【作者】 刘中法；

【导师】 杨勇；

【作者基本信息】 南开大学 ， 光学工程， 2022， 硕士

【摘要】 近年来,光学显微成像技术的飞速发展,使人们能更好的探索与研究生物细胞中更精细的微观结构,甚至可实现对细胞内部的蛋白酶、细胞器等亚细胞结构分布和含量的实时定量分析。然而,不同光学显微成像技术存在的固有局限性,制约了人们对生物细胞图像的正确理解。因此,对获取的细胞或组织图像进行必要的处理,可以增强图像中包含的生物学特征,弥补不同显微成像技术的缺陷。传统的生物医学图像处理技术包括图像分割、细胞计数、图像融合、像素变换等,可以实现对细胞或组织结构中特定区域的定性和定量分析,提高医学诊断或生物研究的准确性。但由于传统的生物医学图像处理技术过度依赖图像的空间区域特征和梯度信息,需要耗费大量人工制定特征提取与图像重建的规则,使这些技术表现出人工主观性强、系统鲁棒性差、效率低等问题。基于上述问题,本文致力于在光学显微成像的基础上,研究应用深度学习技术以增强生物细胞图像质量的方法,分别开展了细胞图像融合和虚拟相衬成像的研究工作。本文提出的生物医学图像处理方法基于不同的深度神经网络,且均为无标记处理,能够显著增强细胞图像中包含的生物学特征。通过主观评价和客观评价方法,验证了所提方法的有效性。具体内容包括:(1)针对细胞折射率显微图像信息损失严重的问题,提出了一种基于Fusion CNN框架将细胞折射率显微图像与明场显微图像相融合的方法。通过卷积神经网络学习、提取两种源图像的关键特征,构建、训练融合模型。该模型可实现端到端的融合图像输出,利用明场显微图像中的形态细节信息,增强折射率显微图像中的生物特征信息。通过与基于小波变换的图像融合方法和基于梯度转移的图像融合方法的比较,主观和客观(PSNR、ENT、AG)评价结果验证了所提方法的有效性。(2)基于Cycle GAN框架,提出了一种可实现无监督的细胞明场显微图像到相衬显微图像的转换方法。对同一条件下采集的细胞明场显微图像和相衬显微图像进行自动配准,形成数据训练集。构建基于Cycle GAN框架的深度学习算法,利用训练集数据,学习特征提取和图像重建规则,获得明场显微图像到相衬显微图像的转换模型,实现了仅需采集明场显微图像即可快速预测出相应的虚拟相衬显微图像的效果。通过主观评价和客观评价(PSNR、COSIN)方法验证了本文所提方法的有效性。取得的创新性研究成果如下:(1)针对获得的石墨烯基细胞折射率显微图像存在分辨率不足、生物特征信息不明显的问题,首次提出了一种采用卷积神经网络算法实现对细胞折射率显微图像与明场显微图像相融合的方法,利用明场显微图像中所包含的丰富的细胞形态信息改善折射率显微图像的质量,有效提高了细胞折射率显微图像中亚细胞结构的生物特征辨识度,在细胞筛选、识别等领域均具有潜在的应用价值。(2)针对普通光学显微镜无法实现相衬成像的问题,首次提出了一种基于深度学习算法实现细胞虚拟相衬显微成像的方法,仅需要获取细胞的明场显微图像,即可快速预测出相应的相衬显微图像。这种方法特别适用于需要同时获取实时明场和相衬显微图像的实验场合,且无需购置价格昂贵的相衬显微镜。此外,该方法在实验教学及快速检测领域也具有潜在应用价值。上述研究工作,拓宽了深度学习算法在视觉图像处理领域的研究思路,探索了利用不同显微成像技术所获得的图像信息进行有效融合及相互转换的方法。此外,本研究结合了本课题组开展的基于石墨烯基光学生物传感器的高特异性显微成像技术,为推动无标记、高灵敏度、高特异性、高空间分辨率的活细胞显微成像技术的发展打下基础。我们相信该技术有望成为一种在分子水平研究细胞生命活动的重要研究手段。更多还原

【Abstract】 In recent years,the rapid development of optical microscopic imaging techniques has enabled better exploration and study the finer microstructures in biological cells,and even real time quantitative analysis of the distribution and content of subcellular structures such as proteases and organelles inside cells.However,the inherent limitations of different optical microscopic imaging techniques constrain the proper understanding of biological cell images.Therefore,by performing the necessary processing of the acquired cell or tissue images,the biological features to be observed can be enhanced and the deficiencies of the different microscopic imaging techniques can be compensated.Traditional biomedical image processing techniques include image segmentation,cell counting,image fusion,pixel transformation,etc.,which can achieve qualitative or quantitative analysis of specific regions of a cell or tissue structure and improve the accuracy of medical diagnosis or biological research.However,because these techniques rely excessively on the spatial region features and gradient information of the images,they require a lot of manual work to develop rules for feature extraction and image reconstruction,making these techniques exhibit problems such as high manual subjectivity,poor system robustness and low efficiency.Based on the above-mentioned problems,this paper is devoted to the application of deep learning techniques to enhance the quality of biological cell images based on optical microscopy,and carries out research work on cell image fusion and virtual phase contrast imaging,respectively.The proposed biomedical image processing methods are based on different deep neural networks,both label-free,and can significantly enhance the biological features in the cell images.The effectiveness of the proposed method is verified by using subjective and objective evaluation methods.The specific contents include:(1)To address the problem of severe information loss in cell refractive index microscopic images,a method is proposed to fuse cell refractive index microscopic images with bright-field microscopic images based on the Fusion CNN framework.The trained fusion model is constructed by learning and extracting the key features of the two source images through convolutional neural networks.The model enables end-toend fused image output,using morphological detail information in bright-field microscopic images to enhance biometric information in refractive index microscopic images.The effectiveness of the proposed method is verified by comparing the subjective and objective(PSNR,ENT,AG)evaluation results with the wavelet transform(WL)based image fusion method and the gradient transfer(GTF)based image fusion method.(2)Based on the Cycle GAN framework,a method is proposed to achieve unsupervised conversion of cellular bright-field microscopic images to phase contrast microscopic images.The bright-field microscopic images and phase contrast microscopic images of cells acquired under the same conditions are automatically aligned to form a training set.A deep learning algorithm based on the Cycle GAN framework was constructed to learn feature extraction and image reconstruction rules using the training set data to obtain a conversion model from bright-field microscopic images to phase contrast microscopic images.The corresponding virtual phase contrast microscopic images can be quickly predicted by acquiring only bright-field microscopic images.The effectiveness of the proposed method is verified by subjective evaluation and objective evaluation(PSNR,COSIN)methods.The innovative research results achieved are as follows:(1)To address the problems of insufficient resolution and inconspicuous biometric information in the obtained graphene-based cell refractive index microscopic images,a method of fusion of cell refractive index microscopic images with bright-field microscopic images using convolutional neural network algorithm is proposed for the first time,which improves the quality of refractive index microscopic images by using the rich cell morphology information contained in the bright-field microscopic images,and effectively improves the recognition of biometric features of subcellular structures in cell refractive index microscopic images.The method can improve the quality of refractive index micrographs by using the rich cell morphology information contained in the bright-field micrographs,and effectively improve the biometric recognition of subcellular structures in micrographs,which has potential application values in cell screening and identification.(2)To address the problem that ordinary optical microscopy cannot realize phase contrast imaging,a method based on deep learning algorithm is proposed to realize virtual phase contrast images of cells for the first time,which only requires the brightfield microscopic images and can quickly predict the corresponding phase contrast microscopic images.This method is particularly suitable for experimental situations where real-time bright-field and phase contrast images need to be acquired simultaneously,and expensive symmetric microscopes are not required.In addition,the method has potential application values in the fields of experimental teaching and rapid detection.The above research work broadens the research ideas of deep learning algorithms in the field of visual image processing,and explores the method of effective fusion and interconversion of image information obtained by using different microscopic imaging techniques.In addition,this study combines the high specificity microscopic imaging technology based on graphene-based optical biosensors carried out by our group,which lays the foundation for promoting the development of label-free,high sensitivity,high specificity,and high spatial resolution microscopic imaging technology of living cells.We believe that the technology is expected to provide an important research tool for studying cellular life activities at the molecular level.更多还原