Cryo-em Image Denoising Based On Generative Adversarial Network

Cryo-EM(cryo-electron microscopy)is a technology that can build the three-dimensional reconstruction of bio-macromolecule.It can obtain high-resolution biological macromolecular structures through cryo-electron microscopy images and three-dimensional reconstruction software.As a result of the particularity of cryo-em imaging,the collected cryo-em images contain a lot of noise,which has a great impact on the accuracy of 3D reconstruction.The main method of current denoising of the cryo-em images is to increase the signal-to-noise ratio of single-particle images by picking a large number of particles and summing them in Fourier space.However,the denoising effect in this way cannot effectively remove noise,and the low signal-to-noise ratio and low contrast of the rawcryo-em image make it difficult to picksingle particles.Therefore,it is necessary to find an effective denoising algorithm for the original cryo-EM images.Based on the characteristics of generative adversarial networks(GANs),this paper regards generative adversarial networks as the process of image encoding and decoding.By combining gray-scale constraints and wasserstein distance,the features of the image are effectively extracted for image restoration,and the problem that the loss can not guide the training during training is avoided,and the cryo-EM images are greatly denoised.The details are as follows:1.An improved discriminant model base on wasserstein distance was proposed.In order to solve the problem of unstable training and loss of the GANs and unable to effectively guide the model training,it is proposed to use the wasserstein distance to improve the discriminator from idea on WGAN.It solves the problems of traditional GANs model training difficulty,the grnerated image is unstable,and the training gradient cannot be transfereed.The model parameters are improved by the training of Cryo-em image simulation data,and the GANs can be effectively iterated by a stable loss gradient.The experiment results show that the model proposed in this paper can increase the PSNR(Peak Signal to Noise Ratio)of Cryo-em simulation image by 6.7d B,and it can better save the model structure and veins signal in the original image.2.An improved generative model by add gray constraint was proposed.In order to reduce the overall gray difference between the result image with the target image.The model can compare the denoised result image with the target image,compare grayscale the result image and the real image,and adjust the parameter for improve generated model.To reduce the gray difference between the denoising result and the target result.In addition,it can improving remove background of biomacromolecule.The orignal image's contour signal can be retained,and the information of biological macromolecules is separated from the noise information,which increases the contrast of the image after denoising and improves the image quality.Noise results are similar in structure to target results.The experimental results show that the improved model can improve SSIM(Structural Similarity Index)of the denoised image results by 0.143.An improved generative model by using manifold learning was proposed.In order to solve the problem that hidden structural relationship between the data after the model encoded be damaged,improve the encoding structure of the generative model by manifold learning,so that the encoding result is always kept on a manifold plane.This effectively relieves the problem that the generated model results in unstable encoding results caused by different encodings during the encoding process of the generated model,and further improves the quality of the denoised image.