Deep Learning-enabled Image Restoration For Fluorescence Microscopy In Scattering Tissues

In the field of biophotonic imaging,scattering mainly accounts for the degeneration of microscopic imaging quality,since biological tissues are anisotropic and opaque without tissue clearing methods,which is the main obstacle to high-quality microscopic imaging and functional research.This thesis proposes a deep-learning-based biological microscopy imaging restoration method,which is suitable for a variety of microscopy imaging modes and biological samples,that significantly improves imaging quality without requiring complex hardware equipment.The improvement in resolution and signal-to-noise ratio of images makes it possible for further life science research of the sample.The main contents of this thesis:(1)We constructed training datasets for different samples through wide-field,light-sheet,and two-photon fluorescence microscopy imaging techniques.In order to obtain pairs of scattering and target images,we designed sample flipping experiments and scattering medium superposition experiments on the basis of light sheet microscopy imaging and two-photon microscopy imaging,respectively.By the image registration method based on interest points detection,we finally obtained the pixel-wise aligned scattering datasets.(2)We used a three-dimensional convolutional neural network(3DCNN)based on the Tensorflow framework and U-Net structure,which was suitable for the recovery of biological tissue scattering,and applied it to our self-built scattering datasets.Through the encodingdecoding model of the U-Net network,the inverse mapping process from the scattered images to the label images was constructed.Furthermore,some good reconstruction results were achieved.(3)We applied the three-dimensional affine transformation algorithm to enhance the original datasets,used the specified filter to remove no signal regions for data cleaning,etc.,to improve the performance of the network.(4)In this thesis,the accuracy of the method was verified by experiments of USAF resolution target and fluorescent beads,and the applicability for in-field biological tissues imaging was verified by experiments of Drosophila embryos and mouse brain neurons.It was proved that our method could be widely used in different microscopic imaging system for various biological samples.