超低信噪比冷冻电镜图像的深度学习去噪算法—DWT-CAE

Noise Reduction Algorithm on the Extremely Low Signal-noise Ratio CryoEM Images Based on Deep Learning Methods-DWT-CAE

冷冻电镜成像技术是获取蛋白质等生物分子结构的重要途径之一,对研究蛋白质功能特性以及在制药、医疗、疾病防治等方面的应用有着重要意义.针对冷冻电镜图像的大数据量和超低信噪比特征,本文着重研究了冷冻电镜图像去噪和颗粒挑选的方法.结合卷积神经网络模型和自动编码机模型,提出了用于去噪的EM-CAE(Electron Microscopy-Convolutional AutoEncoder)方法,并在实验中验证了算法的效果.针对原始图像噪声的复杂特点,本文对EM-CAE方法做进一步改进,将自动编码机模型与小波变换相结合,提出DWT-CAE(Discrete Wavelet Transform-Convolutional AutoEncoder)算法.由于现实中被标注好的粒子图像十分缺乏,本文根据已解析出结构的蛋白质,设计生成算法构造了人工图像protein-projection数据集.实验中DWTCAE方法在protein-projection数据集和真实数据集上均取得了良好的效果.最后,本文进行了一系列对比实验,进一步证明了DWT-CAE方法在图像去噪和颗粒边缘确定方面的优势.

Cryo-EM imaging is one of the important ways to obtain the structure of proteins and other biomolecules. It is of great significance for studying the functional properties of proteins and their applications in pharmaceuticals,medicine,and disease prevention and control. In view of the large amount of data and ultra-low signal-to-noise ratio of cryo-electron microscopy images,this paper focuses on the methods of de-noising and particle selection of cryo-electron microscopy images. Combining the convolutional neural network model and the automatic coder model,an EM-CAE( Electron Microscopy-Convolutional AutoEncoder) method for denoising is proposed,and the effectiveness of the algorithm is verified in experiments. For the complex features of the original image noise,this paper further improves the EM-CAE method,combines the automatic coder model with the wavelet transform,and proposes the DWTCAE(Discrete Wavelet Transform-Convolutional AutoEncoder) algorithm. Due to the lack of marked particle images in reality,this paper designs an artificial image protein-projection data set based on the protein that has been solved. The DWT-CAE method has achieved good results in both the protein-projection dataset and the real dataset. Finally,a series of comparative experiments were conducted in this paper,which further proved the advantages of DWT-CAE in image denoising and particle edge determination.