A deep learning-based cryo-em image denoising method and system

Abstract

The application relates to the technical field of image processing, and discloses a cryo-EM image denoising method and system based on deep learning, which comprises the following steps: S1. obtaining an original cryo-EM image data set and performing pretreatment to obtain a training sample set; S2. constructing a cryo-EM image denoising model comprising a noise mapping module and a denoising module; S3. inputting the training sample set into the noise mapping module to perform noise mapping learning and obtaining a noise mapping graph; S4. inputting the noise mapping graph into the denoising module to obtain a denoised cryo-EM image; S5. training the cryo-EM image denoising model according to the denoised cryo-EM image; and S6. denoising the cryo-EM image through the trained cryo-EM image denoising model. The application solves the problems of great denoising training difficulty and a complicated process in the prior art, and has the characteristics of high signal-to-noise ratio and high efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and more particularly, to a cryo-EM image denoising method and system based on deep learning. BACKGROUND

[0002] Cryo-EM is widely used in biological and vaccine research and development, and protein structure analysis is the most important field. The high-resolution three-dimensional structure of proteins is determined by Fourier central section theorem calculation on a large number of two-dimensional cryo-EM images. This process requires selecting particles in the image, determining the direction of the particles according to the conformational state of the selected particle images, then classifying and iterating, and finally used for three-dimensional reconstruction. The success of these steps mainly depends on the signal-to-noise ratio (SNR) of the cryo-EM image. Because long-time electron beam irradiation will cause radiation damage to the original functional structure of the sample, the intensity of the electron beam is controlled during irradiation to avoid this problem. However, too low electron beam intensity will result in very low signal-to-noise ratio of the obtained cryo-EM image, and the typical signal-to-noise ratio is only 0.1. Such low signal-to-noise ratio will reduce the confidence of the structure and cause problems such as loss of particle direction, and will also make it difficult to distinguish protein particles from noise, affecting subsequent operations.

[0003] Previously, traditional methods were mostly used for cryo-EM image denoising, such as low-pass filtering, Wiener filtering, and BM3D based on block matching and filtering, etc. These algorithms usually use Fourier transform to transform the image into the frequency domain, and through designing algorithms in the frequency domain to remove noise or introduce additional image priors to repair missing and degraded image information. Although certain effects have been achieved, for different types of cryo-EM images, the parameter values need to be adjusted accordingly to obtain better denoising effect, and the pre-defined image priors used in traditional methods cannot correctly adapt to the noise model in cryo-EM images, so these methods cannot effectively solve the specific noise characteristics of cryo-EM photos. For most cryo-EM images, there is a problem of insufficient background noise removal and missing particle details.

[0004] Deep learning has strong feature learning and representation ability, and in recent years it has also been used for cryo-EM image denoising and has achieved certain results. Buchholz et al. first applied CNN to cryo-EM image denoising. This method combines the idea of Noise2Noise

[15] and proposes a two-layer U-Net-based Cryo-CARE network. It is trained on pairs of cryo-EM noisy images rather than pairs of noisy images and clean images. Compared with traditional methods, it increases the contrast of the image and improves the interpretability of the collected data. However, its model is relatively simple, and the feature extraction and recovery ability is limited, the background noise removal is insufficient, and the particle morphology, noise and training images have large differences; In 2020, Bepler et al. [further proposed the Topaz denoising model, which is a six-layer neural network based on U-net structure. It is trained on a dataset composed of thousands of micrographs collected under various imaging conditions. The trained general model can denoise new datasets without additional training, increasing the generalization ability of the method and further improving the denoising effect. However, when facing images with more particles and complex noise, it is easy to be too smooth and lose particle details, affecting subsequent three-dimensional reconstruction; In 2022, Li et al. proposed NT2C, which uses InSilicoTEM to simulate clean and noise-free electron microscope images, then uses a generative adversarial network (GAN)

[19] to learn the noise distribution of pure noise blocks in cryo-EM, and designs a contrast-guided noise and signal weighting algorithm to realize the synthesis of real noise images. Finally, the simulated clean and real noise images are used for supervised denoising training, and good results are achieved in background noise removal. However, InSilicoTEM needs to simulate clean electron microscope images corresponding to the three-dimensional structure file of the protein, which greatly limits the exploration of cryo-EM images without reconstructed three-dimensional protein structures, and the image acquisition of pure noise blocks and the synthesis of real noise images are also relatively cumbersome, increasing the difficulty of training.

[0005] Due to the problems of existing image denoising technology that denoising training is difficult and the process is complicated, how to invent an image denoising method with small denoising training difficulty and simple process is a technical problem that needs to be solved in the technical field. SUMMARY

[0006] The present application provides a cryo-EM image denoising method and system based on deep learning, which has the characteristics of high signal-to-noise ratio and high efficiency.

[0007] To achieve the above-mentioned purposes of the present application, the technical solutions adopted are as follows:

[0008] A cryo-EM image denoising method based on deep learning, comprising the following steps:

[0009] S1. obtaining an original cryo-EM image dataset, and preprocessing the original cryo-EM image dataset to obtain a training sample set;

[0010] S2. constructing a cryo-EM image denoising model comprising a noise mapping module and a denoising module;

[0011] S3. inputting the training sample set into the noise mapping module to perform noise mapping learning, and obtaining a noise mapping image;

[0012] S4. inputting the noise mapping image into the denoising module to obtain a denoised cryo-EM image;

[0013] S5. training the cryo-EM image denoising model according to the denoised cryo-EM image;

[0014] S6. denoising a cryo-EM image by using the trained cryo-EM image denoising model.

[0015] Preferably, in the step S1, the original cryo-EM image dataset is preprocessed to obtain the training sample set, specifically, a pair of cryo-EM images with the same superimposed signal but uncorrelated noise in even and odd frames in the original cryo-EM image dataset is synthesized, and data set augmentation is realized by horizontal, vertical flipping and rotation to obtain the training sample set.

[0016] Further, the noise mapping module is a full convolutional neural network for learning the noise mapping of the image; in the step S3, the training sample set is inputted into the noise mapping module to perform noise mapping learning, and a noise mapping image is obtained, specifically, the training sample set is inputted into the noise mapping module, and convolution is performed through four full convolutional layers; the number of feature channels of each convolutional layer is 32, the size of the convolution kernel is 3x3, and padding is set to be equal to 1; after the training sample set is convolved by each layer of the convolution kernel, the noise mapping image is obtained by activating through the ReLU as an activation function.

[0017] Further, the calculation formula for learning the noise mapping by the noise mapping module is:

[0018]

[0019] wherein, represents the noise mapping image, y represents the input noise image, W N represents the parameters of the noise estimation network.

[0020] Further, the denoising module is composed of a three-layer U-Net type encoder-decoder structure; the encoder adds multi-scale connection; the convolution block of each down-sampling layer and up-sampling layer of the encoder-decoder structure is replaced by a wide activation dense residual block; in step S4, the noise map is input into the denoising module to obtain the denoised cryo-EM image, specifically: the noise map is input into the denoising module, after PixelUnShuffle down-sampling, the feature map corresponding to the encoder level is cascaded after Bilinear down-sampling, and the denoised cryo-EM image is obtained.

[0021] Further, the denoising formula used by the denoising module is:

[0022]

[0023] wherein, represents the denoised image, y represents the input noise image, represents the noise level map, W D represents the parameters of the denoising network.

[0024] Further, the wide activation dense residual block includes a dense connection layer, a plurality of wide activation super-resolution layers, and a feature fusion layer with residual learning; the wide activation super-resolution splits the large convolution kernel after the ReLU activation function into two small convolution kernels, so that the residual block obtains wider feature maps before the ReLU function under the same calculation overhead, allowing more feature information to pass through the layer;

[0025] After inputting the noise map into the wide activation dense residual block, the output features are obtained through a plurality of wide activation super-resolution layers connected in sequence, and the local dense features are extracted by adaptively saving the output features through the dense connection layer;

[0026] After extracting the multi-layer local dense features, global feature fusion is performed through the feature fusion layer to adaptively retain the hierarchical features in a global manner, thereby generating implicit deep supervision.

[0027] Further, after inputting the noise map into the wide activation super-resolution layer,

[0028] The 1x1 convolution operation is used to expand the input noise map to a higher dimensional space;

[0029] The ReLU activation function is used for the expanded feature map, and then another 1x1 convolution operation is performed to reduce the depth of the feature map;

[0030] Finally, a 3x3 convolution layer is used for regular convolution operation, and the feature map after convolution is connected with the original input feature map in residual to obtain the output feature.

[0031] Further, in step S5, the cryo-EM image denoising model is trained, specifically:

[0032] The loss function of the training is constructed:

[0033] Loss=αLoss rec +βLoss TV

[0034] Wherein, the total variation loss function Loss TV The denoising quality of the output image is improved by constraining the smoothness of the noise map and keeping the image contour unsmooth. And Indicates the gradient operator along the horizontal and vertical directions of the image, Indicates the noise feature map;

[0035] The reconstruction function Loss rec is used to calculate the pixel-level mean square error between the output image and the input image:

[0036]

[0037] Wherein, represents The output denoised cryo-EM image, x represents the input cryo-EM image;

[0038] The Adagrad algorithm is used to optimize the cryo-EM image denoising model, and the parameters of the cryo-EM image denoising model are constantly adjusted according to the learning rate. When the loss function is minimum, the training of the cryo-EM image denoising model is completed.

[0039] A cryo-EM image denoising system based on deep learning, comprising a data processing module, a model construction module, a model training module, and an image denoising model;

[0040] The data processing module is used to obtain the original cryo-EM image data set and preprocess the original cryo-EM image data set to obtain a training sample set.

[0041] The model construction module is used to construct a cryo-EM image denoising model comprising a noise mapping module and a denoising module.

[0042] The model training module is used to input the training sample set into the noise mapping module for noise mapping learning to obtain a noise map, input the noise map into the denoising module to obtain a denoised cryo-EM image, and train the cryo-EM image denoising model according to the denoised cryo-EM image.

[0043] The image denoising model is used to denoise the cryo-EM image through the trained cryo-EM image denoising model.

[0044] The beneficial effects of the present application are as follows:

[0045] The application discloses a cryo-EM image denoising method based on deep learning. The application aims at the problems that the signal-to-noise ratio of a cryo-EM image is extremely low, and existing denoising methods cannot effectively remove complex noise, and proposes a cryo-EM image denoising model including a noise mapping module and a denoising module, which can effectively remove background noise and better restore particle details, thereby solving the problems of great denoising training difficulty and complicated process in the prior art, and has the characteristics of high signal-to-noise ratio and high efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of the cryo-EM image denoising method based on deep learning;

[0047] Figure 2 A main flowchart of training of the cryo-EM image denoising model of the cryo-EM image denoising method based on deep learning;

[0048] Figure 3 A network architecture diagram of the noise mapping module of the cryo-EM image denoising method based on deep learning;

[0049] Figure 4 A network architecture diagram of the denoising module of the cryo-EM image denoising method based on deep learning;

[0050] Figure 5 A network architecture diagram of the wide activation dense residual block in the denoising module of the cryo-EM image denoising method based on deep learning;

[0051] Figure 6 A network architecture diagram of the wide activation super-resolution layer of the wide activation dense residual block in the denoising module of the cryo-EM image denoising method based on deep learning;

[0052] Figure 7 A comparison diagram of denoising results of EMPIAR-10028 of the cryo-EM image denoising method based on deep learning.

[0053] Figure 8 A comparison diagram of denoising results of EMPIAR-10025 of the cryo-EM image denoising method based on deep learning. DETAILED DESCRIPTION

[0054] The application will be described in detail below in combination with the drawings and specific embodiments.

[0055] Embodiment 1

[0056] As shown in Figure 1 A deep learning-based cryo-EM image denoising method includes the following steps:

[0057] S1. Obtain an original cryo-EM image dataset, and preprocess the original cryo-EM image dataset to obtain a training sample set;

[0058] S2. Construct a cryo-EM image denoising model including a noise mapping module and a denoising module;

[0059] S3. Input the training sample set into the noise mapping module to perform noise mapping learning, and obtain a noise mapping graph;

[0060] S4. Input the noise mapping graph into the denoising module to obtain a denoised cryo-EM image;

[0061] S5. Train the cryo-EM image denoising model according to the denoised cryo-EM image;

[0062] S6. Denoise the cryo-EM image through the trained cryo-EM image denoising model.

[0063] Embodiment 2

[0064] As shown in Figure 2 A deep learning-based cryo-EM image denoising method includes the following steps:

[0065] S1. Obtain an original cryo-EM image dataset, and preprocess the original cryo-EM image dataset to obtain a training sample set;

[0066] S2. Construct a cryo-EM image denoising model including a noise mapping module and a denoising module;

[0067] S3. Input the training sample set into the noise mapping module to perform noise mapping learning, and obtain a noise mapping graph;

[0068] S4. Input the noise mapping graph into the denoising module to obtain a denoised cryo-EM image;

[0069] In this embodiment, the denoising module is built based on U-Net, and a multi-scale cascade structure and a wide activation dense residual block are added to improve the expression ability and feature extraction and recovery ability of the denoising module;

[0070] S5. Train the cryo-EM image denoising model according to the denoised cryo-EM image;

[0071] In this embodiment, during training, the total variation loss function is used to constrain the smoothness of the noise map while preserving the particle characteristics, to improve the denoising quality of the output image, and to calculate the image denoising model loss function according to the corresponding even frozen electron microscope image of the denoised image and the original input image; the Adagrad algorithm is used to optimize the frozen electron microscope image denoising model, and the parameters of the frozen electron microscope image denoising model are constantly adjusted according to the learning rate, and when the loss function is minimized, the training of the frozen electron microscope image denoising model is completed.

[0072] S6. Denoising the frozen electron microscope image by the trained frozen electron microscope image denoising model.

[0073] In one specific embodiment, in step S1, the original frozen electron microscope image dataset is preprocessed to obtain a training sample set, specifically: the even and odd frame superposition signal same but noise unrelated frozen electron microscope image pairs in the original frozen electron microscope image dataset are synthesized, and the data set is augmented by horizontal, vertical flipping and rotation to obtain the training sample set.

[0074] In this embodiment, the original frozen electron microscope image dataset is the original data obtained from EMPIAR; EMPIAR contains thousands of datasets taken by a frozen electron microscope, each dataset of frozen electron microscope images belongs to the same kind of protein particles, and the frozen electron microscope images provided are stored in the form of mrc or tiff files; in this embodiment, 200 mrc files are selected from them, and the dataset is further preprocessed;

[0075] In this embodiment, the dataset is further preprocessed, specifically: in the selected EMPIAR-10025 dataset, each mrc file contains 18 frames of frozen electron microscope images, a python script is needed to separately superimpose the even and odd frames in the mrc to synthesize two frozen electron microscope images with the same signal but unrelated noise and save them in different mrc files, and the dataset is augmented by horizontal, vertical flipping and rotation. To avoid interpolation artifacts, the image is only rotated by 90°, 180° and 270°, and finally 800 mrc files are obtained, thereby obtaining the final training required dataset image pair.

[0076] In one specific embodiment, as Figure 3As shown, the noise mapping module is a full convolutional neural network for learning the noise mapping of the image; in step S3, the training sample set is input into the noise mapping module to learn the noise mapping and obtain the noise mapping graph, specifically: the training sample set is input into the noise mapping module and convolved through four full convolutional layers; the number of feature channels of each convolutional layer is 32, the size of the convolution kernel is 3*3, and padding is set to 1; after the training sample set is convolved through each layer of the convolution kernel, the noise mapping graph is obtained through the activation of the ReLU activation function.

[0077] In one specific embodiment, the calculation formula for learning the noise mapping by the noise mapping module is:

[0078]

[0079] wherein, represents the noise mapping graph, y represents the input noise image, W N represents the parameters of the noise estimation network.

[0080] In one specific embodiment, as shown in Figure 4 the denoising module is composed of a three-layer U-Net type encoder-decoder structure; the encoder adds a multi-scale connection; the convolutional block of each down-sampling layer and up-sampling layer of the encoder-decoder structure is replaced by a wide activation dense residual block; in step S4, the noise mapping graph is input into the denoising module to obtain the denoised cryo-EM image, specifically: the noise mapping graph is input into the denoising module, and after down-sampling by PixelUnShuffle, it is concatenated with the feature map corresponding to the encoder level down-sampled by Bilinear, to obtain the denoised cryo-EM image.

[0081] In one specific embodiment, the denoising formula adopted by the denoising module is:

[0082]

[0083] wherein, represents the denoised image, y represents the input noise image, represents the noise level mapping graph, W D represents the parameters of the denoising network.

[0084] In one specific embodiment, as shown in Figure 5 the wide activation dense residual block includes a dense connection layer, a plurality of wide activation super-resolution layers, and a feature fusion layer with residual learning; the wide activation super-resolution splits the large convolution kernel after the ReLU activation function into two small convolution kernels, so that the residual block obtains wider feature maps before the ReLU function under the same calculation overhead, allowing more feature information to pass through the graph layer.

[0085] After inputting the noise map into the wide-activation dense residual block, the output features are obtained through a plurality of wide-activation super-resolution layers connected in sequence, and the local dense features are extracted by adaptively saving the output features through the dense connection layer;

[0086] After extracting the multi-layer local dense features, the global feature fusion is performed through the feature fusion layer to adaptively retain the hierarchical features in a global manner, thereby generating an implicit deep supervision.

[0087] In one specific embodiment, as shown in Figure 6 After inputting the noise map into the wide-activation super-resolution layer,

[0088] The input noise map is expanded to a higher dimensional space by using a 1x1 convolution operation;

[0089] The expanded feature map is activated by using a ReLU activation function, and then another 1x1 convolution operation is performed to reduce the depth of the feature map;

[0090] Finally, a 3x3 convolution layer is used for regular convolution operation, and the feature map after convolution is connected with the original input feature map in a residual manner to obtain the output feature.

[0091] In one specific embodiment, in step S5, the cryo-EM image denoising model is trained, specifically:

[0092] The loss function for training is constructed:

[0093] Loss = aLoss rec + bLoss TV

[0094] wherein the total variation loss function Loss TV The denoising quality of the output image is improved by constraining the smoothness of the noise map to maintain the non-smoothness of the image outline; and denote the gradient operators along the horizontal and vertical directions of the image, denote the noise feature map;

[0095] The reconstruction function Loss rec is used to calculate the pixel-level mean square error between the output image and the input image:

[0096]

[0097] wherein represents The output denoised cryo-EM image, x represents the input cryo-EM image;

[0098] Adagrad algorithm is used to optimize the cryo-EM image denoising model, the parameters of the cryo-EM image denoising model are continuously adjusted according to the learning rate, and when the loss function is minimum, the training of the cryo-EM image denoising model is completed.

[0099] The present application aims at the problem that the signal-to-noise ratio of a cryo-EM image is extremely low, and existing denoising methods cannot effectively remove complex noise, and proposes a cryo-EM image denoising method based on deep learning. Figure 7 and Figure 8 As shown in the formula, EMPIAR-10028 and EMPIAR-10028 are denoised, it can be found that the method can effectively remove background noise while better restoring particle details, the signal-to-noise ratio (SNR) is optimal, and the number of positive particles is also improved.

[0100] The application discloses a cryo-EM image denoising method based on deep learning.

[0101] Embodiment 3

[0102] A cryo-EM image denoising system based on deep learning includes a data processing module, a model construction module, a model training module, and an image denoising model.

[0103] The data processing module is used to obtain an original cryo-EM image data set and pre-process the original cryo-EM image data set to obtain a training sample set.

[0104] The model construction module is used to construct a cryo-EM image denoising model including a noise mapping module and a denoising module.

[0105] The model training module is used to input the training sample set into the noise mapping module for noise mapping learning to obtain a noise mapping graph, input the noise mapping graph into the denoising module to obtain a denoised cryo-EM image, and train the cryo-EM image denoising model according to the denoised cryo-EM image.

[0106] The image denoising model is used to denoise the cryo-EM image through the trained cryo-EM image denoising model.

[0107] Obviously, the above-mentioned embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. Any modification, equivalent replacement and improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A deep learning-based cryo-EM image denoising method, characterized by: The method comprises the following steps: S1. Obtain an original cryo-EM image dataset, and pre-process the original cryo-EM image dataset to obtain a training sample set; S2. Construct a cryo-EM image denoising model comprising a noise mapping module and a denoising module; the denoising module is composed of a three-layer U-Net type encoder-decoder structure; the encoder is provided with multi-scale connection; and the convolution block of each down-sampling layer and up-sampling layer of the encoder-decoder structure is replaced by a wide activation dense residual block; S3. Input the training sample set into the noise mapping module to perform noise mapping learning, and obtain a noise mapping graph; S4. Input the noise mapping graph into the denoising module to obtain a denoised cryo-EM image, specifically: input the noise mapping graph into the denoising module, perform PixelUnShuffle down-sampling, and then concatenate the feature map corresponding to the encoder level after Bilinear down-sampling to obtain the denoised cryo-EM image; S5. Train the cryo-EM image denoising model according to the denoised cryo-EM image; S6. Denoise the cryo-EM image through the trained cryo-EM image denoising model.

2. The deep learning based cryo-EM image denoising method of claim 1, wherein: In the step S1, the original cryo-EM image dataset is pre-processed to obtain the training sample set, specifically: the even-numbered frames and odd-numbered frames in the original cryo-EM image dataset are stacked to synthesize cryo-EM image pairs with the same signal but uncorrelated noise, and data set augmentation is realized through horizontal, vertical flipping and rotation to obtain the training sample set. ​ 3. The deep learning based cryo-EM image denoising method of claim 1, wherein: The noise mapping module is a full convolutional neural network, which is used for learning the noise mapping of the image; in the step S3, the training sample set is input into the noise mapping module to perform noise mapping learning, and a noise mapping graph is obtained, specifically: the training sample set is input into the noise mapping module, and convolution is performed through four full convolutional layers; the number of feature channels of each convolutional layer is 32, the size of the convolution kernel is 3*3, and padding is set to 1; after the training sample set is convolved through each convolution kernel, the noise mapping graph is obtained through the activation of the ReLU activation function.

4. The deep learning based cryo-EM image denoising method of claim 3, wherein: The calculation formula for learning the noise mapping by the noise mapping module is: wherein, representative of a noise map, representative of an input noise image, representative of parameters of a noise estimation network.

5. The deep learning based cryo-EM image denoising method of claim 1, wherein: The denoising formula used by the denoising module is: wherein, denotes a denoised image, denotes an input noisy image, denotes a noise level map, denotes parameters of the denoising network.

6. The deep learning based cryo-EM image denoising method of claim 5, wherein: The wide activation dense residual block comprises a dense connection layer, a plurality of wide activation super-resolution layers and a feature fusion layer with residual learning; the wide activation super-resolution splits a large convolution kernel after the ReLU activation function into two small convolution kernels, so that the residual block obtains wider feature maps before the ReLU function under the same calculation overhead, and allows more feature information to pass through the graph layer; ​ After the noise mapping graph is input into the wide activation dense residual block, the output features are obtained through a plurality of wide activation super-resolution layers connected in sequence, and the local dense features are extracted by adaptively saving the output features through the dense connection layer; After the multi-layer local dense features are extracted, the global feature fusion is performed through the feature fusion layer to adaptively retain the hierarchical features in a global manner, thereby generating implicit deep supervision.

7. The deep learning-based cryo-EM image denoising method according to claim 6, characterized in that: After inputting the noise map into the wide-activated super-resolution layer, The input noise map is expanded to a higher dimensional space using a 1x1 convolution operation; A ReLU activation function is applied to the expanded feature map, followed by another 1x1 convolution operation to reduce the depth of the feature map; Finally, a 3x3 convolution layer is used for regular convolution operation, and the feature map after convolution is connected with the original input feature map to obtain the output feature.

8. The deep learning based cryo-EM image denoising method of claim 7, wherein: In step S5, the cryo-EM image denoising model is trained, specifically: The loss function for training is constructed: wherein the total variation loss function The denoising quality of the output image is improved by restraining the smoothness of the noise map to keep the non-smoothness of the image contour. reconstruction function to compute the pixel-wise mean squared error between the output image and the input image: wherein represents outputting the de-noised cryo-EM image, represents an input cryo-EM image; The cryo-EM image denoising model is optimized using the Adagrad algorithm, and the parameters of the model are constantly adjusted according to the learning rate. When the loss function is minimized, the training of the cryo-EM image denoising model is completed. 9.A deep learning based cryo-EM image denoising system, characterized in that: For implementing the method of any one of claims 1-8, comprising a data processing module, a model construction module, a model training module, and an image denoising model. The data processing module is used to obtain the original cryo-EM image dataset and preprocess the original cryo-EM image dataset to obtain the training sample set. The model construction module is used to construct a cryo-EM image denoising model comprising a noise mapping module and a denoising module. The model training module is used to input the training sample set into the noise mapping module for noise mapping learning to obtain a noise map; input the noise map into the denoising module to obtain a denoised cryo-EM image; and train the cryo-EM image denoising model according to the denoised cryo-EM image. The image denoising model is used to denoise the cryo-EM image through the trained cryo-EM image denoising model.

Images