Deep Learning-enabled Restoration Of Light-sheet Fluorescence Microscopy Image Of Scattering Medium

Optical scattering occurs when a light beam passes through scattering medium,such as haze in the sky,seawater and biological tissues,resulting in a reduction in the focusing ability of the light beam through scattering medium.In this case,the scattering prevents the detector from obtaining information from the object,thereby limiting its applications in optical fiber communication,biomedical imaging and other fields.How to realize high-quality imaging through scattering medium has become an important challenge to be addressed.In this dissertation,a deep-learning based method was studied,which can restore image degradation caused by tissue scattering when imaging thick biological specimens using light-sheet fluorescence microscopy(LSFM).This approach can significantly improve the singal-to-noise ratio(SNR)and spatial resolution of 3D LSFM image,whle eliminate the need of complicated optical setup or iterative recovery algorithms.What we have achieved in this dissertation include:A deep neural network was designed to learn the mapping from degraded light-scattered image to the high-resolution label image.Then we used multi-view light-sheet fluorescence microscopy in conjuncation with 3D image registration,to obtain training dataset that contain registered low-quality scatterd data and high-quality label data of the same biological samples.After training of the network,it is capable of restoring degradation by light scattering in new3 D image,and providing a high-SNR high-resolution output.The performance of this scattering neural network were veried by restoring light-sheet fluorescence microscopy image of Drosophila embryo,clarified mouse brain and 3D cultured cell cluster,demonstrating its broad applications in different biological specimens.Finally,the quality of network restoration was also quantitatively analyzed.