hyperspectral LSFM (HLSFM) that is based on DMD-only light shaping and neural network reconstruction (DNR). This computational technique combines structured illumination with a physical model of the acquisition process, a denoising neural network and non negative unmixing, to enable robust fluorophore quantification. DNR-HLSFM does not rely on optical filters, which overcomes limitations in spectral resolution and signal loss. We experimentally demonstrate our method by imaging zebrafish embryos in three dimensions, achieving effective autofluorescence removal and precise separation of two spectrally overlapping red fluorophores. This work may pave the way for versatile, high-resolution, quantitative hyperspectral imaging in biomedical research, significantly expanding the capabilities of LSFM. To promote reproducibility and foster further development, we have made our raw data, reconstructed data, and code publicly available.