Deep Design Produces ‘Butterfly’ Phase Mask for Light Sheet Fluorescence Microscopy

Researchers have presented a solution to a problem in light sheet fluorescence microscopy: a new illumination beam designed based on deep learning with a trainable phase mask. This work eliminates the need for advanced optical design tools and allows optimization to be applied directly to improve image contrast.

At the heart of their approach is the integration of optical propagation modeling with deep neural networks. This optimization simultaneously updates parameters of both the deep learning network and the illumination beam, resulting in superior image quality. The group’s research was published July 4 in the journal Intelligent Computing.

The authors demonstrated the effectiveness of their approach through both simulations and optical experiments, showing a significant improvement in image quality compared to conventional Gaussian light sheets. This method could potentially simplify the design of novel illumination beams, even for those without optical expertise.

This approach is analogous to an assembly line: Traditional deep learning approaches represent skilled workers operating within an established assembly line. In contrast, the new co-optimization approach incorporates worker input during the design of the assembly line to achieve better results faster.

Rather than simply analyzing images, the new deep learning model designs unexpected shapes for the illumination beam to achieve better results. Specifically, the model generated the butterfly-shaped beam by optimizing “hundreds of thousands of variables” in a phase mask.

Light sheet fluorescence microscopy has become the predominant method to image large tissue-cleared samples in 3D due to its optical sectioning, reduced photodamage, and rapid acquisition. Image quality is highly dependent on the properties of the illumination beam. Recent designs of thin, non-diffracting beams such as Bessel, Airy, and lattice light sheets have enabled uniform, high-contrast images, but new beam shapes have the potential to improve image quality, a major need for biological samples.

Collaborators on the study include Chen Li, Mani Ratnam Rai, Troy Ghashghaei, Yuheng Cai, Adele Moatti and Alon Greenbaum from the UNC/NCSU Joint Department of Biomedical Engineering.