Deep-learning-enhanced Light-sheet Microscopy For High-throughput Imaging Of Large Biological Tissue

Fluorescence microscopy has become a crucial tool for the life science research,owing to its strong ability to visualize the specifically-labeled cellular structures in biological specimens.Nowadays,high-resolution imaging of entire macro-scale samples,has posed big challenge to current fluorescence microscopy,especially for the applications in tissue pathology,neurobiology,etc.In the past decades,light-sheet microscopy has emerged as tool,as its selective plane illumination can provide higher axial resolution and throughput.However,conventional Gaussian light-sheet microscope still suffers from a tradeoff between the high axial resolution and the large field-of-view,thereby causing limited optical throughput suboptimal for the intoto imaging of large biological samples.The generation of thin-and-wide light-sheet illumination and development of efficient algorithm that can computationally enhance the image resolution are both potential solutions to the above-mentioned issues.This dissertation describes how to improve the throughput of current light-sheet microscopy based on the novel light-sheet illumination design and deep-learning-enhanced image restoration algorithm,which together allow fast and high-resolution 3D imaging of diverse large biological samples,such as mouse brain and lung,with throughput increased by 1-2 orders of magnitude.The main accomplishments in this dissertation can be summarized as following three parts:Firstly,an add-on light-sheet microscopy with deep-learning-enhanced axial resolution is proposed to enable fast,isotropic light-sheet fluorescence imaging on a conventional widefield microscope.Following a minimized device that transforms an inverted wide-field microscope into a light-sheet microscope,a deep-learning algorithm is developed to improve the axial resolution of the raw images.The combination of add-on light-sheet microscopy with the convolution neuron network achieves isotropic 3 um spatial resolution,thus reaching a five times higher imaging throughput than the traditional Gaussian light-sheet microscopy.Secondly,an axially swept light-sheet microscopy with the deep-learning-enhanced 3D resolution is proposed to improve the low axial resolution of the conventional static Gaussian light sheet for imaging intact organs and tissues.Utilizing a customized spinning disk,the proposed system achieves rapid scanning of the beam waist of the light sheet along the propagation direction,enlarging the confocal range of the conventional Gaussian light sheet and realizing high-throughput isotropic 3D imaging.The spatial resolution has been improved by a factor of 4 in all three dimensional using the convolution network,providing a 2～3 order higher optical imaging throughput than conventional Gaussian light-sheet microscopy.Benefited by the ultra-high optical imaging throughput,whole mouse brain imaging can be achieved on a sub-minute time scale.At last,a non-diffraction light-sheet microscope based on the double-ring mask and the deep learning algorithm for eliminating the side lobes is proposed.Combined with a customized sample holder,a double-sided illumination non-diffractive light-sheet imaging system based on the double-ring mask is designed and can realize high-speed imaging with continuous field-of-view ranging from 1.26 × to 12.6 ×.A deep-learning-based image processing algorithm is then applied to enhance the image resolution while eliminating the side lobes.Integrating the whole pipeline in Python,an automatic program of image enhanment,screening and analyzing increases the throughput of the entire process by 3 orders of magnitude,as compared with traditional methods.The proposed deep learning based light-sheet microscopy enable high-throughput imaging of biological organs and tissues,satisfying the requirement of single-cell resolution imaging and providing powerful tools for the studies of neuroscience,histology,and pathology in the field of biomedical research.