Method and system for quality restoration of cryo-electron tomogram images

Abstract

The application discloses a quality recovery method and system for a frozen electron tomographic reconstruction image; wherein the method comprises the following steps: acquiring a two-dimensional projection sequence of a frozen electron tomography; performing three-dimensional reconstruction on the two-dimensional projection sequence to obtain a three-dimensional reconstruction image; processing the three-dimensional reconstruction image by using a trained noise generation model to obtain noise; superimposing the noise on the three-dimensional reconstruction image to obtain a noise superimposed image; and inputting the noise superimposed image into a trained image quality recovery model to output an image quality recovered image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cryo-electron tomography, and particularly relates to a quality restoration method and system for cryo-electron tomography reconstruction images. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Cryo-electron tomography (Cryo-ET) imaging technology is a powerful imaging tool for reconstructing and visualizing sub-microscopic structures in biological cells in a natural state. However, due to the limitations of imaging instruments and biological samples, the projection image sequence of Cryo-ET and its three-dimensional reconstruction image often exhibit extremely low signal-to-noise ratio (SNR) and complex noise, which will affect the downstream analysis tasks in the Cryo-ET processing flow. In particular, due to the changes in the parameter configuration of cryo-electron microscopy for different biological samples, it is difficult to obtain a large amount of image data of the same biological sample during the imaging process. Therefore, it is an important research problem in this field to realize the quality restoration of Cryo-ET reconstruction under the condition of limited data amount.

[0004] For the quality restoration of Cryo-ET reconstruction images, researchers have proposed a large number of traditional (non-learning) algorithms based on statistical and iterative strategies. The first idea of this kind of method is based on non-local mean: using the neighborhood pixels around each pixel, the pixel value is estimated by weighted average, such as the non-local mean algorithm. The second idea is based on the grouping and collaborative filtering technology of image blocks according to the local similarity measure, so as to realize the fusion of non-local information, such as the BM4D algorithm. The establishment conditions of traditional methods include the following two points: 1. the noise mean tends to zero; 2. enough image blocks can be obtained, but these two conditions are not always true in the real data environment, thus causing great limitations. In addition, the calculation amount of traditional methods is often large, and the time complexity of the algorithm is high, which is difficult to integrate into the main flow of Cryo-ET.

[0005] With the development of deep neural network technology, supervised deep learning quality restoration algorithms have achieved good results in natural images. Supervised learning quality restoration algorithms usually require a large amount of (low-quality image, clean image) training data pair, so compared with traditional methods, such methods have more reliable image structure restoration performance. However, the supervised learning quality restoration algorithm also has a very obvious problem: first, the supervised learning algorithm has high requirements for training data, and needs to explicitly introduce the information provided by the clean image. However, in a real environment, it is difficult to obtain such training data; second, since the method depends largely on the distribution of the training data, in many cases there is a problem of insufficient generalization. If the distance between the test data and the training data distribution is too large, the performance of the model will be greatly reduced.

[0006] In order to solve the problem caused by the limitation of supervised learning, researchers have proposed self-supervised learning quality restoration algorithms. Self-supervised learning quality restoration algorithms do not need to explicitly use clean images as labels for training, based on the assumptions introduced by image modeling and statistics, directly learn and predict the structural information and contrast information of images from low-quality noisy images, such as Noise2Noise, Noise2Void algorithms, and self-supervised learning algorithms support single-sample training of fluorescence microscope images or Cryo-ET images. However, the performance of such algorithms under single-sample training is still limited. The premise for self-supervised learning algorithms to approach the performance of supervised learning algorithms is to have a large amount of low-quality image data. SUMMARY

[0007] In order to solve the problems of the prior art, the present application provides a quality restoration method and system for cryo-electron tomography images. In the case of low signal-to-noise ratio, low data volume and inability to obtain Ground Truth as training data for Cryo-ET images, a noise modeling and self-supervised learning framework based on contrast guide term and first-order sparse guide term are used to realize the quality restoration of images, to solve the low robustness problem of traditional methods and single-sample self-supervised learning for complex noise in cryogenic electron tomography Cryo-ET, and the overfitting problem caused by the dependence of supervised learning methods on training data.

[0008] In a first aspect, the present application provides a quality restoration method for cryo-electron tomography images;

[0009] The quality restoration method for cryo-electron tomography images comprises:

[0010] Obtaining a two-dimensional projection sequence of cryo-electron tomography;

[0011] Three-dimensional reconstruction is performed on the two-dimensional projection sequence to obtain a three-dimensional reconstructed image;

[0012] noise to the three-dimensional reconstruction image to obtain a noise superimposed image;

[0013] input the noise superimposed image into the trained image quality recovery model to output an image quality recovered image.

[0014] In a second aspect, the present application provides a quality recovery system for a frozen electron tomography reconstruction image.

[0015] The quality recovery system for the frozen electron tomography reconstruction image comprises:

[0016] an acquisition module configured to acquire a two-dimensional projection sequence of a frozen electron tomography;

[0017] a three-dimensional reconstruction module configured to perform three-dimensional reconstruction on the two-dimensional projection sequence to obtain a three-dimensional reconstruction image;

[0018] a noise superimposing module configured to obtain noise by processing the three-dimensional reconstruction image using a trained noise generation model, and superimpose the noise to the three-dimensional reconstruction image to obtain a noise superimposed image;

[0019] a quality recovery module configured to input the noise superimposed image into the trained image quality recovery model to output an image quality recovered image.

[0020] In a third aspect, the present application further provides an electronic device comprising:

[0021] a memory for non-transiently storing computer readable instructions; and

[0022] a processor for running the computer readable instructions,

[0023] wherein the computer readable instructions are run by the processor to perform the method of the first aspect.

[0024] In a fourth aspect, the present application further provides a storage medium for non-transiently storing computer readable instructions, wherein the non-transient computer readable instructions are executed by a computer to perform the instructions of the method of the first aspect.

[0025] In a fifth aspect, the present application further provides a computer program product comprising a computer program for implementing the method of the first aspect when run on one or more processors.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] 1. A new Cryo-ET reconstructed image quality recovery self-supervised learning network based on noise modeling and sparsity constraint is proposed. Noise modeling of Cryo-ET images is achieved through a generative model; a guided image is generated using a Gaussian low-pass filter, and sparsity in a highly smooth image is defined using the first derivative of the image, and both are introduced into a new loss function to achieve sparsity constraint, ensuring the quality recovery effect of the image in the case of single-sample training data.

[0028] 2. A generative model is used to learn the noise distribution in Cryo-ET reconstructed images, solving the problem of defining the noise distribution in Cryo-ET with an analytical method, so that the model has prior knowledge of noise, which helps to recover the structure information of the image;

[0029] 3. The training strategy of self-supervised learning solves the problem of difficult to obtain Ground Truth in Cryo-ET image data, and reduces the stringent requirements for Ground Truth in training data;

[0030] 4. A new constraint condition is used to define the structure information and smoothness in the image from the perspectives of contrast and sparsity, to achieve a single-sample training quality recovery network and obtain ideal denoising results. BRIEF DESCRIPTION OF DRAWINGS

[0031] The drawings constituting a part of the specification of the present application are used to provide a further understanding of the present application, the illustrative embodiments of the present application and the description thereof serve to explain the present application, and do not constitute an improper limitation on the present application.

[0032] Figure 1 The model overall architecture and data flow of example one;

[0033] Figure 2 The basic structure of the image noise extraction and generator of example one. DETAILED DESCRIPTION

[0034] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0035] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0036] The embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict.

[0037] All data acquisition of the present embodiment is based on the compliance with laws and regulations and user consent, and the legal application of data.

[0038] In the process of cryo-electron tomography, in order to ensure the original state of the biological sample as much as possible, the electron dose in the imaging device and the ice layer thickness of the sample have strict requirements, so that the signal-to-noise ratio of the two-dimensional projection taken and the three-dimensional reconstruction image generated based on the projection image is very low. In addition, the influence of electronic noise and reconstruction artifacts, the noise type in the image is very complex.

[0039] Embodiment one

[0040] The present embodiment provides a quality recovery method for cryo-electron tomography reconstruction images;

[0041] The quality recovery method for cryo-electron tomography reconstruction images comprises:

[0042] S101: acquiring a two-dimensional projection sequence of cryo-electron tomography;

[0043] S102: three-dimensional reconstruction is performed on the two-dimensional projection sequence to obtain a three-dimensional reconstruction image;

[0044] S103: using the trained noise generation model to process the three-dimensional reconstruction image to obtain noise; superimposing the noise into the three-dimensional reconstruction image to obtain a noise superimposed image;

[0045] S104: inputting the noise superimposed image into the trained image quality recovery model to output an image quality recovered image.

[0046] Further, the network structure of the trained noise generation model comprises:

[0047] The generator and the discriminator are connected to each other.

[0048] Further, the trained noise generation model, the training process of which comprises:

[0049] A first training set and a first test set are constructed; wherein the first training set and the first test set are three-dimensional reconstruction images of known real noise blocks; the real noise block is obtained by extracting a plurality of noise slices from a three-dimensional reconstruction image and then cutting each noise slice;

[0050] The random noise block is input into the generator to generate a fake noise block;

[0051] The real noise block and the fake noise block are input into the discriminator to train the discriminator; when the loss function value stops decreasing, the training is stopped, and a trained noise generation model is obtained.

[0052] Further, the loss function value is expressed as:

[0053]

[0054] is the data distribution of the generator, is the noise distribution, is the distribution obtained by linear sampling between data points in and .

[0055] It should be understood that the loss function of the noise generator: the goal of the noise generator is to train the generator to generate noise blocks, and to train the discriminator to determine whether the generated image is fake noise. As a result of adversarial learning, the noise blocks generated by the generator will not be considered as fake noise by the discriminator.

[0056] Noise generator network training parameters: the network is initialized by a Gaussian distribution with a strength of 0.002, an Adam optimizer with parameters β1=0.5 and β2=0.999 is used, the total iteration of the training data set is 10000 cycles, and the network learning rate is 0.0002. During the training process, the image size is 256x256x256.

[0057] Further, the network structure of the trained image quality restoration model comprises:

[0058] The up-sampling layer, the convolution block a1, the max-pooling layer c1, the convolution block a2, the max-pooling layer c2, the convolution block a3, the max-pooling layer c3, the convolution block a4, the max-pooling layer c4, the convolution block a5, the max-pooling layer c5, the convolution block a6, the interpolation layer z1, the convolution block b1, the convolution block b2, the interpolation layer z2, the convolution block b3, the convolution block b4, the interpolation layer z3, the convolution block b5, the convolution block b6, the interpolation layer z4, the convolution block b7, the convolution block b8, the interpolation layer z5, the convolution block b9, the three-dimensional convolution layer w1, the down-sampling layer and the output layer are sequentially connected;

[0059] The output end of the up-sampling layer is connected with the input end of the interpolation layer z5; the output end of the max-pooling layer c1 is connected with the input end of the interpolation layer z4; the output end of the max-pooling layer c2 is connected with the input end of the interpolation layer z3; the output end of the max-pooling layer c3 is connected with the input end of the interpolation layer z2; and the output end of the max-pooling layer c4 is connected with the input end of the interpolation layer z1.

[0060] The internal structures of the convolution block a1, the convolution block a2, the convolution block a3, the convolution block a4, the convolution block a5 and the convolution block a6 are the same, and each includes a three-dimensional convolution layer, a batch normalization layer and a rectifier unit connected with each other.

[0061] The internal structures of the convolution block b1, the convolution block b2, the convolution block b3, the convolution block b4, the convolution block b5, the convolution block b6, the convolution block b7, the convolution block b8 and the convolution block b9 are the same, and each includes a three-dimensional convolution layer, a batch normalization layer and a rectifier unit connected with each other.

[0062] Further, the trained image quality restoration model, the training process of which comprises:

[0063] A second training set and a second test set are constructed, wherein the second training set and the second test set are both original two-dimensional projection sequences of frozen electron tomography with known image quality restoration results;

[0064] The original two-dimensional projection sequence of the second training set is three-dimensionally reconstructed to obtain a three-dimensional reconstructed image, and the three-dimensional reconstructed image is filtered to obtain a filtered reconstructed image;

[0065] The original two-dimensional projection sequence of the second training set is denoised to obtain a smoothed projection sequence, the smoothed projection sequence is three-dimensionally reconstructed to obtain a smoothed reconstructed image, and the smoothed reconstructed image is sparsely extracted to obtain local sparse information;

[0066] The three-dimensional reconstructed image is processed by using the trained noise generator to obtain noise, and the noise is added to the three-dimensional reconstructed image to obtain an input image with superimposed noise;

[0067] The input image with superimposed noise is input into the image quality recovery model to train the model, and the training is stopped when the total loss function value of the model no longer decreases, and a trained image quality recovery model is obtained.

[0068] The total loss function is the sum of the first loss function, the second loss function and the third loss function.

[0069] The first loss function is the loss value of the model output image and the three-dimensional reconstructed image.

[0070] The second loss function is the loss value of the model output image and the local sparse information.

[0071] The third loss function is the loss value of the model output image and the contrast information of the filtered reconstructed image.

[0072] Further, the formula of the total loss function is:

[0073]

[0074] Further, the formula of the first loss function is:

[0075]

[0076] wherein V po (x) is the corresponding patch of the smooth reconstructed image, V ps (x) is the network input reconstructed image with superimposed noise, represents the gradient-guided sparsity constraint function, represents the mathematical expectation; D(f θ (V ps (x))) represents the gradient information of the network output image patch; D(V po (x)) represents the gradient information of the smooth reconstructed image patch;

[0077] The mathematical expression of the first derivative is as follows:

[0078]

[0079] G x is a three-dimensional derivative convolution kernel defined on the x-axis, G y is a three-dimensional derivative convolution kernel defined on the y-axis, is a convolution operator, V o (x) represents the smooth reconstructed image.

[0080] Gradient guided sparsity constraint function: according to the definition of the first derivative, the present application can derive a property that "the first derivative image corresponding to a sufficiently sparse image is also sufficiently sparse, and the sparsity of the image can be described". According to this property, the first derivative of the output image can be approximated to the first derivative of the highly smooth reconstructed image, and the sparsity of the image is ensured, so as to remove the noise.

[0081] Further, the second loss function is expressed as:

[0082]

[0083] Wherein, V pf (x) is the corresponding patch of the filtered reconstructed image, V ps (x) is the network input reconstructed image superimposed with noise, represents the filtered guide item loss function, represents mathematical expectation; f θ (V ps (x)) represents the network output image;

[0084] Wherein,

[0085] is denoted as a three-dimensional Gaussian filter function with scale s and intensity σ, and V(x) represents the original reconstructed image; V f (x) represents the filtered reconstructed image.

[0086] The image obtained by the Gaussian low-pass filter with strong parameters has sufficient smoothness and faithful contrast, and can well constrain the contrast of the denoised image, so it can be used as a guide item of the network loss function to guide the training.

[0087] Further, the third loss function is expressed as:

[0088]

[0089] Wherein, f θ is a quality recovery network; V ps (x) is the network input reconstructed image superimposed with noise; V p (x) is the image patch of the original reconstructed image; represents the self-supervised structure reconstruction loss function, f θ (V ps (x)) represents the network output image patch; represents mathematical expectation.

[0090] Quality recovery network training parameters: the application uses the Adam optimizer with parameters β1=0.5 and β2=0.999, the total iteration of the training data set is 40 cycles, and the network learning rate is 0.001. During the training process, the image patch size is 128*128*128. The parameters of the Gaussian filter function are scale s=11 and intensity σ=2.6. For the training loss function, the application takes the weighted weights λ1=1.0, λ2=0.05 and λ3=0.5.

[0091] The application proposes a solution of a self-supervised learning model, which realizes the recovery of a high-quality reconstructed image of a single sample by combining noise modeling, contrast guidance and sparsity constraint, and the process is as follows:

[0092] The two-dimensional electron tomography original projection sequence is taken as the system input.

[0093] The original projection sequence is operated as follows: a. directly reconstructing three-dimensionally to obtain an original three-dimensional reconstruction; b. performing two-dimensional denoising on the projection sequence to obtain an over-smoothed projection sequence.

[0094] The original three-dimensional reconstruction is operated as follows: c. extracting a noise region outside the sample as training data for a noise generator; d. using a three-dimensional low-pass filter to generate a filtered reconstruction image containing high-contrast information.

[0095] The extracted noise is used to train a stable noise generator.

[0096] The obtained noise generator is used to generate a large amount of noise data, and the noise data is superimposed on the original three-dimensional reconstruction image to obtain a training input image with superimposed noise.

[0097] The over-smoothed projection sequence is reconstructed into a smooth reconstruction image.

[0098] The original three-dimensional reconstruction image is taken as the supervision information, the obtained filtered reconstruction image and the obtained smooth reconstruction image are taken as the sparsity constraint and the contrast constraint, and the quality recovery network is trained in a self-supervised learning manner.

[0099] The obtained quality recovery network is used to calculate the original three-dimensional reconstruction image as the network input, so as to finally obtain a denoised image.

[0100] Figure 1 The network architecture and overall data flow designed by the application are shown. The input of the network is the original projection sequence, and the following processing is performed:

[0101] 1) directly using the original sequence to reconstruct to generate an original noisy reconstruction;

[0102] 2) The noise sequence is subjected to noise suppression by a two-dimensional denoising model to obtain a smoothed projection sequence, and reconstruction is performed to generate a smoothed reconstruction.

[0103] The original reconstruction image is used to model the noise by a noise generator, and the learned noise data is linearly superimposed on the original reconstruction to obtain a reconstruction image with superimposed noise, which is used as the input of the quality recovery network.

[0104] Local sparse information extracted from the smoothed reconstruction and contrast-guided information in the filtered reconstruction image are used to guide the training of the quality recovery network. Using the model obtained after training, reliable denoised images can be obtained in a single-sample training mode.

[0105] Figure 2 The structure of the noise extraction and generation module is shown. In the noise extraction and generation module, the size of the noise block is s x s.

[0106] The noise generator network adopts the basic architecture of two-dimensional WGAN-GP.

[0107] The generator consists of 5 basic units, each containing a 2D deconvolution layer, a batch normalization layer and a linear rectifier unit, and the output layer is a 2D deconvolution layer with a step size of 2.

[0108] The discriminator consists of 5 basic units, each containing a 2D convolution layer, a batch normalization layer and an L-linear rectifier unit (slope k = 0.2), and the output layer is a 2D convolution layer with a step size of 1.

[0109] The method of generating noise is as follows: using the trained noise generator, n x s 2D noise image blocks with a size of s x s are generated each time, and the s noise blocks are stacked into n three-dimensional image blocks with a size of s x s x s, which are used to generate training images with superimposed noise.

[0110] The image quality recovery model is an encoder-decoder network based on the UNet architecture with a depth of 5. The network introduces a 2-fold nearest neighbor interpolation layer at the front end to enhance the structure of the data, which increases the receptive field and further preserves the structural information.

[0111] Each basic convolution module of the network includes a 3D convolution layer with a size of 3 x 3 x 3 and a step size of 1, a 3D batch normalization layer, and an L-linear rectifier unit with a slope k = 0.1. The size of the max pooling layer is 2 x 2 x 2. The interpolation layer is also a nearest neighbor interpolation with a multiple of the image size output by the encoder convolution layer at the same layer.

[0112] In order to ensure the size consistency of the final output image, a 0.5-fold downsampling unit is connected to the last layer to restore the image to the original size.

[0113] The input image of the superimposed noise is generated in the following way:

[0114] V s (x) = V(x) + aN s (x)

[0115] V(x) is the original reconstructed image, N s (x) is the generated noise, V s (x) is the network input image of the superimposed noise, a is the superimposition coefficient, which determines the amount of superimposed noise.

[0116] The sparse information of the smoothed reconstructed image is extracted to obtain local sparse information, and the operator is realized by using the first derivative definition. The calculation formula is as follows:

[0117]

[0118] Where G x is the first derivative operator in the x-axis direction, and G y is the first derivative operator in the y-axis direction.

[0119] Embodiment two

[0120] The embodiment provides a quality recovery system for a frozen electron tomography reconstructed image;

[0121] The quality recovery system for the frozen electron tomography reconstructed image comprises:

[0122] An acquisition module configured to acquire a two-dimensional projection sequence of frozen electron tomography;

[0123] A three-dimensional reconstruction module configured to perform three-dimensional reconstruction on the two-dimensional projection sequence to obtain a three-dimensional reconstructed image;

[0124] A noise superposition module configured to process the three-dimensional reconstructed image by using a trained noise generation model to obtain noise, and superimpose the noise on the three-dimensional reconstructed image to obtain a noise superimposed image;

[0125] A quality recovery module configured to input the noise superimposed image into a trained image quality recovery model to output an image quality recovered image.

[0126] It should be noted that the acquisition module, the three-dimensional reconstruction module, the noise superposition module and the quality recovery module correspond to steps S101 to S104 in Embodiment One, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment One. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0127] The description of each of the above embodiments focuses on different aspects. The parts not described in detail in one embodiment can be found in the relevant description of the other embodiments.

[0128] The proposed system can be implemented in other ways. For example, the above-described system embodiments are only illustrative. For example, the division of the above modules is only a logical functional division. In actual implementation, there can be another division manner. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0129] Embodiment three

[0130] The embodiment also provides an electronic device, including one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected with the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device executes the method in the above embodiment one.

[0131] It should be understood that in the embodiment, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, ready programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0132] The memory can include read-only memory and random access memory, and provide instructions and data to the processor. A part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0133] In the implementation process, each step of the above method can be completed by integrated logic circuits of hardware in the processor or instructions in the form of software.

[0134] The method in the embodiment one can be directly embodied as hardware processor execution completion, or combined with hardware and software modules in the processor to complete execution. The software module can be located in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0135] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0136] Embodiment four

[0137] The embodiment also provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in embodiment one.

[0138] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for quality restoration of a frozen electron tomogram image, characterized by, The method comprises the following steps: acquiring a sequence of two-dimensional projections of a frozen electron tomography; performing three-dimensional reconstruction on the sequence of two-dimensional projections to obtain a three-dimensional reconstructed image; processing the three-dimensional reconstructed image by using a trained noise generation model to obtain noise; superimposing the noise on the three-dimensional reconstructed image to obtain a noise superimposed image; inputting the noise superimposed image into a trained image quality recovery model to output an image quality recovered image; the training process of the trained image quality recovery model comprises the following steps: constructing a second training set and a second test set; wherein the second training set and the second test set are both original two-dimensional projection sequences of frozen electron tomography with known image quality recovery results; performing three-dimensional reconstruction on the original two-dimensional projection sequences of the second training set to obtain three-dimensional reconstructed images; and performing filtering on the three-dimensional reconstructed images to obtain filtered reconstructed images; performing denoising on the original two-dimensional projection sequences of the second training set to obtain smoothed projection sequences; performing three-dimensional reconstruction on the smoothed projection sequences to obtain smoothed reconstructed images; and extracting local sparse information from the smoothed reconstructed images; processing the three-dimensional reconstructed images by using the trained noise generation model to obtain noise; and adding the noise to the three-dimensional reconstructed images to obtain noise superimposed input images; inputting the noise superimposed input images into the image quality recovery model to train the model; when the total loss function value of the model no longer decreases, stopping the training to obtain the trained image quality recovery model; the total loss function is the sum of a first loss function, a second loss function and a third loss function; the first loss function is the loss value of the model output image and the three-dimensional reconstructed image; the second loss function is the loss value of the model output image and the local sparse information; the third loss function is the loss value of the contrast information of the model output image and the filtered reconstructed image; the formula of the total loss function is: ； Second loss function is expressed by the formula: wherein, is a corresponding patch of the smooth reconstructor image, is a network input reconstructor image superimposed with noise, denotes a gradient guided sparsity constraint function, denotes a mathematical expectation; denotes gradient information of a network output image patch; denotes gradient information of a smooth reconstructor image patch.

2. The quality restoration method for a frozen electron tomogram image according to claim 1, wherein, the training process of the trained noise generation model comprises the following steps: constructing a first training set and a first test set; wherein the first training set and the first test set are both three-dimensional reconstructed images with known real noise blocks; the real noise block is obtained by extracting a plurality of noise slices from the three-dimensional reconstructed image and then cutting each noise slice; inputting random noise blocks into a generator to generate false noise blocks; The real noise block and the false noise block are input into the discriminator, and the discriminator is trained until the loss function value stops decreasing, and a trained noise generation model is obtained.

3. The quality restoration method for a frozen electron tomogram image according to claim 1, wherein, the network structure of the trained image quality recovery model comprises: an up-sampling layer, a convolution block a1, a maximum pooling layer c1, a convolution block a2, a maximum pooling layer c2, a convolution block a3, a maximum pooling layer c3, a convolution block a4, a maximum pooling layer c4, a convolution block a5, a maximum pooling layer c5, a convolution block a6, an interpolation layer z1, a convolution block b1, a convolution block b2, an interpolation layer z2, a convolution block b3, a convolution block b4, an interpolation layer z3, a convolution block b5, a convolution block b6, an interpolation layer z4, a convolution block b7, a convolution block b8, an interpolation layer z5, a convolution block b9, a three-dimensional convolution layer w1, a down-sampling layer and an output layer, which are sequentially connected. The output end of the upsampling layer is connected with the input end of the interpolation layer z5; the output end of the maximum pooling layer c1 is connected with the input end of the interpolation layer z4; the output end of the maximum pooling layer c2 is connected with the input end of the interpolation layer z3; the output end of the maximum pooling layer c3 is connected with the input end of the interpolation layer z2; and the output end of the maximum pooling layer c4 is connected with the input end of the interpolation layer z1.

4. The quality restoration method for a frozen electron tomogram image according to Claim 1, wherein, Third loss function is expressed by the formula: ​ wherein, is a corresponding patch of the filtered reconstructed image, is a network input reconstructed image superimposed with noise, denotes a filtered guide term loss function, denotes a mathematical expectation; denotes a network output image; wherein ; a three-dimensional Gaussian filter function noted and intensity as parameters, denotes the original reconstructed body image; The first loss function is expressed by the formula: wherein, is a quality restoration network; is a superimposed noisy network input reconstructor image; is an image patch of the original reconstructor image; denotes a self-supervised structure reconstruction loss function, denotes a network output image patch; denotes a mathematical expectation.

5. A quality restoration system for cryo-electron tomogram images, characterized by, Comprise: An acquisition module configured to acquire a sequence of two-dimensional projections of a cryo-electron tomogram; A three-dimensional reconstruction module configured to perform three-dimensional reconstruction on the sequence of two-dimensional projections to obtain a three-dimensional reconstructed image; A noise superposition module configured to process the three-dimensional reconstructed image using a trained noise generation model to obtain noise; Superimpose the noise into the three-dimensional reconstructed image to obtain a noise superimposed image; A quality recovery module configured to input the noise superimposed image into a trained image quality recovery model to output an image quality recovered image; the training process of the trained image quality recovery model comprises: Constructing a second training set and a second test set; wherein the second training set and the second test set are both original two-dimensional projection sequences of a cryo-electron tomogram with known image quality recovery results; Performing three-dimensional reconstruction on the original two-dimensional projection sequences of the second training set to obtain three-dimensional reconstructed images; and performing filtering on the three-dimensional reconstructed images to obtain filtered reconstructed images; Performing denoising on the original two-dimensional projection sequences of the second training set to obtain smoothed projection sequences; performing three-dimensional reconstruction on the smoothed projection sequences to obtain smoothed reconstructed images; and extracting local sparse information from the smoothed reconstructed images; Processing the three-dimensional reconstructed images using the trained noise generation model to obtain noise; and adding the noise into the three-dimensional reconstructed images to obtain noise superimposed input images; Inputting the noise superimposed input images into the image quality recovery model to train the model; stopping the training when the total loss function value of the model no longer decreases, to obtain the trained image quality recovery model; The total loss function is the sum of the first loss function, the second loss function and the third loss function; The first loss function is the loss value of the model output image and the three-dimensional reconstructed image; The second loss function is the loss value of the model output image and the local sparse information; The third loss function is the loss value of the contrast information of the model output image and the filtered reconstructed image; The formula expression of the total loss function is: ； Second loss function is expressed by the formula: wherein, is a corresponding patch of the smooth reconstructor image, is a network input reconstructor image superimposed with noise, denotes a gradient guided sparsity constraint function, denotes a mathematical expectation; denotes gradient information of a network output image patch; denotes gradient information of a smooth reconstructor image patch. 6.An electronic device, comprising: a memory for non-transiently storing computer readable instructions; and a processor for running the computer readable instructions, wherein the computer readable instructions, when run by the processor, perform the method of any one of claims 1-4.

7. A storage medium characterized by, Non-transiently store computer readable instructions, wherein when the non-transient computer readable instructions are executed by a computer, the instructions perform the method of any one of claims 1-4.

Images