Construction method and device of structured light illumination microscope image reconstruction model

Abstract

The invention discloses a construction method and device of a structured light illumination microscope image reconstruction model, and belongs to the technical field of image processing. The method comprises the following steps: acquiring a paired data set of the structured light illumination microscope; preprocessing the data set, and dividing the data set into a training set and a test set; constructing a deep neural network model reconstructed by the structured light illumination microscope, deploying the deep neural network model as a Bayesian neural network, and outputting a reconstruction mean value and a reconstruction standard deviation of reconstruction super-resolution image distribution by the deep neural network model; inferring posterior distribution of parameters of the Bayesian neural network to obtain multiple groups of parameter sets; after training, obtaining a trained Bayesian neural network with a plurality of groups of different parameter sets; and inputting the test data and each group of parameters into the trained Bayesian neural network for model reasoning to obtain a corresponding reconstructed super-resolution image, the reconstructed data uncertainty and the reconstructed model uncertainty. And the uncertainty of quantized reconstruction of the super-resolution image is realized.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and more specifically, relates to a method and device for constructing a structured light illumination microscope image reconstruction model. Background Art

[0002] Due to the diffraction effect of light and the limited aperture of the optical system, traditional optical microscopes have a resolution limit of about λ / 2NA, where λ is the wavelength of the emitted light and NA is the numerical aperture of the objective lens. Generally speaking, the resolution limit of traditional optical microscopes is around 250 nanometers, which severely limits biologists' observation of life processes at a smaller scale.

[0003] In order to break this shackle, a series of super-resolution microscopy methods have emerged. These methods obtain sub-diffraction information of samples beyond the resolution limit through special optical path design and other methods to achieve super-resolution imaging. The emergence of super-resolution microscopy methods has brought a series of major breakthroughs in the field of life sciences, and its inventors have won the 2014 Nobel Prize in Chemistry. Among the many super-resolution imaging methods, structured light illumination microscopy is a technology that uses cosine structured light to optically modulate samples to achieve a two-fold imaging resolution improvement. It has the advantages of low excitation power and high photon efficiency, and is very suitable for living cell imaging. However, as a computational imaging method, the imaging quality of structured light illumination microscopy is highly dependent on the back-end reconstruction algorithm. The reconstruction algorithm is responsible for extracting the modulated super-resolution information from multiple frames of original images and reconstructing an isotropic super-resolution image.

[0004] In recent years, deep learning technology has been widely used to reconstruct structured light microscopy images. However, due to the black box nature of deep learning, the reliability of its reconstruction cannot be guaranteed in practice. At the same time, existing image reconstruction methods cannot quantify the uncertainty of the reconstructed image and cannot provide users with additional reconstruction confidence information. Summary of the invention

[0005] In view of the defects of the related art, the purpose of the present invention is to provide a method and device for constructing a structured light illumination microscope image reconstruction model, aiming to solve the problem that the reliability of the reconstructed image is low and the uncertainty of the reconstructed image cannot be quantified.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a structured light illumination microscope image reconstruction model, comprising:

[0007] S1. Collect paired data sets of cells under structured light illumination microscopy D = {I raw ,I GT}, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios rawAnd the true value image I GT ;

[0008] S2. Construct a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution;

[0009] S3, inferring the posterior distribution of the parameters of the Bayesian neural network, and training the parameters of the Bayesian neural network in S2 by using the training data D1 and a preset loss function to obtain multiple sets of different parameter sets;

[0010] S4. Input the test data D2 into the trained Bayesian neural network for model inference to obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstructed data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

[0011] Optionally, S1 specifically includes:

[0012] S11. Under the same field of view, change the direction and phase of the structured light, collect data of the cell sample under the structured light illumination microscope, and obtain a set of original images I raw ;

[0013] S12, adjusting the excitation power or exposure time of the structured light to obtain multiple groups of original images with different signal-to-noise ratios in the same field of view. raw , get the original image data set of the same field of view;

[0014] S13, select a group of original images with the highest signal-to-noise ratio from the original image data set of the same field of view, and reconstruct the true value image I GT ;

[0015] S14, adjust the field of view of the structured light illumination microscope, repeat S11-S13, collect multiple sets of paired data, and form a data set D = {I raw ,I GT}.

[0016] Optionally, the preprocessing includes background removal, normalization and data enhancement.

[0017] Optionally, the deep neural network model is constructed based on a residual convolutional neural network architecture.

[0018] Optionally, S3 specifically includes:

[0019] S31, inferring the posterior distribution of the parameters of the Bayesian neural network, by using the training data D1 to train the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2, to obtain multiple groups of different parameter sets;

[0020] Among them, other posterior inference algorithms can also be used, such as variational inference, MC-dropout or Deepensemble.

[0021] S32, freeze the network parameters trained in S31, use the training data D1 to train the standard deviation quantization layer parameters of the Bayesian neural network in the Bayesian model deployment module, and obtain a trained Bayesian neural network with multiple groups of different parameter sets.

[0022] Optionally, the feature mapping layer includes a convolution layer to map the input original image to a high-dimensional feature space.

[0023] Optionally, S4 specifically includes:

[0024] S41, input the test data D2 into the trained Bayesian neural network to perform a model inference, and obtain K reconstruction means μ and reconstruction standard deviations σ;

[0025] S42, calculating the mean of the K groups of reconstruction mean values ​​μ as the reconstruction result of the super-resolution image:

[0026]

[0027] Calculate the mean of multiple reconstruction standard deviations σ as the data uncertainty of the reconstruction:

[0028]

[0029] The standard deviation of the multiple reconstruction means μ is calculated as the model uncertainty of the reconstruction:

[0030]

[0031] In a second aspect, the present invention further provides a device for constructing a structured light illumination microscope image reconstruction model, comprising:

[0032] The data acquisition module is used to collect paired data sets D = {I raw ,I GT}, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios raw And the true value image I GT ;

[0033] A model building module is used to build a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution;

[0034] The Bayesian model training module is used to infer the posterior distribution of the parameters of the Bayesian neural network, and train the parameters of the Bayesian neural network in the model building module by using the training data D1 and the preset loss function to obtain multiple sets of different parameter sets;

[0035] The image reconstruction model testing module is used to input the test data D2 into the trained Bayesian neural network for model inference, and obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstruction data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

[0036] Compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects:

[0037] 1. A method for constructing a structured light illumination microscope image reconstruction model. The neural network reconstruction model in this scheme is deployed as a Bayesian neural network, and adopts a block training method to train the feature mapping layer, deep feature extraction layer and mean reconstruction layer of the output reconstruction mean, and the standard deviation quantization layer of the output reconstruction standard deviation. Due to the unique modeling method, training method and posterior distribution sampling average, it has excellent reconstruction performance and can parse clear, reliable and artifact-free super-resolution structures even under low signal-to-noise ratio imaging conditions. Model uncertainty can reflect the generalization performance of the trained model, and data uncertainty reflects the ambiguity from input to output; the uncertainty of reconstructing super-resolution images is quantified by outputting model uncertainty and data uncertainty.

[0038] 2. A method for constructing a structured light illumination microscope image reconstruction model. The reconstruction data uncertainty and reconstruction model uncertainty obtained by this scheme can help users effectively identify unreliable results and improve the credibility of the structured light illumination microscope. The reconstruction uncertainty result can evaluate the reconstruction error without a reference image.

[0039] 3. A method for constructing a structured light illumination microscope image reconstruction model. Given that computational imaging models can easily lead to unreliable erroneous structures or reconstruction artifacts in many cases, the present invention identifies unreliable results of the model by quantifying the uncertainty in the image reconstruction process, and ultimately enables computational imaging based on deep learning to avoid such problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of a method for constructing a structured light illumination microscope image reconstruction model provided by an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of the results of a method for constructing a structured light illumination microscope image reconstruction model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0043] The contents involved in the above embodiment are described below in conjunction with a preferred embodiment.

[0044] Embodiment 1

[0045] The present invention provides a method for constructing a structured light illumination microscope image reconstruction model, comprising:

[0046] S1. Collect paired data sets of cells under structured light illumination microscopy D = {I raw ,I GT}, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios raw And the true value image I GT ;

[0047] S2. Construct a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution;

[0048] S3, inferring the posterior distribution of the parameters of the Bayesian neural network, and training the parameters of the Bayesian neural network in S2 by using the training data D1 and a preset loss function to obtain multiple sets of different parameter sets;

[0049] S4. Input the test data D2 into the trained Bayesian neural network for model inference to obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstructed data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

[0050] Among them, S1 specifically includes:

[0051] S11. Under the same field of view, change the direction and phase of the structured light, collect data of the cell sample under the structured light illumination microscope, and obtain a set of original images I raw ;

[0052] S12, adjusting the excitation power or exposure time of the structured light to obtain multiple groups of original images with different signal-to-noise ratios in the same field of view. raw , get the original image data set of the same field of view;

[0053] S13, select a group of original images with the highest signal-to-noise ratio from the original image data set of the same field of view, and reconstruct the true value image I GT ;

[0054] S14, adjust the field of view of the structured light illumination microscope, repeat S11-S13, collect multiple sets of paired data, and form a data set D = {I raw ,I GT}.

[0055] refer to Figure 1 , Figure 1 This is the overall flow chart of this program.

[0056] Using sinusoidal interference fringes to illuminate cell samples for imaging, the imaging model is:

[0057]

[0058] Where θ is the azimuth angle of the structured light, p θ is the wave vector of structured light, c θ is the structured light modulation depth, is the initial phase of the structured light, f(x) represents the fluorescence distribution of the sample, h(x) is the point spread function of the imaging system, Represents a convolution operation.

[0059] For the same field of view, the structured light is repeated three times in the spatial direction with an interval angle of 120°. The initial phase is also repeated three times, that is, m = 0, ±1. The structured light is imaged nine times with different directions and phases to obtain a set of nine frames of diffraction-limited original images I raw By adjusting the excitation power or exposure time, multiple sets of original images with different signal-to-noise ratios can be collected. At the same time, a set of original images with extremely high signal-to-noise ratios are collected and reconstructed using traditional reconstruction methods to obtain a super-resolution image with a resolution of about 100 nanometers. GT Repeat the above operation for different fields of view of the cell sample to collect multiple sets of paired data to form a data set D = {I raw ,I GT}.

[0060] The collected data set D is preprocessed, and the preprocessing includes background removal, normalization and data enhancement. GT The camera background is subtracted from the image; the background removal algorithm is then used to remove the out-of-focus background fluorescence; then the I raw and I GT Normalization is performed to scale the dynamic range of the image to the range of [0,1] to facilitate subsequent neural network processing; then, data enhancement techniques such as random cropping and random flipping are used to increase data diversity to avoid overfitting in subsequent neural network training; finally, the preprocessed data set is randomly divided into training set D1 and test set D2 in a ratio of 4:1, which are used for subsequent model training and testing, respectively.

[0061] A deep neural network model for structured light illumination microscopy reconstruction is constructed. Its input is set as a set of diffraction-limited original images (a set of nine original images), and the output target is the distribution of the reconstructed super-resolution image. The mapping dimension is

[0062] The distribution of the reconstructed super-resolution image is a spatially heteroscedastic Gaussian distribution. The distribution of the reconstructed super-resolution image can be parameterized as a spatially varying mean and standard deviation, so the actual output of the neural network model is the mean and standard deviation of the distribution of the reconstructed super-resolution image.

[0063] Optionally, S3 specifically includes:

[0064] S31, using the SGLD algorithm to infer the posterior distribution of the parameters of the Bayesian neural network, and using the training data D1 to train the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2 to obtain multiple groups of different parameter sets;

[0065] S32, freeze the network parameters trained in S31, use the training data D1 to train the standard deviation quantization layer parameters of the Bayesian neural network in the Bayesian model deployment module, and obtain a trained Bayesian neural network with multiple groups of different parameter sets.

[0066] The constructed neural network model is based on the residual convolutional neural network architecture, which uses convolutional layer, ReLU activation function, attention mechanism, and jump connection as basic units. It consists of four parts: feature mapping layer, deep feature extraction layer, mean reconstruction layer, and standard deviation quantization layer. Among them, the feature mapping layer is a convolutional layer, which maps nine frames of original images to a high-dimensional feature space. The deep feature extraction layer is composed of several residual group cascades and jump connections. The deep feature extraction layer includes multiple residual groups, each residual group is composed of several residual blocks, the mean reconstruction layer includes an upsampling layer and a convolution layer, and the standard deviation quantization layer includes a convolution layer and an activation function. The feature mapping layer maps the input original image to a high-dimensional feature space, the deep feature extraction layer extracts the modulated super-resolution information in the high-dimensional feature space, and the mean reconstruction layer and the standard deviation quantization layer are used to output the mean and standard deviation of the super-resolution image respectively. The data type of the trainable parameters of the neural network is single-precision floating point (32bit floating point number). Further, the deep neural network model in this scheme is not limited to the residual convolutional neural network architecture.

[0067] The input data passes through the feature mapping layer and the deep feature extraction layer to obtain the deep features:

[0068] F=Conv 1→64 (I raw )+RG 64→64 ((…RG 64→64 (Conv 1→64 (I raw )))

[0069] Among them, Conv represents the convolutional layer, RG represents the residual group, and the subscript represents the change in the number of feature channels;

[0070] Each residual group consists of several residual blocks:

[0071] RG 64→64 (x) = x + RB 64→64 (…RB 64→64 (x)

[0072] Each residual block is specifically:

[0073] RB 64→64 (x) = x + CA (Conv 64→64 (ReLU(Conv 64→64 (x))))

[0074] Among them, ReLU is the activation function and CA is the channel attention mechanism.

[0075] The mean reconstruction layer uses upsampling modules and convolutional layers to reconstruct the extracted deep features into a super-resolution image:

[0076] μ=Conv 64→1 (PS(F))

[0077] Among them, PS means PixelShuffle upsampling.

[0078] The standard deviation quantization layer is also based on deep features, using convolutional layers and activation functions to predict the standard deviation of the reconstructed image distribution:

[0079] σ=ReLU(Conv 64→64 (ReLU(Conv 64→64 (F))))

[0080] The constructed neural network model is modeled as a Bayesian neural network, that is, the network parameters are no longer a set of fixed values, but obey a certain probability distribution. In Bayesian inference, this parameter distribution is called the posterior distribution, that is, φ~p(φ|D1). Any sample φ~p(φ|D1) in the posterior distribution can fit the training set data D1 well.

[0081] The inference of the posterior distribution will be implemented using the SGLD algorithm in the subsequent training steps. Specifically, the SGLD algorithm is used to infer the posterior distribution of the parameters of the Bayesian neural network. During the inference process, the feature mapping layer parameters, deep feature extraction layer parameters, and mean reconstruction layer parameters of the Bayesian neural network in S2 are trained using the training data D1 and the dual-domain loss function to obtain multiple sets of different parameter sets. After each iterative update of the network parameters, the SGLD algorithm samples a Gaussian noise and applies it to the network parameters to approximate the true posterior distribution.

[0082] Among them, random Gaussian noise: η t ~N(0,v t ), the subscript indicates the tth iteration, v t represents the network learning rate at the tth iteration

[0083] Among them, the loss function used in training is the output reconstruction mean μ and the training target I GT In the dual-domain error of the spatial domain and the frequency domain, the dual-domain loss function is: in, represents Fourier transform.

[0084] The parameters obtained through training are the network parameters related to feature extraction and mean reconstruction in the Bayesian neural network, and the trained parameters are frozen.

[0085] The divided training data D1 is used to train the standard deviation quantization layer in the Bayesian neural network. The loss function used in the training is Gaussian heteroscedastic loss:

[0086] The test data D2 and each set of parameters are input into the trained Bayesian neural network for reconstruction. Different sets of parameters from the posterior distribution sampled using the SGLD algorithm can be used to infer the distribution of multiple reconstructed super-resolution images, and the mean of multiple distribution means is calculated as the final reconstruction result. The mean of multiple distribution standard deviations is calculated as the final reconstruction standard deviation (also known as data uncertainty), and the standard deviation of multiple distribution means is calculated as model uncertainty.

[0087] Optionally, S4 specifically includes:

[0088] S41, input the test data D2 into the trained Bayesian neural network to perform a model inference, and obtain K reconstruction means μ and reconstruction standard deviations σ;

[0089] S42, calculating the mean of the K groups of reconstruction mean values ​​μ as the reconstruction result of the super-resolution image:

[0090]

[0091] Calculate the mean of multiple reconstruction standard deviations σ as the data uncertainty of the reconstruction:

[0092]

[0093] The standard deviation of the multiple reconstruction means μ is calculated as the model uncertainty of the reconstruction:

[0094]

[0095] like Figure 2 As shown, compared with the reconstruction results of the existing methods, the reconstruction results of the present invention have higher resolution and better analysis of spatial details. In addition, the data uncertainty quantified by the present invention indicates the standard deviation of the super-resolution image distribution, and the credibility of each area of ​​the reconstructed image can be obtained by combining with subsequent statistical analysis. The model uncertainty quantified by the present invention describes the consistency of the reconstruction results obtained by different sampling of the posterior distribution of the model parameters, and the amplitude of the model uncertainty can be used to judge the matching degree between the neural network model and the input data.

[0096] The embodiment of the present invention deploys the constructed neural network model as a Bayesian neural network to make it obey the unknown posterior distribution, so that multiple sets of parameter sets can be obtained when training parameters, so that multiple results can be obtained when performing model reasoning; at the same time, the distribution of the reconstructed super-resolution image is designed to be a Gaussian distribution with spatial heteroskedasticity, so that the actual output of the neural network model is the mean and standard deviation of the reconstructed super-resolution image distribution, so that it can be calculated based on multiple means and standard deviations, and the mean of multiple distribution means is obtained as the final reconstruction result, the mean of multiple distribution standard deviations is used as the final reconstruction standard deviation, representing data uncertainty, and the standard deviation of multiple distribution means is used as model uncertainty. The technical problems of low feasibility of reconstructed images and inability to quantify the uncertainty of reconstructed images are solved, accurate image reconstruction is achieved, and the uncertainty of reconstructed super-resolution images can be quantified. The obtained reconstruction data uncertainty and reconstruction model uncertainty can help users effectively identify the beneficial effect of unreliable results.

[0097] Embodiment 2

[0098] The present invention also provides a device for constructing a structured light illumination microscope image reconstruction model, which is used to execute the method in the first embodiment, including:

[0099] The data acquisition module is used to collect paired data sets D = {I raw ,I GT}, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios raw And the true value image I GT ;

[0100] A model building module is used to build a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution;

[0101] The Bayesian model training module is used to infer the posterior distribution of the parameters of the Bayesian neural network, and train the parameters of the Bayesian neural network in the model building module by using the training data D1 and the preset loss function to obtain multiple sets of different parameter sets;

[0102] The image reconstruction model testing module is used to input the test data D2 into the trained Bayesian neural network for model inference, and obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstruction data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

[0103] A device for constructing a structured light illumination microscope image reconstruction model provided in an embodiment of the present invention is used to execute a method for constructing a structured light illumination microscope image reconstruction model provided in any first embodiment of the present invention, and has corresponding functional modules and beneficial effects.

[0104] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a structured light illumination microscope image reconstruction model, characterized in that: include: S1. Collect paired data sets of cells under structured light illumination microscopy D = {I raw , I GT }, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios raw And the true value image I GT ; S2. Construct a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution; S3, inferring the posterior distribution of the parameters of the Bayesian neural network, and training the parameters of the Bayesian neural network in S2 by using the training data D1 and a preset loss function to obtain multiple sets of different parameter sets; S4. Input the test data D2 into the trained Bayesian neural network for model inference to obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstructed data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

2. The method according to claim 1, characterized in that S1 specifically includes: S11. Under the same field of view, change the direction and phase of the structured light, collect data of the cell sample under the structured light illumination microscope, and obtain a set of original images I raw ; S12, adjusting the excitation power or exposure time of the structured light to obtain multiple groups of original images with different signal-to-noise ratios in the same field of view. raw , get the original image data set of the same field of view; S13, select a group of original images with the highest signal-to-noise ratio from the original image data set of the same field of view, and reconstruct the true value image I GT ; S14, adjust the field of view of the structured light illumination microscope, repeat S11-S13, collect multiple sets of paired data, and form a data set D = {I raw , I GT }.

3. The construction method according to claim 1, characterized in that: The preprocessing includes background removal, normalization and data enhancement.

4. The construction method according to claim 1, characterized in that: The construction of the deep neural network model is based on the residual convolutional neural network architecture.

5. The construction method according to claim 1, characterized in that: S3 specifically includes: S31, inferring the posterior distribution of the parameters of the Bayesian neural network, by using the training data D1 to train the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2, to obtain multiple groups of different parameter sets; S32, freeze the network parameters trained in S31, use the training data D1 to train the standard deviation quantization layer parameters of the Bayesian neural network in the Bayesian model deployment module, and obtain a trained Bayesian neural network with multiple groups of different parameter sets.

6. The construction method according to claim 5, characterized in that: The feature mapping layer includes a convolution layer, which maps the input original image to a high-dimensional feature space.

7. The method according to claim 1, characterized in that S4 specifically includes: S41, input the test data D2 into the trained Bayesian neural network to perform a model inference, and obtain K reconstruction means μ and reconstruction standard deviations σ; S42, calculating the mean of the K groups of reconstruction mean values ​​μ as the reconstruction result of the super-resolution image: Calculate the mean of multiple reconstruction standard deviations σ as the data uncertainty of the reconstruction: The standard deviation of the multiple reconstruction means μ is calculated as the model uncertainty of the reconstruction:

8. A device for constructing a structured light illumination microscope image reconstruction model, characterized in that: include: The data acquisition module is used to collect paired data sets D = {I raw , I GT }, and is divided into a training set D1 and a test set D2; the data set D includes original images I with different signal-to-noise ratios raw And the true value image I GT ; A model building module is used to build a deep neural network model for structured light illumination microscopy reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the reconstructed super-resolution image distribution; The Bayesian model training module is used to infer the posterior distribution of the parameters of the Bayesian neural network, and train the parameters of the Bayesian neural network in the model building module by using the training data D1 and the preset loss function to obtain multiple sets of different parameter sets; The image reconstruction model testing module is used to input the test data D2 into the trained Bayesian neural network for model inference, and obtain the reconstruction mean μ and reconstruction standard deviation σ of multiple groups of reconstructed super-resolution image distributions; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstruction data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstruction model uncertainty.

Images