Deep Learning-based Optical Sectioning Microscopy

As a common tool of modern biologists,light microscopes have solved many mysteries of biological structure for humans.Wide-field microscope has the basic structure of the light microscope.The out-of-focus signal captured by the wide-field microscope will reduce the contrast of the image.Optical-sectioning can acquire a clear focal plane image by activating only in-focus signal or removing the out-of-focus signal.However,traditional opticalsectioning systems require additional components in the optical path,which reduces the stability of the system and is difficult to apply under special conditions.Although the deconvolution algorithm does not require any additional components,it often requires repeated iterations to obtain results,which takes a long calculation time.Using deep learning to achieve optical sectioning can minimize the required devices,is suitable for imaging in a variety of special conditions,and reduce the restrictions on biological specimen.However,there are still challenges in applying deep learning to optical sectioning:(1)Collecting a large amount of training data required for deep learning will take a lot of manpower and material resources.(2)The speed of deep learning reconstruction is not sufficient for realtime reconstruction of high-throughput imaging.(3)It is sometimes difficult for the optical system to collect paired images as training data.In order to solve these challenges,several targeted algorithms are proposed to optimize accordingly,which lowers the threshold for implementing deep learning to optical sectioning.Single pair image learning algorithm is proposed in response to the difficulty that deep learning requires a large amount of training data.The image splitting size and neural network structure are designed according to the point spread function to minimize the required training data.The resolution and optical sectioning ability of the single pair image learning algorithm are verified using fluorescent beads and biological specimens.Moreover,the signal-to-background ratio and noise level are quantitatively compared.The reconstruction results of multiple biological specimens and multiple imaging systems are tested,and the universality of the algorithm for pictures of different microscopes and different objective lenses is studied.High-throughput optical sectioning algorithm is developed in response to accelerating the speed of deep learning reconstruction.The reconstruction throughput is greatly improved by optimizing the structure,size and reconstruction process of neuron network.The reconstructed whole brain data is compared with the structured illumination microscope imaging data,and the cell count results are compared to verify the accuracy of the reconstruction results.The effects of different labeling strategies are tested,and the uncertainty of the algorithm is analyzed,which proves the ability of the high-throughput optical sectioning algorithm to output the correct results quickly and stably for a long time.Unpaired-image learning through adversarial network algorithm is proposed in response to the problem that paired data is difficult to collect.Adversarial learning solves the problem that unpaired images cannot be trained.Training using the same specimen and different specimen data are tested,and the accuracy of the unpaired learning algorithm is studied.The signal-to-background ratio improvement of the optically sectioned image is quantitatively evaluated.Two typical applications of assisting recognizing the brain region by cytoarchitecture and uniforming imaging quality of different types of specimens demonstrate the great potential of unpaired learning algorithm.