Application Research Of Deep Learning In High-throughput Optical Microscopy Imaging

As an essential tool for the life science research,microscopy imaging techniques intrinsically bare a limited optical throughput,as reflected in the compromise between the imaging speed and the resolution.Faster and higher-resolution imaging techniques have been desired for a long time.Deep learning single-image super-resolution techniques try to alleviate this problem by enhancing low-resolution,large field-of-view raw measurements into high-resolution ones while maintaining a large field-of-view and/or a high imaging speed,so that the imaging throughput could be promoted.However,there are still two problems remaining.On one hand,prior deep learning-based microscopy super-resolution techniques fit the inverse of the imaging process using a single convolutional neural network,ignoring the complex physical image degradation.As a result,microsopy images,especially three-dimensional images captured by the light-sheet fluorescence microscopy can not be enhanced effitively.On the other hand,most deep learning-based microscopy super-resolution techniques rely heavily on the high-resolution imags as references to train the deep neural networks in a supervised way,and fail to reconstruct high-resolution structures correctly when the high-resolution images are not available in the supervised learning.For the first problem,an image degradation model that simulates the physical process of the light sheet imaging is proposed in this study,based on which we report a dual-stage processing approach that divides the super-resolution task into two sub-problems.Using a single low-resolution,low signal-to-noise ratio image as input,the proposed method first extracts true singals from the badly contaminated measurement,and then interpolates the signals into an output with notably higher resolution and improved signal-to-noise ratio.This method is implemented with two deep convolutional neural networks with a supervised learning task,and the image degradation model is used to synthesize registered high-resolution and low-resolution image pairs as the training dataset.Combining with this technique,conventional light-sheet fluorescence microscopy reachs a two-order higher imaging throughput than before.For the second problem,this study futher performs an applicaction research of unsupervised learning in the super-resultion microscopy.An adaptive super-resolution algorithm is proposed for biological applications where the high-resolution images as the references are not available for supervised training,and where the images for training are very different from the real measurements.The proposed method uses the domain adaptation to overcome the domain shift,introduces a saliency loss and a super-resolution loss to help maintaining the structures of the signals,and successfully applies the super-resolution processing to the real measurements,reconstructing high-resolution images with high fidelity.Combined with a contact imaging device,the proposed method captures the proliferation dynamics of cells and achieves a subcellular spatial resolution of 1.6 μm over a large field-of-view.The promotion to the image resolution also benefits the downstream biological analyses like cell identification and segmentation.Without introducing any hardware modifications,the proposed methods have successfully promoted the imaging throughput of current microscopy systems by enhancing the image resolution,indicating that the proposed methods could be potential tools for a variety of biomedical assays in the field of brain science,neuroscience and cell biology.